You are on page 1of 22

This copy is for personal use only. To order printed copies, contact reprints@rsna.

org
1676
| GENITOURINARY IMAGING

Artificial Intelligence–assisted
Prostate Cancer Diagnosis:
Radiologic-Pathologic Correlation
Lidia Alcalá Mata, MD
Juan Antonio Retamero, MD The classic prostate cancer (PCa) diagnostic pathway that is based
Rajan T. Gupta, MD on prostate-specific antigen (PSA) levels and the findings of digital
Roberto García Figueras, MD, PhD rectal examination followed by systematic biopsy has shown mul-
Antonio Luna, MD, PhD tiple limitations. The use of multiparametric MRI (mpMRI) is now
RADIOLOGY-PATHOLOGY COLLECTION

widely accepted in men with clinical suspicion for PCa. In addition,


Abbreviations: ADC = apparent diffusion clinical information, PSA density, risk calculators, and genomic and
coefficient, AI = artificial intelligence, CAD = other “omics” biomarkers are being used to improve risk stratifica-
computer-aided diagnosis, csPCa = clinically
significant PCa, DRE = digital rectal examina- tion. On the basis of mpMRI and MRI-targeted biopsies (MRI-
tion, DWI =diffusion-weighted imaging, EAU = TBx), new diagnostic pathways have been established, aiming to
European Association of Urology, ISUP = In-
ternational Society of Urological Pathology,
improve the limitations of the classic diagnostic approach. However,
mpMRI = multiparametric MRI, MRI-TBx = these pathways still show limitations associated with mpMRI and
MRI-targeted biopsy, ncsPCa = non–clinically MRI-TBx. Definitive PCa diagnosis is made on the basis of histo-
significant PCa, NLP = natural language pro-
cessing, PCa = prostate cancer, PI-RADS = pathologic Gleason grading, which has demonstrated an excellent
Prostate Imaging Reporting and Data System, correlation with clinical outcomes. However, Gleason grading is
PSA = prostate-specific antigen, TRUS-Bx =
transrectal US-guided biopsy
done subjectively by pathologists and involves poor reproducibility,
and PCa may have a heterogeneous distribution of histologic pat-
RadioGraphics 2021; 41:1676–1697
terns. Thus, important discrepancies persist between biopsy tumor
https://doi.org/10.1148/rg.2021210020 grading and final whole-organ pathologic assessment after radi-
Content Codes: cal prostatectomy. PCa offers a unique opportunity to establish
From the Department of Radiology, Clínica Las a real radiologic-pathologic correlation, as whole-mount radical
Nieves, HT Médica, Calle Carmelo Torres Núm prostatectomy specimens permit a complete spatial relationship
2, 23007 Jaén, Spain (L.A.M., A.L.); Paige.AI,
New York, NY (J.A.R.); Department of Radiol-
with mpMRI. Artificial intelligence is increasingly being applied to
ogy, Duke University Medical Center, Durham, radiologic and pathologic images to improve clinical accuracy and
NC (R.T.G.); and Department of Radiology, efficiency in PCa diagnosis. This review delineates current PCa di-
Complexo Hospitalario Universitario de San-
tiago de Compostela, Santiago de Compostela, agnostic pathways, with a focus on the role of mpMRI, MRI-TBx,
Spain (R.G.F.). Recipient of a Certificate of and pathologic analysis. An overview of the expected improvements
Merit award for an education exhibit at the 2020
RSNA Annual Meeting. Received January 31, in PCa diagnosis derived from the use of artificial intelligence, inte-
2021; revision requested March 9 and received grated radiologic-pathologic systems, and decision support tools for
May 26; accepted May 27. For this journal-based
SA-CME activity, the authors J.A.R., R.T.G.,
multidisciplinary teams is provided.
and A.L. have provided disclosures (see end of
article); all other authors, the editor, and the re- An invited commentary by Purysko is available online.
viewers have disclosed no relevant relationships. Online supplemental material is available for this article.
Address correspondence to L.A.M. (e-mail:
©
l.alcala.o@htime.org). RSNA, 2021 • radiographics.rsna.org
©
RSNA, 2021

SA-CME LEARNING OBJECTIVES


Introduction
After completing this journal-based SA-CME Prostate cancer (PCa) is the second most diagnosed cancer and the
activity, participants will be able to:
fourth cause of cancer-related death in men worldwide (1). Despite
„ State the different current diagnostic
pathways for prostate cancer, including
its high incidence, no screening programs have resulted in signifi-
their advantages and limitations. cant mortality decrease (2,3). The most widespread tools used in
Describe the existing AI tools that
„ determining patient risk are serum prostate-specific antigen (PSA)
streamline the prostate cancer diagnostic and digital rectal examination, followed by systematic transrectal
process. US-guided biopsy (TRUS-Bx), which samples regions of the gland
Explain the benefits of digital patho-
„ systematically without a defined target.
logic analysis and integrated radiologic-
pathologic diagnosis in prostate cancer.
See rsna.org/learning-center-rg.
RG  •  Volume 41  Number 6 Alcalá Mata et al  1677

TEACHING POINTS
„ The use of MRI to help determine which patients undergo
MRI-TBx is known as the MRI pathway. Alternatively, MRI
combined with TRUS-Bx is used in the combined biopsy path-
way, in which patients with negative MRI findings undergo
TRUS-Bx, and those with suspicious lesions at MRI undergo
both TRUS-Bx and MRI-TBx.
„ Deep learning and CNNs are currently the most used in the
development of AI tools for clinical diagnosis. Also, radiomics
involves the extraction of quantitative data from radiologic
images and, when combined with machine learning tech-
niques, it permits the identification of image characteristics
beyond those obtained by radiologists.
„ There are two main disadvantages to the development of
these AI tools. First, the large amount of labeled data neces-
sary for adequate AI training requires notable input and time
from experienced radiologists. Second, the false-positive rate
of these tools is up to 50%, which requires radiologist input
during reading to validate or reject the lesions presented by
AI.
Figure 1.  Outline of PCa early detection program objectives.
„ The advantages of a digital pathology workflow are multiple,
The main objective of an early detection program is to increase
including improved efficiency and better user experience the csPCa detection rate while decreasing the detection rate for
compared with light microscopy. Digital histologic images ncsPCa. The purpose is to reduce overtreatment and its com-
can be viewed anywhere by using a computer instead of a mi- plications, the number of unnecessary examinations (MRI and
croscope, which makes them accessible beyond the pathol- biopsy), and the morbidity and costs derived from them.
ogy laboratory. Also, digital pathology permits the application
of AI tools to histologic images to aid diagnosis, as it occurs
in radiology. Therefore, digital pathology is the foundation of reading systems like Prostate Imaging Reporting
computational pathology and CAD in pathology.
and Data System (PI-RADS) version 2.1 (8,9).
„ To solve this problem and facilitate the diagnosis, treatment,
In spite of these limitations, the use of mpMRI
and follow-up of patients with PCa by multidisciplinary teams,
decision support tools are created, based on, and validated
is widely accepted in men suspected of having
by clinical guidelines or expert management. These tools fa- csPCA in both biopsy-naive patients and those
cilitate finding relevant data in the clinical history, including who have undergone TRUS-Bx. Currently, the
analytical values, previous studies, histologic results, and treat- role of MRI-targeted biopsy (MRI-TBx) is grow-
ments received.
ing, although there is variability in the timing and
type of biopsy to be used.
The use of MRI to help determine which
PCa is a heterogeneous disease with vari- patients undergo MRI-TBx is known as the
able prevalence and a wide range of prognoses MRI pathway. Alternatively, MRI combined with
(4). Therefore, it is important to differentiate TRUS-Bx is used in the combined biopsy path-
between indolent or non–clinically significant way, in which patients with negative MRI find-
PCa (ncsPCa), which does not increase mortal- ings undergo TRUS-Bx, and those with suspi-
ity, requires no immediate treatment, and can be cious lesions at MRI undergo both TRUS-Bx and
managed with active surveillance, and clinically MRI-TBx (10).
significant PCa (csPCa), which can become The current lack of uniformity in PCa diagno-
metastatic, can cause death, and requires defini- sis is partly due to differences between U.S. and
tive treatment. The emphasis should be directed European guidelines, with differences regarding
at detecting the latter, aiming to decrease its un- timing and use of imaging techniques, risk calcu-
derdiagnosis and undertreatment, while reduc- lators, and MRI-guided biopsy (11).
ing the detection rate of the former to minimize Historically, Gleason grading has demonstrated
overdiagnosis and unnecessary treatment (Fig an excellent correlation with clinical outcomes.
1). Age, ethnicity, genetic mutations, and clinical Several revisions have improved patient risk strati-
history are helpful data in determining csPCa fication and better predicted their outcome. On
patient risk. Also, PSA density, genomic and the basis of the Gleason score, the International
other “omics” biomarkers, and risk calculators Society of Urological Pathology (ISUP) consensus
may play a key role (5–7). guideline recommends grade grouping to facilitate
Multiparametric MRI (mpMRI) is gaining im- communication of pathologic diagnoses to urolo-
portance in csPCa diagnosis, although it requires gists and patients (Table 1) (12). However, some
experienced radiologists and has a steep learn- unresolved areas regarding diagnosis and quanti-
ing curve despite the existence of standardized fication of high-grade patterns remain (13). Also,
1678  October Special Issue 2021 radiographics.rsna.org

Gleason grading is subjectively done by patholo- Table 1: ISUP Grades and Gleason Scores
gists and involves inter- and intrareader vari-
ability and poor reproducibility (14). In addition, ISUP Grade Gleason Score
PCa may be multifocal and have a heterogeneous 1 2–6
Gleason pattern distribution within the same 2 7 (3+4)
patient (15). Likewise, mpMRI often underes- 3 7 (4+3)
timates pathologic tumor size, particularly in 4 8 (4+4), (3+5), (5+3)
lesions with lower PI-RADS scores and smaller 5 9–10 (4+5), (5+4), (5+5)
size (16). Tumor heterogeneity and mpMRI
limitations, along with biopsy error, lead to dis- Source.—Reference 12.
crepancies between biopsy tumor grade and final
whole-organ pathologic assessment after radical
prostatectomy (17). Natural language processing (NLP) tools
Artificial intelligence (AI) is being applied to read, understand, and classify data in text such
medical imaging, both in radiology and pathology as clinical history or diagnostic reports. Machine
(18). The purpose is to increase clinical efficiency learning creates mathematical algorithms that can
and the interpretability of results and support automatically learn from certain tasks, requir-
decision making, both in the detection and risk ing minimal or no human intervention (23). In
stratification of disease in patients with PCa, and supervised learning, computers learn from train-
after treatment follow-up. In mpMRI PCa diagno- ing data that were previously labeled by humans,
sis, several AI tools may help radiologists in auto- along with the expected results. A greater amount
matic prostate gland segmentation as well as lesion and variability of data used during the training
detection and characterization (19). Similarly, stage typically results in better learning and more
there is growing evidence of improving diagnostic accurate outputs. Algorithms then undergo a vali-
performance when pathologists are aided by AI dation process whereby unlabeled data are intro-
(20,21). Also, PCa offers a unique opportunity to duced, and algorithmic output is contrasted with
establish radiologic-pathologic correlation, as radi- the results generated by human operators. Ideally,
cal prostatectomy specimens permit a complete this validation is done by exposing algorithms to
spatial relationship with mpMRI. Thus, AI-based data from multiple sources. After validation, the
tools and integrated radiologic-pathologic di- algorithm is calibrated to adjust to local variables,
agnosis, along with clinical and biomarker data, to improve accuracy and increase sensitivity and
show promise in reducing interreader variability, specificity (18). In unsupervised learning, the
both among radiologists and pathologists, and system learns directly from unlabeled raw data
solve some of the multiple weak links that exist in without human-made annotations, separating
PCa diagnosis. AI-assisted decision support tools the samples into different classes on the basis of
for multidisciplinary teams may facilitate com- the characteristics of the training data and the
munication among radiologists, pathologists, and random variables under study (20).
urologists and analyze multimodal data by using Subtypes of machine learning include deep
algorithms that are able to manage large datasets. learning, neural networks, and convolutional
These solutions may recognize hidden relation- neural networks (CNNs), in which the systems
ships among different biomarkers and suggest the are formed by layers that interact with each other
most appropriate therapy (20,22). and can learn complex processes. Deep learning
This review delineates current PCa diagnostic and CNNs are currently the most used in the de-
pathways with a focus on the role of mpMRI, velopment of AI tools for clinical diagnosis (23).
MRI-TBx, and pathologic analysis. It also pro- Also, radiomics involves the extraction of quan-
vides an overview of the role of AI in PCa radiol- titative data from radiologic images, and when
ogy and pathology. Finally, the future of PCa combined with machine learning techniques, it
diagnosis and management based on integrated permits the identification of image characteristics
diagnostic systems is discussed. beyond those obtained by radiologists.
AI-based tools may improve the diagnosis,
General Concepts of AI prognosis, and treatment of PCa (Fig 2, Table 2).
AI is the field of computer science that develops Machine learning solutions can perform auto-
systems capable of performing tasks that are matic prostatic mpMRI segmentation, identify
commonly associated with human intelligence suspicious areas, propose a PI-RADS score,
(18). It is gaining interest in image analysis ap- and estimate risk of histologic aggressiveness.
plied to radiology and pathology, thanks to its Algorithms have demonstrated a similar area
ability to automatically extract information from under the curve (AUC) to humans in the char-
medical images and make diagnostic predictions. acterization of csPCa with mpMRI by using the
RG  •  Volume 41  Number 6 Alcalá Mata et al  1679

Figure 2.  Diagram shows the treatment process in patients with potential PCa risk and the possible role of AI (machine learning
[ML], deep learning, and radiomics) in this process. NLP tools can be used to help enroll patients in the screening program. Risk calcu-
lators based on big data suggest tests to be performed. AI also assists the diagnostic process by automatically detecting and classifying
lesions according to PI-RADS version 2.1. Automatic prostatic and suspicious lesion segmentation with mpMRI can be done before
MRI-TBx. The histologic assessment of the biopsy sample, tumor detection, and the corresponding Gleason score can be aided by AI
algorithms. In addition, the whole process can be visualized and managed by multidisciplinary teams through integrated diagnostic
platforms (decision support tools) that can propose next diagnostic and treatment steps according to updated clinical guidelines by
using optimization and planning algorithms. DRE = digital rectal examination.

PI-RADS system (19). Also, machine learning recommended no further PSA testing in the
algorithms can optimize fusion MRI/US biopsies general population in 2012 (grade D recom-
by improving image alignment and considering mendation) (3). In 2017, the revised USPSTF
prostate gland deformation during the procedure recommendation advised that men between the
(20). Another growing area in AI is the identifi- ages of 50 and 69 years be informed of the risks
cation and validation of genetic signatures and and benefits of PSA screening for PCa (grade C
other “omics” biomarkers that can help in the recommendation), recommending PSA determi-
diagnosis and prognosis of PCa (6). Computer- nations only after patients are informed of their
aided diagnosis (CAD) in PCa histopathologic individual risk for csPCa. This introduces the
analysis by using AI is expected to reduce reading concept of early PCa detection, becoming es-
times, improving reproducibility and accuracy sential to quantify individual csPCa patient risk
(20,32). AI may play a role in PCa treatment, to minimize overdiagnosis, overtreatment, and
building three-dimensional models that could associated morbidity (2).
be integrated in augmented and virtual reality Epidemiologic factors that increase the
solutions, improving tumor visualization during likelihood of csPCa include ethnicity (ie, Black
surgery or image-guided focal therapy procedures patients [49%] and White patients [26%]), tumor
and providing a more accurate lesion segmenta- volume, younger age of onset, and BRCA2 muta-
tion and its relationship with neighboring organs tion carriers (35,36). Furthermore, the genomic
to optimize radiation therapy treatments (33). landscape of csPCa is complex and heteroge-
neous and is associated with TP53 and DNA
Screening and Early PCa Detection repair gene aberrations. Therefore, the genomic
analysis of primary PCa biopsies shows promise
Current Status for early identification of tumors that evolve into
The use of PSA as a screening tool for PCa more aggressive variants (37).
detection led to an incidence increase of PCa There are discrepancies and similarities
(20%–50%), which was mostly associated with regarding risk factor management between the
ncsPCa, and was accompanied by a rise in European Association of Urology (EAU) and
biopsy and treatment-related complications but their U.S. counterpart, the National Comprehen-
without an overall survival increase or mortal- sive Cancer Network (NCCN) (38,39) (Table
ity rate reduction (2,3,34). For this reason, the 3). Both guidelines agree that patients undergo-
U.S. Preventive Services Task Force (USPSTF) ing early detection programs should be informed
1680  October Special Issue 2021 radiographics.rsna.org

Table 2: Advantages and Limitations of AI Solutions in PCa Diagnosis

AI Solution Tool Type Advantages Limitations


Automatic risk factor NLP Greater sensitivity and specificity in Software costs
detection in clinical identifying patients at risk for csPCa Integration with EMR may
history (PSA or DRE) Less clinician time required be challenging
Fewer false-negative findings due to EMR language heteroge-
human oversight neity may lead to false-
negative findings
PCa diagnostic predic- CNNs Higher sensitivity, specificity, PPV, Multiple tools using differ-
tion tools, combining NPV, and AUC in PCa prediction ent variables may reach
PSA with other ana- than with conventional tools similar results
lytical, genetic, and Decreased false-positive findings due
“omics” biomarkers to high PSA
Potential to reduce unnecessary
biopsies
May help decrease false-positive rate
in screening programs
Integration of several biomarkers
may lead to improve PCa diagno-
sis, risk stratification, and outcome
Risk calculator Logistic regression Fewer unnecessary biopsies Result variability depends
models Detect patients at risk in need of on choice of tool
Big data biopsy with negative findings at Variable heterogeneity
imaging (ethnicity, genetic vari-
Fewer unnecessary MRI studies ants, mpMRI data)
CAD for MRI lesion CNNs Fewer false-negative findings Low PPV due to high
detection and clas- Radiomics Less radiologist time required, accel- false-positive rate
sification Texture analysis erated learning curve for nonex- Large amount of labeled
pert radiologists data required for train-
Increased sensitivity in detection of ing
TZ lesions, increased sensitivity in Data heterogeneity
PI-RADS 3 lesions No regulatory clearance
for diagnostic use
Automatic segmenta- Machine learning Less radiologist time Radiologist needs to check
tion of prostate gland Texture analysis Provide quantitative data such as segmentation accuracy
and suspicious lesions prostate and lesion volume Limited results in mpMRI
with T2-weighted Generate files needed to perform of suboptimal quality
sequences MRI-TBx Imprecise in posttreatment
prostates or those with
heterogeneous signal
intensity
Real-time 3D deform- Machine learning Improved results in MRI/US fusion Need for controlled rate of
able image registra- Radiomics biopsy elastic deformation
tion algorithms Permits precise radiation therapy Expensive integration
planning software
Aids robotic surgery
AI in pathology Machine learning Improves sensitivity, specificity, Relies on digital pathology
predictive values, and diagnostic Software integration may
concordance be difficult
Better biomarker quantification
AI-assisted decision Machine learning Better communication among radiol- Software costs
support tools NLP ogists, pathologists, and urologists Difficult workflow integra-
Optimization and Proposes management plan follow- tion
planning ing automatically updated clinical Tools under development
Big data guidelines and not yet cleared for
Time savings with uncomplicated primary diagnosis
cases in tumor boards
Sources.—References 6, 23, and 24–31. AUC = area under the curve, CAD = computer-aided diagnosis, CNN =
convolutional neural network, EMR = electronic medical record, FDA = U.S. Food and Drug Administration,
NPV = negative predictive value, PPV = positive predictive value, 3D = three-dimensional, TZ = transitional zone.
RG  •  Volume 41  Number 6 Alcalá Mata et al  1681

Table 3: Patient Populations Included in PCa Early Detection Programs

NCCN Criteria EAU Criteria


Men aged 45–75 years Men aged ≥50 years
NA Men aged ≥45 years with family history of PCa
African American men aged ≥40 years Men of African descent aged ≥45 years old
Men aged >40 years carrying BRCA2 or BRCA1 muta- Men aged >40 years carrying BRCA2 mutations
tions
Sources.—References 38 and 39.
Note.— EAU = European Association of Urology, NA = not applicable, NCCN = National Comprehensive
Cancer Network.

about what that program consists of, its risks and biopsy and low-risk patients who should avoid
benefits, and the exclusion of patients with low it, and the choice of biopsy type that optimizes
life expectancy. csPCa detection.
Classic independent indicators of the need
Future of PCa Screening for biopsy such as abnormal DRE and elevated
It is essential to find tools to decrease the high PSA levels show evident limitations (43). When
false-positive rate associated with elevated PSA csPCa is suspected, there is mixed evidence on
levels in PCa diagnosis. Several studies have shown the use of screening biomarkers such as Pros-
that CNN models using different combinations tate Health Index (PHI), SelectMDx, 4KScore,
of data on age, total and free PSA, digital rectal ExoDx Prostate Test, or prostate cancer antigen 3
examination (DRE), and prostate volume can (PCA3) before biopsy (38,39). The most updated
significantly diminish the false-positive findings guidelines suggest mpMRI before biopsy (38,39).
linked to elevated PSA and help prevent unneces- Patients with a PI-RADS score less than or equal
sary biopsies (24,40,41). to 2 at mpMRI have low PCa suspicion and, if
After the USPSTF’s recommendation to elim- the patient agrees, a biopsy may not be necessary.
inate routine PSA testing, the incidence of PCa Patients with a PI-RADS score greater than or
during the following 5 years in the United States equal to 3 should undergo a TRUS-Bx together
increased at the expense of regional and distant with MRI-TBx of the suspicious lesion, accord-
disease (by 1.5% and 3.5%, respectively) and a ing to EAU guidelines (39). This contrasts with
decrease in the proportion of local-stage disease the PI-RADS version 2.1 guideline for patients
(7.6%) (42). This indicates the need for a new with a PI-RADS score of 3, who may undergo
strategy in the treatment of men between 50 and biopsy only if deemed appropriate according to
69 years of age and younger men with additional clinical data and risk stratification (9) (Fig 3).
risk factors. A promising strategy introduces Clinical guidelines support an mpMRI-based ap-
screening intervals based on PCa risk and biopsy proach, but some issues need to be addressed to
recommendation after a risk-based calculation by make this technique more broadly used (11,43–
using risk calculators and mpMRI. This approach 45) (Table 4).
promotes follow-up in patients with low or inter- To refine PCa diagnosis, new tools such as risk
mediate PCa risk (43) (Figs 3–5). calculators or PSA density have been introduced.
Another promising approach to improve selec- Risk calculators are multivariable-based tools for
tion of at-risk patients is the use of NLP tools in the prediction of individual PCa and csPCa risk
primary care that can automatically detect risk- and help clarify the need for biopsy. They calcu-
factor data in the patients’ clinical record and cre- late individual risk for each patient on the basis
ate alerts to enroll these patients in an early detec- of populational big data after histologic verifica-
tion program. Similarly, they can classify patients tion. Risk calculators present greater discrimina-
into risk groups automatically after obtaining PSA tion power in PCa prediction than do PSA levels
levels and DRE findings (23). and DRE alone, with concordance between 70%
and 77% (7). There are many risk calculators
Clinical and Analytical Diagnosis for PCa, although those with the highest csPCa
of PCa prediction are the European Randomized Study
of Screening for Prostate Cancer (ERSPC) risk
Current Status calculator, which was created from a European
The two main problems are the accurate dis- patient population, and the Prostate Cancer Pre-
tinction between patients who should undergo vention Trial (PCPT) risk calculator, which was
1682  October Special Issue 2021 radiographics.rsna.org

Figure 3.  Early detection strategy


algorithm for csPCa with risk strati-
fication by using mpMRI according
to current clinical guidelines.

Figure 4. Proposed
strategy algorithm in
early csPCa detection
that is based on risk
stratification by using
a risk calculator (RC)
and PSA density (PSAd)
before mpMRI. This
approach could avoid
unnecessary mpMRI in
low-risk patients (or-
ange pathway).

created from a North American study, with the predictor is gaining interest since it is easy to ac-
ERSPC risk calculator seemingly having greater quire, and there is growing evidence demonstrat-
PCa prediction (7). ing the relationship between elevated PSA density
Another important tool in PCa diagnosis is values and csPCa. PSA density values greater than
PSA density (total PSA divided by prostatic 0.10 ng/mL2 are associated with a detection rate of
volume), which is a csPCa predictive marker that, 77% of PCa with a Gleason score greater than or
although not currently included in clinical guide- equal to 7 (5).
lines, is used in active surveillance and unofficially
used in mpMRI interpretations at some centers. It Future of Clinical PCa Diagnosis
can now be calculated by AI tools and included in mpMRI is a tool to help improve risk strati-
the patient’s report. Its usefulness as a csPCa risk fication in patients with suspected csPCa, but
RG  •  Volume 41  Number 6 Alcalá Mata et al  1683

Figure 5.  Proposed strategy flowchart for early csPCa detection based on risk stratification by using a risk calculator
(RC) and PSA density (PSAd) before and after mpMRI. This approach could avoid unnecessary mpMRI in low-risk pa-
tients and unnecessary biopsies in patients with PI-RADS 3 lesions with low risk of csPCa and, more importantly, could
avoid underdiagnosing csPCa in patients with negative MRI findings at medium or high risk of csPCA (orange pathway).

Table 4: Problems Arising from mpMRI Pathway

Problem Prevalence
Low accessibility and high cost of mpMRI NA
Overdiagnosis of ncsPCa in PI-RADS 3 lesions Up to 92%
csPCa undiagnosed because of false-negative findings at 4%–8%
mpMRI (PI-RADS ≤2)
False-positive findings at mpMRI (acute prostatitis, 67% by PI-RADS 3, 31% by PI-RADS 4, and 6% with
atypical hyperplastic nodules in TZ, bleeding areas in PI-RADS 5
PZ)
Increased complications (hematuria and hematosper- Increase in complication risk of 1.11% per core taken
mia) after TRUS-Bx plus MRI-TBx
mpMRI reading requires experienced radiologists and NA
has a steep learning curve
Sources.—References 8,11, and 43–45.
Note.—NA = not applicable, PZ = peripheral zone, TZ = transitional zone.

unnecessary mpMRI should be avoided. Recent may lead to unnecessary biopsy (Fig 4). After
publications advocate the use of risk calcula- mpMRI, risk calculators could avoid unneces-
tors in this regard. European guidelines recom- sary biopsies in patients with indeterminate-risk
mended the use of risk calculators in patients lesions (PI-RADS 3), who could benefit from
with PSA values between 2 and 10 ng/mL follow-up (Fig 5). More importantly, they could
(39,43). Risk calculators can act as gatekeepers help identify high-risk patients with negative
in improving patient stratification before and mpMRI findings (PI-RADS ≤ 2), in which a
after mpMRI. Before mpMRI, risk calculators TRUS-Bx should be considered, helping avoid
could identify low-risk patients, helping avoid underdiagnosis of csPCa in patients with false-
probable mpMRI false-positive findings that negative findings at mpMRI (10,11,44,46).
1684  October Special Issue 2021 radiographics.rsna.org

Figure 6.  mpMRI and subsequent MRI-TBx in a 64-year-old White man with no previous
biopsy, a PSA level of 5.46 ng/mL, normal DRE findings, normal transrectal US findings, a
prostatic volume of 42 mL, and low risk according to risk calculators. (A) Axial T2-weighted
multiparametric MR image shows a hypointense 10.8-mm lesion with partially occult borders
that was detected in the right anterior midglandular transitional zone (a region of interest
[ROI]) with a PI-RADS score of 3. (B, C) Axial diffusion-weighted MR image (B) depicts marked
lesion diffusion restriction, demonstrating high signal intensity with a b value of 1500 sec/mm2
(an ROI) and with low signal intensity on a corresponding ADC map (an ROI) (C), which corre-
sponded to a PI-RADS score of 4. (D) MRI-TBx of the lesion was performed with MRI/US fusion
biopsy. The biopsy needle was in the center of the lesion. (E) Low-power photomicrograph of
a core-needle biopsy sample shows a Gleason 3+3 adenocarcinoma in the main core (oval).
(Hematoxylin-eosin [H-E] stain; original magnification, 31.) (F) The algorithm highlights zones
of a Gleason 3 pattern while ghosting out the benign surrounding tissue (31). Histologic im-
ages obtained with Paige Prostate (Paige.AI).

Interestingly, PSA density combined with


mpMRI findings has improved PI-RADS nega-
tive predictive value (NPV) and has the promise
to help reduce unnecessary biopsies (10,11,44,47).
Patients at low risk of csPCa as determined by
both variables could benefit from clinical follow-
up, thus avoiding probable mpMRI false-positive of biopsies would be avoided and overdiagnosis of
findings that would lead to unnecessary biopsy ncsPCa reduced by 45% (Fig 6).
(11). According to Boesen et al (48), restricting Both risk calculators and PSA density should
biopsy to patients with a PSA density greater than also be used routinely after mpMRI when the re-
or equal to 0.15 ng/mL2 and a PI-RADS MRI sult is a PI-RADS score that is less than or equal
score greater than or equal to 4 is the best strategy, to 3, potentially avoiding unnecessary biopsies
as only 5% of csPCas would be missed, but 41% in this group who would benefit from multidis-
RG  •  Volume 41  Number 6 Alcalá Mata et al  1685

Table 5: mpMRI Protocol

Sequence When to Perform Sequence Parameters


T2WI Always At least 2D RARE T2WI in two planes
(axial and sagittal or coronal) and 3D
RARE T2WI
DWI Always At least two b values (one low: 0 or 50
sec/mm2; one intermediate: 800 or
1000 sec/mm2) and calculated or ac-
quired high b values ≥1400 sec/mm2
ADC map Always Created only from a low b value (50 sec/
mm2 preferably) and a b value of 800
or 1000 sec/mm2
DCE-MRI Negative prior biopsy results and unexplained high PSA Same plane as in DWI
In active surveillance or have undergone treatment Temporal resolution ≤15 sec
Previous biparametric MRI and persisting PCa suspicion
Negative prior biopsy results and strong family history or
genetic predisposition
Metallic prostheses present
Source.—Reference 9.
Note.—DCE = dynamic contrast-enhanced, RARE = rapid acquisition with relaxation enhancement, T2WI =
T2-weighted imaging, 2D = two-dimensional.

ciplinary team follow-up (10,11,44). The use of mpMRI allows initial tumor staging in the case of
risk calculators and PSA density is potentially positive biopsy results (9).
more important in identifying high-risk patients
with negative mpMRI findings (PI-RADS ≤ 2), in Future of mpMRI in PCa Diagnosis
which a TRUS-Bx could be considered, avoiding A very promising field is the use of AI in the
csPCa underdiagnosis in that group due to false- development of diagnostic probabilistic maps,
negative mpMRI findings (11). This approach which can identify lesions at mpMRI and es-
can be considered a refinement of the combined tablish a correlation with histologic grading and
biopsy pathway of PCa diagnosis. clinical outcome (20). The shortage of experi-
Novel “omics” biomarkers may help improve enced radiologists can potentially be alleviated
PCa diagnosis and risk stratification. Here, AI has by AI algorithms, reducing time required by the
a role in the identification and validation of new radiologist and minimizing false-negative find-
biomarkers, as machine learning tools accelerate ings that are associated with more inexperienced
the identification of new applications (ie, high readers. Alternatively, these solutions can be used
Ki67 expression may be used to select candidates as an independent second reader to help improve
for focal therapy) (6,49). interreader reproducibility.
Radiomics has been used in the development
Multiparametric MRI of CAD tools for PCa detection and PI-RADS
classification of mpMRI lesions. Furthermore,
Current Situation these solutions can provide information on tu-
At present, mpMRI should be performed in men mor heterogeneity that has histologic correlation
who are suspected of having csPCa in both bi- by examining edge characteristics and texture
opsy-naive patients and after TRUS-Bx, accord- analysis (52).
ing to PI-RADS version 2.1 recommendation CAD tools typically use texture analysis and
(Table 5) (9,50). mpMRI improves the detection radiomics with one or several MRI sequences,
and location of PCa with an ISUP grade greater analyzing lesion borders and signal intensity more
than or equal to 2, with a sensitivity of 91% and commonly with T2-weighted sequences or diffu-
specificity of 37%. For tumors with an ISUP sion-weighted imaging (DWI) and corresponding
grade greater than or equal to 3, the sensitivity apparent diffusion coefficient (ADC) maps. They
of mpMRI is 97% and specificity is 35%, with a are built by using supervised learning, in which
low detection rate of ISUP grade 1 PCa (< 30%) experienced radiologists label the lesions and
(51). Also, mpMRI is necessary to perform MRI- score them according to PI-RADS. These data
TBx, which significantly increases the detection can be correlated with histologic results, which are
rate of csPCa in comparison with TRUS-Bx (38). then used in the AI training process. Later, new
1686  October Special Issue 2021 radiographics.rsna.org

Figure 7.  CAD AI in prostatic MRI (MR


Prostate AI included in syngo.via Fron-
tier; Siemens Healthineers). The AI tool
assists with mpMRI reading and prostate
gland segmentation, including the pe-
ripheral zone and transitional zone, and
the detection and classification of lesions
with a PI-RADS score of 3 or greater by
using T2-weighted, DWI, and postcon-
trast dynamic contrast-enhanced (DCE)
sequences. The AI tool creates both rigid
and elastic registers of the sequences to
generate synthetic or calculated b values
of 1500 sec/mm2 and ADC maps from
original b values of 50 and 1000 sec/
mm2. (A) On these images, the automatic
glandular segmentation is done, which
is then validated or manually edited by
the radiologist. (B) After validation, the
prostatic volume is calculated, and detec-
tion and analysis of suspicious lesions are
performed. Information offered includes
volume, greater diameter, average ADC
value, 10th percentile of the ADC map,
average Ktrans value, and lesion location.
(C) An automatic PI-RADS scoring in the
three analyzed sequences is then done,
and a global PI-RADS scoring is gener-
ated. (D) The radiologist validates this
classification, and a report is automati-
cally generated with the data listed in the
legend for B, which includes key images
highlighting the lesion in the quadrant
map. Both gland segmentation and suspi-
cious lesions can be exported to the hard-
ware to perform MRI-TBx.

unlabeled training data may be presented to the (between 75%–80%) by CAD tools in clinical
algorithm (Fig 7) (20,52). trials, which perform similarly to radiologists
However, there are two main disadvantages to (19,26). In a multicenter multireader study, CAD
the development of these AI tools. First, the large sensitivity was similar to that of a radiologist but
amount of labeled data necessary for adequate was associated with a reading time reduction.
AI training requires notable input and time from Moderately experienced readers required 6.3
experienced radiologists (25,53). Second, the minutes, but this time was reduced to 4.4 min-
false-positive rate of these tools is up to 50%, utes when CAD assisted (27). For experienced
which requires radiologist input during reading readers, this was 3.5 minutes without CAD as-
to validate or reject the lesions presented by AI sistance, which was reduced to 2.7 minutes with
(25,53). Another hurdle to perform quantitative CAD assistance. The sensitivity of index lesion
analysis of mpMRI is the heterogeneity of data detection (PI-RADS ≥3) was 78%, but where
that is due to differences in acquisition protocols, CAD stood out was in the increase in sensitivity
magnet field strengths, and MRI system ven- in the detection of transition zone lesions by less
dors. To minimize these variations, some authors experienced readers, reaching 83.8% compared
propose the normalization of data to ensure that to 66.9% without CAD (Fig E1).
intensity values within the prostate gland are Although in most series, the sensitivity for de-
consistent among different studies (54). tection and classification of PI-RADS lesions with
There is growing evidence, including from a score greater than or equal to 3 and the specific-
meta-analyses, describing significant robustness ity for PI-RADS lesions with a score greater than
in the mpMRI detection rate of suspicious lesions or equal to 4 is similar to that of radiologists, posi-
RG  •  Volume 41  Number 6 Alcalá Mata et al  1687

Figure 8.  False-positive findings in CAD-aided interpretation of mpMRI in a 67-year-old man with a PSA level of 12.1 ng/mL (same
patient as in Figure E2). (A–D) The algorithm (MR Prostate AI) detected a small hypointense bandlike lesion (orange region of interest
[ROI]) at T2-weighted MRI in the apical posterolateral left peripheral zone (A), with high signal intensity at DWI (B), intermediate
ADC values (C), and without early enhancement at dynamic contrast-enhanced (DCE) MRI (D), suggesting an inflammatory or
benign nature. (E) The lesion was assigned a PI-RADS 2 score by a radiologist but was misclassified as PI-RADS 3 by the algorithm.

tive predictive value is low because of false-positive sional MRI are associated with higher-grade PCa,
findings, which would increase the number of making this a promising tool in the prediction of
unnecessary biopsies (25). As algorithms continue aggressive PCa (56,57) (Fig 9).
to improve, the false-positive rate should continue In summary, the literature shows how AI
to decrease (53) (Fig 8) (Fig E2). tools can accurately identify suspicious prostatic
Recent studies point to the association be- lesions at mpMRI and can reasonably estimate
tween MRI signal intensity with T2-weighted PI-RADS and Gleason scores (27,28,58–60).
sequences, DWI characteristics, and the density Improvements are expected once AI algorithms
of the tissue composition, which is derived from can be trained with the relevant digital pathologic
the automatic segmentation of stromal tissue, slides, given the spatial correlation between radi-
epithelium, epithelial nuclei, and lumen that per- cal prostatectomy specimens and mpMRI (61).
mit the creation of radiopathomic maps. A study
has shown that increased epithelial-map density Biopsy
and decreased lumen-mapping density were
associated with high-grade PCa (55). Another Current Status
study concludes that increased epithelial volume Transrectal or transperineal TRUS-Bx is the
and decreased stromal and lumen map volumes mainstay of PCa diagnosis. Given the low sensi-
that are determined by using hybrid multidimen- tivity of this approach in csPCa detection and the
1688  October Special Issue 2021 radiographics.rsna.org

Figure 9.  Radiopathomic maps from mpMRI that were generated with an AI-based radiomic tool (hybrid
multidimensional MRI) in a 56-year-old patient with Gleason 4+3 adenocarcinoma in the right peripheral zone.
(A–C) Stroma volume map (A), lumen volume map (B), and epithelium volume map (C) depict the malignant
lesion (arrow) as having reduced stroma and lumen volume (36.4% and 7.8%, respectively) and elevated epi-
thelium volume (55.8%) compared with surrounding benign tissue. (D) Axial turbo spin-echo T2-weighted MR
image. The hybrid multidimensional MRI map was correctly predicted as csPCa on the cancer map and was a
hypointense region in the corresponding turbo spin-echo T2-weighted MR image (arrow). (Case courtesy of
Aritrick Chatterjee, PhD, Department of Radiology, University of Chicago, Chicago, Ill.)

introduction of mpMRI in the diagnostic path- detection rate by 9.9%, compared with any
way, new biopsy methods have emerged, such as single-technique approach (63). However, in
MRI-TBx, ranging from cognitive biopsy (visu- patients with previous negative biopsy, com-
ally registered targeted biopsy), MRI/US fusion bined biopsy is not recommended (51). TRUS-
biopsy, to in-bore previous MRI-guided biopsy Bx is still important in patients with negative
(Figs 10, 11). Therefore, the role of TRUS-Bx is MRI findings and high csPCa suspicion, and in
under review. MRI/US fusion biopsy and MRI- patients with PI-RADS score 3 lesions with-
guided biopsy have demonstrated an increase of out previous biopsy. However, a meta-analysis
30% and of 38%, respectively, in csPCa detec- involving approximately 14 000 patients con-
tion compared with TRUS-Bx, explaining their cluded that excluding TRUS-Bx in these pa-
increasing adoption (62). Also, the use of MRI- tients was associated with a decrease in ncsPCa
TBx seems more consistent with the Gleason diagnosis without affecting the diagnosis of
score in prostatectomy specimens in lesions with csPCa (64). Furthermore, it has been suggested
a PI-RADS score greater than or equal to 3 than that PI-RADS score 4 lesions should have only
in TRUS-Bx (63). However, some controversy MRI-TBx sampling of the target lesion and the
remains about the best biopsy route (transrectal penumbra area (prostatic tissue surrounding the
or transperineal), the optimal number of cores lesion), since in small lesions, the sampling error
per lesion, the best technical solution to perform rate can be high, and this ensures that the suspi-
MRI-TBx, and whether it is performed alone or cious area is biopsied. In large PI-RADS score
combined with TRUS-Bx (39). 5 lesions, only MRI-TBx of the lesion has been
At present, the recommended approach in recommended, but this is an evolving topic (Fig
the EAU and NCCN guidelines is to perform 12) (11).
TRUS-Bx plus MRI-TBx for lesions with a Another question is the difference in csPCa
PI-RADS score greater than or equal to 3 in detection between different MRI-TBx tech-
patients without previous biopsies, as perform- niques. In MRI/US fusion biopsy, in vivo US
ing both together seems to increase the PCa acquisitions are overlaid on previously segmented
RG  •  Volume 41  Number 6 Alcalá Mata et al  1689

Figure 10.  Biopsy types.


(A) TRUS-Bx is the usual biopsy
technique. To compensate for the
poor sensitivity of TRUS-Bx in the
detection of suspicious lesions, the
systematic TRUS-Bx uses sextant
sampling. Six cylinders are ob-
tained from the right peripheral
zone and six from the left, with no
routine sampling of the transitional
zone. (B) Cognitive biopsy is per-
formed after mpMRI with positive
findings. In this case, the suspi-
cious lesion was located in the left
anterior midglandular transitional
zone (arrow). The operator then
performed a transrectal US-guided
biopsy of the area, in which the
mpMRI lesion was located, al-
though in this case, the lesion was
not visualized at US. This tech-
nique is indicated when the equip-
ment to perform MRI-TBx is not
available. (C) In MRI/US fusion bi-
opsy, the gland and suspected le-
sion are first segmented at mpMRI
by using specific software, and the
target lesion is then biopsied by
using hardware that permits the
fusion of in vivo US and previous
mpMRI acquisitions. (MRI/US fu-
sion biopsy performed with Uro-
Nav fusion biopsy system; Philips
Healthcare.) (D) Cognitive biopsy.
The operator then performed a
transrectal US-guided biopsy of
the area, in which the mpMRI le-
sion was located, although in this
case, the lesion was not visualized
at US. This technique is indicated
when the equipment to perform
MRI-TBx is not available.

MRI images of the targeted areas. Currently, dif- or lesions located in the apex or prostatic base,
ferent commercial MRI/US fusion biopsy systems which may be difficult to reach in patients with
use different biopsy routes and show different a large prostate (66,67). Also, the fact that these
technical characteristics and clinical results (65). biopsies are obtained in the same position as
One advantage of MRI/US fusion biopsy the original MRI examination should be associ-
over in-bore MRI-guided biopsy is that it can ated with a better histologic correlation. Thus,
be combined in a single procedure with TRUS- in patients with biopsied lesions with a Gleason
Bx. In addition, in-bore MRI-guided biopsy is a score of 3, but high suspicion at mpMRI with
biopsy technique that is of limited use at present a Gleason score of 4, an in-bore MRI-guided
because of its high cost and restricted avail- biopsy should be considered (67).
ability, but with promising results for certain
lesions. In-bore MRI-guided biopsy shows a Future of MRI-TBx
higher csCPa detection rate of 25% compared Although the analysis of MRI/US fusion bi-
with MRI/US fusion biopsy in patients with opsy systems is beyond the scope of this review,
previous biopsy, lesions smaller than 10 mm, there is room for improvement in the alignment
1690  October Special Issue 2021 radiographics.rsna.org

Figure 11.  In-bore MRI biopsy in a 64-year-old-man with a PSA level of 4.36 ng/mL and no
previous biopsy. (A–C) Axial T2-weighted MR image (A), axial diffusion-weighted MR image
with a b value of 1000 sec/mm2 (B), and axial ADC map obtained with mpMRI (C) reveal a left
peripheral zone–based medial-lateral lesion measuring 4 mm with a PI-RADS 4 score (arrow).
The needle position was represented in real time by the orange line overlaid from the needle
guide and showed the optimal path after the guide was remotely repositioned. (D, E) Axial
diffusion-weighted MR image (D) and sagittal T2-weighted image (E) show the needle track
and core that were obtained with a remote-controlled manipulator (Soteria Medical), which
are represented by the red line. Two cores were taken, revealing a Gleason 3+3 pattern in
both. The cores consisted of 50% and 70% tumor content, respectively. (Case courtesy of Joan
Carles Vilanova, MD, PhD, Clínica Girona, Girona, Spain.)

Figure 12.  Biopsy indication according to MRI results (blue track). New trends are currently emerging on the biopsy
type of choice in PI-RADS 4 lesions, in which the detection rate of csPCA with MRI-TBx due to saturation sampling of
the lesion and its penumbra zone would be similar to that of a combined biopsy (MRI-TBx and TRUS-Bx). In PI-RADS 5
lesions, given their large size, the existence of csPCA in other locations is unlikely, so sampling that lesion only would be
indicated, thus avoiding the complication risk associated with sampling other regions (orange track).

between both imaging modalities and in target tation in mpMRI, while considering the deform-
location, particularly of lesions located anteri- able prostate pseudocapsule. Other AI tools help
orly or very apically, where the prostatic gland speed up mpMRI reading by performing auto-
deformation during the procedure may compli- matic prostatic segmentation and detecting and
cate biopsy. classifying lesions with a PI-RADS score of 3 or
Prostate gland segmentation is a repetitive task higher. Both gland segmentation and suspicious
that is required for prostate volume measure- lesion data can be exported to the hardware to
ment, radiation therapy planning, and MRI-TBx. perform MRI-TBx. In MRI/US fusion biopsy, AI
AI-based solutions can label anatomic structures, may help improve prostatic segmentation in both
helping to perform an accurate prostate segmen- mpMRI and US, and more importantly during
RG  •  Volume 41  Number 6 Alcalá Mata et al  1691

Figure 13.  MRI/US fusion biopsy workflow with real-time 3D deformable image registration algorithms. MRI/US fusion biopsy re-
quires prior software-specific segmentation of the prostatic gland at T2-weighted imaging, which can be performed automatically as
in this example (A), and of the suspicious lesion (B), which is usually done manually or semi-automatically (DynaCAD; Philips Health-
care). These segmentations are transferred to the biopsy hardware, which is a US scanner with an electromagnetic field generator and
a navigation sensor mounted on the ultrasound translator. (Fig 13 continues on next page.)

image registration, whereby prostatic segmenta- Pathology


tion and targets in T2-weighted sequences are
fused with the US images in real time (64). The Current Status
use of real-time 3D deformable image registra- A necessary step for the integration between
tion algorithms can improve MRI/US fusion radiologic and pathologic characteristics is the
biopsy results (29,68) (Fig 13). digital transformation of pathology. Pathology
AI can also help PCa treatment. The coregis- relies on the glass slide as the primary diagnos-
tration of cancer lesions at mpMRI, as detected tic medium. Digital pathology consists of the
by a radiomics-based machine learning classi- creation of a digital replica of this glass slide,
fier together with CT, can help generate accu- called a whole-slide image, by using a scanning
rate radiation therapy plans (69). Also, cancer device. Only recently, technologic and regulatory
regions detected by a machine learning algo- changes have permitted the digitization of pathol-
rithm at mpMRI can be fused with real-time ogy. In 2017, the U.S. Food and Drug Admin-
US to be automatically displayed to surgeons, istration (FDA) permitted the use of the first
providing valuable information during robotic digital pathology solution for primary diagnosis
surgery (70). (71). The adoption of digital pathology is still in
1692  October Special Issue 2021 radiographics.rsna.org

Figure 13. (Continued from pre-


vious page.)  (C) The segmented
T2-weighted images and US im-
ages are fused together by using
an elastic and rigid registration
(UroNav fusion biopsy system; Phil-
ips Healthcare). The purple outline
is the segmented MRI prostatic
contour, which adapts to the de-
formity of the prostate during the
US examination (caused by the
transrectal probe and sedation)
and checks in real time its appropri-
ate positioning during the duration
of the examination. (D) The green
outline shows the contours of the
segmented lesion at MRI, display-
ing the biopsy needle position
in real time, and taking between
three and five cores per target.

its infancy, with few centers worldwide having better user experience compared with light mi-
reported their digital transformation. croscopy (72,30). Digital histologic images can
Core-needle biopsy evaluation is the main his- be viewed anywhere by using a computer instead
topathologic assessment in PCa and is one of the of a microscope, which makes them accessible
most frequent specimen types in surgical pathol- beyond the pathology laboratory. Also, digital
ogy. The histologic evaluation is typically done by pathology permits the application of AI tools to
using hematoxylin-eosin–stained tissue sections, histologic images to aid diagnosis, as it occurs
which are examined under conventional micros- in radiology. Therefore, digital pathology is the
copy, and involves the detection of cancer and its foundation of computational pathology and CAD
corresponding Gleason grading. This process is in pathology.
affected by several problems that could benefit Specific problem areas in PCa diagnosis that
from CAD. Increasing workloads and decreasing are likely to benefit from CAD are reducing
pathologist numbers worldwide put pressure on false-negative rates in histopathologic diagnosis
the existing workforce, as it is critical that diagnos- and better Gleason grading, as this is a subjective
tic quality standards are met. process with suboptimal concordance, particu-
larly when done by general pathologists instead
Future of Pathologic Diagnosis of PCa of specialists (73,74) (Fig 14). Clinical-grade AI
The advantages of a digital pathology workflow systems in PCa already exist and have demon-
are multiple, including improved efficiency and strated increased pathologist sensitivity and opti-
RG  •  Volume 41  Number 6 Alcalá Mata et al  1693

Figure 14.  False-negative histopathologic findings re-


vealed by AI in a 74-year-old man with a PSA level of 4.3
ng/mL. (A) Coronal T2-weighted MR image that was ana-
lyzed by MR Prostate AI shows a 13.7-mm lenticular lesion
located in the right anterior apical transitional zone, which
was hypointense at T2-weighted imaging. (B) Fusion of the
lesion heat map and the T2-weighted image shows how
the level of suspiciousness (LoS) calculated by the algorithm
is 85 on a scale from 1 to 100 and is therefore suggestive
of malignancy, with a PI-RADS score of 4. (C) Low-power
photomicrograph of a prostatic core-needle biopsy speci-
men shows abundant prostatic adenocarcinoma, acinar
type. The primary histologic diagnosis was Gleason score
3+3. (H-E stain; original magnification, 32.) (D) Tumor
probability heat map includes an algorithm that highlights
tumor areas according to their mathematical probability
of being malignant. One of the drawbacks of this feature
is that the tissue details are no longer visible. The quanti-
fication window (arrow) displays the tumor size and per-
centage of tumor present in the core, as calculated by AI.
(E, F) Gleason pattern 3 detection (E) and Gleason pat-
tern 4 (F) detection maps can highlight different tumor
areas according to the Gleason pattern. The tumor areas
are clearly visible now and stand out from the benign tissue
background. This facilitates visual assessment of the overall
Gleason score that is also calculated by AI and displayed
(arrow). In this case, AI indicated a Gleason score of 3+4,
which was discordant with the primary histologic diagnosis,
which was subsequently reviewed and updated. Histologic
images were obtained with Paige Prostate Clinical.

mal negative predictive value and have led to the Radiopathologic Integration and
reevaluation of lesions that were misdiagnosed Integrated Diagnostics
clinically as benign (32,21). This illustrates how
AI could complement pathologists in tasks that Current Status
are difficult and time consuming. Therefore, the The PCa diagnostic process has some weak links.
role of technology in the future of histopathologic False-positive findings in mpMRI interpretation,
PCa diagnosis seems warranted to help alleviate associated with prostatitis foci, extruded stromal
problems such as intra- and interreader variabil- nodules and hyperplastic nodules, or nodules with
ity and long reading times. hemorrhagic content in the transitional zone that
1694  October Special Issue 2021 radiographics.rsna.org

Figure 15.  An example of a tumor board decision support tool (AI-Pathway Companion; Siemens Healthineers) that reflects all
clinical data of the patient and the actions performed. This application also suggests possible management pathways according to
clinical guidelines.

are classified as PI-RADS lesions with a score of reducing confusion and distress in clinicians and
3 or greater, generate a multitude of equivocal patients, similar to what occurs in the treatment
biopsies. Erroneous segmentation of suspicious of breast cancer (31) (Fig E3). An appropri-
lesions may lead to biopsying the wrong location ate way to help speed up radiologists’ mpMRI
at MRI-TBx. Inadequate biopsy sampling may learning curve could be based on integrated
occur in lesions located in regions difficult to radiologic-pathologic diagnosis, so lesions with
reach transrectally with MRI/US fusion biopsy, intermediate PI-RADS scores or those situated
especially in patients with a large prostate, with in more complex locations could be discussed in
an outcome of false-negative findings or discor- a multidisciplinary setting to clarify diagnostic
dant radiologic-histopathologic results. disparities. Similarly, in disease with a PI-RADS
The biopsy processing in the pathology labora- score of 4 or greater with radiologic-pathologic
tory consists of fixing the obtained tissue core in disagreement, any diagnostic errors such as
formalin, which is then embedded in a paraffin mpMRI false-positive findings or biopsy sam-
block. This is a 3D structure, and sampling only pling errors could be resolved. Finally, it is im-
its most superficial levels may result in most of portant to ponder the type of biopsy, considering
the tissue remaining unsampled, and thus unex- that saturation biopsies of suspicious lesions with
amined by the pathologist. A biopsy result that mpMRI could reduce the number of procedures,
is discrepant with the radiologic findings should thus reducing the risk of complications.
prompt a thorough examination of the tissue Cancer centers are recommended to discuss all
remaining in the paraffin block. Finally, histologic their cases in multidisciplinary teams. However,
interpretation errors may result in cancer foci the time available for complex cases is reduced by
being either incorrectly graded or missed entirely, standard case management. To solve this problem
which could lead to underdiagnosis of mpMRI and facilitate the diagnosis, treatment, and follow-
lesions that are suggestive of csPCa. up of patients with PCa by multidisciplinary
teams, decision support tools are created, based
Future on, and validated by clinical guidelines or expert
A joint radiology-pathology tumor board would management (22). These tools facilitate find-
help detect and resolve these diagnostic inconsis- ing relevant data in the clinical history, includ-
tencies before a final joint report is issued, thus ing analytical values, previous studies, histologic
RG  •  Volume 41  Number 6 Alcalá Mata et al  1695

results, and treatments received. Currently, novel article: lectured for Siemens Healthineers, Philips, and Canon;
received royalties from Springer for book editing. Other activi-
NLP tools can locate relevant data in the patient ties: disclosed no relevant relationships.
file such as family history, PSA levels, mpMRI
findings, and pathologic reports and generate
alerts, as well as suggesting actions based on up- References
1. Ferlay J, Colombet M, Soerjomataram I, et al. Estimating the
dated clinical guidelines. Finally, these platforms global cancer incidence and mortality in 2018: GLOBOCAN
integrate data from different sources (ie, clinical sources and methods. Int J Cancer 2019;144(8):1941–1953.
history, laboratory, genomics, or medical images), 2. Fenton JJ, Weyrich MS, Durbin S, Liu Y, Bang H,
Melnikow J. Prostate-Specific Antigen-Based Screening
providing the opportunity to apply AI solutions for Prostate Cancer: Evidence Report and Systematic
that analyze these data and compare them with Review for the US Preventive Services Task Force. JAMA
populational big data to improve individual risk 2018;319(18):1914–1931.
3. Hayes JH, Barry MJ. Screening for prostate cancer with the
stratification and better predict patient outcomes, prostate-specific antigen test: a review of current evidence.
in a practical example of personalized and preci- JAMA 2014;311(11):1143–1149.
sion medicine (22,75) (Fig 15). 4. Barbieri CE, Bangma CH, Bjartell A, et al. The muta-
tional landscape of prostate cancer. Eur Urol 2013;64(4):
567–576.
Conclusion 5. Nordström T, Akre O, Aly M, Grönberg H, Eklund M.
PCa diagnosis is experiencing a revolution with Prostate-specific antigen (PSA) density in the diagnostic
algorithm of prostate cancer. Prostate Cancer Prostatic Dis
the introduction of mpMRI and MRI-TBx, 2018;21(1):57–63.
which is known as the MRI pathway. mpMRI is 6. Van Booven DJ, Kuchakulla M, Pai R, et al. A Systematic
essential in the detection and characterization of Review of Artificial Intelligence in Prostate Cancer. Res
Rep Urol 2021;13:31–39.
csPCa, and MRI-TBx is gaining an important 7. Foley RW, Maweni RM, Gorman L, et al. European
role in PCa diagnosis at the expense of the more Randomised Study of Screening for Prostate Cancer (ER-
common TRUS-Bx. However, TRUS-Bx still SPC) risk calculators significantly outperform the Prostate
Cancer Prevention Trial (PCPT) 2.0 in the prediction
plays an important role in PCa diagnosis when of prostate cancer: a multi-institutional study. BJU Int
combined with mpMRI, which constitutes the 2016;118(5):706–713.
so-called combined biopsy pathway. As such, 8. Rosenkrantz AB, Ayoola A, Hoffman D, et al. The Learn-
ing Curve in Prostate MRI Interpretation: Self-Directed
the introduction of these diagnostic approaches Learning Versus Continual Reader Feedback. AJR Am J
results in frequent and substantial changes in Roentgenol 2017;208(3):W92–W100.
PCa management clinical guidelines. Further 9. Turkbey B, Rosenkrantz AB, Haider MA, et al. Prostate
Imaging Reporting and Data System Version 2.1: 2019
refinement of these diagnostic pathways may be Update of Prostate Imaging Reporting and Data System
granted because of the integration of risk calcula- Version 2. Eur Urol 2019;76(3):340–351.
tors, analytical methods (ie, PSA density), and 10. Schoots IG, Padhani AR, Rouvière O, Barentsz JO, Richen-
berg J. Analysis of Magnetic Resonance Imaging-directed
genomic and other “omics” biomarkers. Biopsy Strategies for Changing the Paradigm of Prostate
AI is gaining an ever-increasing importance Cancer Diagnosis. Eur Urol Oncol 2020;3(1):32–41.
in PCa management. AI tools facilitate prostate 11. Schoots IG, Padhani AR. Delivering Clinical impacts of
the MRI diagnostic pathway in prostate cancer diagnosis.
gland segmentation, lesion detection, and classifi- Abdom Radiol (NY) 2020;45(12):4012–4022.
cation at mpMRI, limiting interreader variability 12. Epstein JI, Egevad L, Amin MB, et al. The 2014 Interna-
and mitigating the potential lack of expertise of tional Society of Urological Pathology (ISUP) Consensus
Conference on Gleason Grading of Prostatic Carcinoma:
less-experienced radiologists. Digital pathology is Definition of Grading Patterns and Proposal for a New
the basis for radiologic-pathologic integration and Grading System. Am J Surg Pathol 2016;40(2):244–252.
of CAD in pathology, helping improve efficiency, 13. Sehn JK. Prostate Cancer Pathology: Recent Updates and
Controversies. Mo Med 2018;115(2):151–155.
diagnostic accuracy, and concordance. The role of 14. Allsbrook WC Jr, Mangold KA, Johnson MH, et al. Interob-
machine learning tools is expected to grow in ra- server reproducibility of Gleason grading of prostatic carci-
diologic and pathologic PCa diagnosis, facilitating noma: urologic pathologists. Hum Pathol 2001;32(1):74–80.
15. Aihara M, Wheeler TM, Ohori M, Scardino PT. Heteroge-
radiologic-pathologic integration. Finally, the use neity of prostate cancer in radical prostatectomy specimens.
of decision support tools and AI-based analysis of Urology 1994;43(1):60–66; discussion 66–67.
fused data may be the next step in the PCa diag- 16. Pooli A, Johnson DC, Shirk J, et al. Predicting Pathological
Tumor Size in Prostate Cancer Based on Multiparametric
nostic pathway, as their use is expected to improve Prostate Magnetic Resonance Imaging and Preoperative
risk stratification and patient outcome. Findings. J Urol 2021;205(2):444–451.
17. Cohen MS, Hanley RS, Kurteva T, et al. Comparing the
Disclosures of Conflicts of Interest.—J.A.R. Activities related Gleason prostate biopsy and Gleason prostatectomy grading
to the present article: disclosed no relevant relationships. Activi- system: the Lahey Clinic Medical Center experience and an
ties not related to the present article: current employee of Paige. international meta-analysis. Eur Urol 2008;54(2):371–381.
AI; holds stock options with Paige.AI; former employee of 18. Park SH, Han K. Methodologic Guide for Evaluating
Philips. Other activities: disclosed no relevant relationships. Clinical Performance and Effect of Artificial Intelligence
R.T.G. Activities related to the present article: disclosed no rele- Technology for Medical Diagnosis and Prediction. Radiol-
vant relationships. Activities not related to the present article: con- ogy 2018;286(3):800–809.
sultant to Invivo and DxTx Medical. Other activities: disclosed 19. Cuocolo R, Cipullo MB, Stanzione A, et al. Machine
no relevant relationships. A.L. Activities related to the present learning for the identification of clinically significant
article: editorial board member of RadioGraphics (not involved prostate cancer on MRI: a meta-analysis. Eur Radiol
in the handling of this article). Activities not related to the present 2020;30(12):6877–6887.
1696  October Special Issue 2021 radiographics.rsna.org

20. Goldenberg SL, Nir G, Salcudean SE. A new era: artificial 41. Stephan C, Cammann H, Semjonow A, et al. Multicenter
intelligence and machine learning in prostate cancer. Nat evaluation of an artificial neural network to increase the
Rev Urol 2019;16(7):391–403. prostate cancer detection rate and reduce unnecessary
21. Raciti P, Sue J, Ceballos R, et al. Novel artificial intel- biopsies. Clin Chem 2002;48(8):1279–1287.
ligence system increases the detection of prostate cancer 42. Jemal A, Culp MB, Ma J, Islami F, Fedewa SA. Prostate
in whole slide images of core needle biopsies. Mod Pathol Cancer Incidence 5 Years After US Preventive Services
2020;33(10):2058–2066. Task Force Recommendations Against Screening. J Natl
22. van Wijk Y, Halilaj I, van Limbergen E, et al. Decision Sup- Cancer Inst 2021;113(1):64–71.
port Systems in Prostate Cancer Treatment: An Overview. 43. Van Poppel H, Hogenhout R, Albers P, van den Bergh
BioMed Res Int 2019;2019:4961768. RCN, Barentsz JO, Roobol MJ. Early Detection of Prostate
23. Martín Noguerol T, Paulano-Godino F, Martín-Valdivia Cancer in 2020 and Beyond: Facts and Recommendations
MT, Menias CO, Luna A. Strengths, Weaknesses, Op- for the European Union and the European Commission.
portunities, and Threats Analysis of Artificial Intelligence Eur Urol 2021;79(3):327–329.
and Machine Learning Applications in Radiology. J Am 44. Schoots IG, Padhani AR. Personalizing prostate cancer
Coll Radiol 2019;16(9 Pt B):1239–1247. diagnosis with multivariate risk prediction tools: how
24. Djavan B, Remzi M, Zlotta A, Seitz C, Snow P, Marberger should prostate MRI be incorporated? World J Urol
M. Novel artificial neural network for early detection of 2020;38(3):531–545.
prostate cancer. J Clin Oncol 2002;20(4):921–929. 45. Wegelin O, Exterkate L, van der Leest M, et al. Complica-
25. Schelb P, Kohl S, Radtke JP, et al. Classification of Cancer tions and Adverse Events of Three Magnetic Resonance
at Prostate MRI: Deep Learning versus Clinical PI-RADS Imaging-based Target Biopsy Techniques in the Diagnosis of
Assessment. Radiology 2019;293(3):607–617. Prostate Cancer Among Men with Prior Negative Biopsies:
26. Hambrock T, Vos PC, Hulsbergen-van de Kaa CA, Barentsz Results from the FUTURE Trial, a Multicentre Randomised
JO, Huisman HJ. Prostate cancer: computer-aided diagnosis Controlled Trial. Eur Urol Oncol 2019;2(6):617–624.
with multiparametric 3-T MR imaging--effect on observer 46. Radtke JP, Wiesenfarth M, Kesch C, et al. Combined
performance. Radiology 2013;266(2):521–530. Clinical Parameters and Multiparametric Magnetic Reso-
27. Gaur S, Lay N, Harmon SA, et al. Can computer-aided nance Imaging for Advanced Risk Modeling of Prostate
diagnosis assist in the identification of prostate cancer on Cancer-Patient-tailored Risk Stratification Can Reduce
prostate MRI? a multi-center, multi-reader investigation. Unnecessary Biopsies. Eur Urol 2017;72(6):888–896.
Oncotarget 2018;9(73):33804–33817. 47. Distler FA, Radtke JP, Bonekamp D, et al. The Value
28. Sunoqrot MRS, Nketiah GA, Selnæs KM, Bathen TF, of PSA Density in Combination with PI-RADS™ for
Elschot M. Automated reference tissue normalization of the Accuracy of Prostate Cancer Prediction. J Urol
T2-weighted MR images of the prostate using object rec- 2017;198(3):575–582.
ognition. MAGMA 2021;34(2):309–321. 48. Boesen L, Nørgaard N, Løgager V, et al. Prebiopsy Bipa-
29. Anas EMA, Mousavi P, Abolmaesumi P. A deep learn- rametric Magnetic Resonance Imaging Combined with
ing approach for real time prostate segmentation in Prostate-specific Antigen Density in Detecting and Ruling
freehand ultrasound guided biopsy. Med Image Anal out Gleason 7-10 Prostate Cancer in Biopsy-naïve Men.
2018;48:107–116. Eur Urol Oncol 2019;2(3):311–319.
30. Retamero JA, Aneiros-Fernandez J, Del Moral RG. Mi- 49. Green WJ, Ball G, Hulman G, et al. KI67 and DLX2
croscope? No, Thanks: User Experience With Complete predict increased risk of metastasis formation in pros-
Digital Pathology for Routine Diagnosis. Arch Pathol Lab tate cancer-a targeted molecular approach. Br J Cancer
Med 2020;144(6):672–673. 2016;115(2):236–242.
31. Tan MP, Ong EM, Amy D, Tot T. Integrating anatomy, 50. Padhani AR, Weinreb J, Rosenkrantz AB, Villeirs G, Turk-
radiology, pathology, and surgery: An alternative approach bey B, Barentsz J. Prostate Imaging-Reporting and Data
in resecting multifocal and multicentric breast carcinoma. System Steering Committee: PI-RADS v2 Status Update
Breast J 2017;23(6):663–669. and Future Directions. Eur Urol 2019;75(3):385–396.
32. da Silva LM, Pereira EM, Salles PG, et al. Independent 51. Drost FH, Osses D, Nieboer D, et al. Prostate Magnetic
real-world application of a clinical-grade automated prostate Resonance Imaging, with or Without Magnetic Resonance
cancer detection system. J Pathol 2021;254(2):147–158. Imaging-targeted Biopsy, and Systematic Biopsy for Detect-
33. Tătaru OS, Vartolomei MD, Rassweiler JJ, et al. Artificial ing Prostate Cancer: A Cochrane Systematic Review and
Intelligence and Machine Learning in Prostate Cancer Pa- Meta-analysis. Eur Urol 2020;77(1):78–94.
tient Management-Current Trends and Future Perspectives. 52. Toivonen J, Montoya Perez I, Movahedi P, et al. Radiomics
Diagnostics (Basel) 2021;11(2):354. and machine learning of multisequence multiparametric
34. Ilic D, Neuberger MM, Djulbegovic M, Dahm P. Screen- prostate MRI: Towards improved non-invasive prostate
ing for prostate cancer. Cochrane Database Syst Rev cancer characterization. PLoS One 2019;14(7):e0217702.
2013;(1):CD004720. 53. Padhani AR, Turkbey B. Detecting Prostate Cancer with
35. Sanchez-Ortiz RF, Troncoso P, Babaian RJ, Lloreta J, Deep Learning for MRI: A Small Step Forward. Radiology
Johnston DA, Pettaway CA. African-American men with 2019;293(3):618–619.
nonpalpable prostate cancer exhibit greater tumor volume 54. Isaksson LJ, Raimondi S, Botta F, et al. Effects of MRI im-
than matched white men. Cancer 2006;107(1):75–82. age normalization techniques in prostate cancer radiomics.
36. Page EC, Bancroft EK, Brook MN, et al. Interim Results Phys Med 2020;71:7–13.
from the IMPACT Study: Evidence for Prostate-specific 55. McGarry SD, Hurrell SL, Iczkowski KA, et al. Radio-
Antigen Screening in BRCA2 Mutation Carriers. Eur Urol pathomic Maps of Epithelium and Lumen Density Predict
2019;76(6):831–842. the Location of High-Grade Prostate Cancer. Int J Radiat
37. Mateo J, Seed G, Bertan C, et al. Genomics of lethal prostate Oncol Biol Phys 2018;101(5):1179–1187.
cancer at diagnosis and castration resistance. J Clin Invest 56. Chatterjee A, Bourne RM, Wang S, et al. Diagnosis
2020;130(4):1743–1751. of Prostate Cancer with Noninvasive Estimation of
38. Carlsson S, Castle EP, Catalona WJ, et al. NCCN Guide- Prostate Tissue Composition by Using Hybrid Multidi-
lines Version 2.2020 Prostate Cancer Early Detection. mensional MR Imaging: A Feasibility Study. Radiology
https://www.nccn.org/professionals/physician_gls/pdf/ 2018;287(3):864–873.
prostate_blocks.pdf. Published 2020. Accessed DATE. 57. Chatterjee A, Watson G, Myint E, Sved P, McEntee M,
39. Mottet N, van den Bergh RCN, Briers E, et al. EAU-EANM- Bourne R. Changes in Epithelium, Stroma, and Lumen
ESTRO-ESUR-SIOG Guidelines on Prostate Cancer-2020 Space Correlate More Strongly with Gleason Pattern and
Update. Part 1: Screening, Diagnosis, and Local Treatment Are Stronger Predictors of Prostate ADC Changes than
with Curative Intent. Eur Urol 2021;79(2):243–262. Cellularity Metrics. Radiology 2015;277(3):751–762.
40. Finne P, Finne R, Auvinen A, et al. Predicting the outcome 58. Moradi M, Salcudean SE, Chang SD, et al. Multiparametric
of prostate biopsy in screen-positive men by a multilayer MRI maps for detection and grading of dominant prostate
perceptron network. Urology 2000;56(3):418–422. tumors. J Magn Reson Imaging 2012;35(6):1403–1413.
RG  •  Volume 41  Number 6 Alcalá Mata et al  1697

59. Fehr D, Veeraraghavan H, Wibmer A, et al. Automatic 67. Gurgitano M, Ancona E, Maresca D, et al. In-bore MRI
classification of prostate cancer Gleason scores from mul- targeted biopsy. Acta Biomed 2020;91(10-S):e2020012.
tiparametric magnetic resonance images. Proc Natl Acad 68. Hu Y, Modat M, Gibson E, et al. Weakly-supervised convo-
Sci U S A 2015;112(46):E6265–E6273. lutional neural networks for multimodal image registration.
60. Antonelli M, Johnston EW, Dikaios N, et al. Machine Med Image Anal 2018;49:1–13.
learning classifiers can predict Gleason pattern 4 prostate 69. Shiradkar R, Podder TK, Algohary A, Viswanath S, Ellis
cancer with greater accuracy than experienced radiologists. RJ, Madabhushi A. Radiomics based targeted radiotherapy
Eur Radiol 2019;29(9):4754–4764 [Published correction planning (Rad-TRaP): a computational framework for
appears in Eur Radiol 2020;30(2):1295.]. prostate cancer treatment planning with MRI. Radiat Oncol
61. Harmon SA, Tuncer S, Sanford T, Choyke PL, Türkbey 2016;11(1):148.
B. Artificial intelligence at the intersection of pathology 70. Mohareri O, Ischia J, Black PC, et al. Intraoperative registered
and radiology in prostate cancer. Diagn Interv Radiol transrectal ultrasound guidance for robot-assisted laparo-
2019;25(3):183–188. scopic radical prostatectomy. J Urol 2015;193(1):302–312.
62. Siddiqui MM, Rais-Bahrami S, Turkbey B, et al. Comparison 71. FDA news release: FDA allows marketing of first whole slide
of MR/ultrasound fusion-guided biopsy with ultrasound- imaging system for digital pathology. https://www.fda.gov/
guided biopsy for the diagnosis of prostate cancer. JAMA newsevents/newsroom/pressannouncements/ucm552742.
2015;313(4):390–397. htm. Published April 12, 2017. Accessed January 18, 2021.
63. Ahdoot M, Wilbur AR, Reese SE, et al. MRI-Targeted, 72. Retamero JA, Aneiros-Fernandez J, Del Moral RG. Com-
Systematic, and Combined Biopsy for Prostate Cancer plete Digital Pathology for Routine Histopathology Diagnosis
Diagnosis. N Engl J Med 2020;382(10):917–928. in a Multicenter Hospital Network. Arch Pathol Lab Med
64. Goldberg H, Ahmad AE, Chandrasekar T, et al. Comparison 2020;144(2):221–228.
of Magnetic Resonance Imaging and Transrectal Ultrasound 73. Yang C, Humphrey PA. False-Negative Histopathologic
Informed Prostate Biopsy for Prostate Cancer Diagnosis in Diagnosis of Prostatic Adenocarcinoma. Arch Pathol Lab
Biopsy Naïve Men: A Systematic Review and Meta-Analysis. Med 2020;144(3):326–334.
J Urol 2020;203(6):1085–1093. 74. Steinberg DM, Sauvageot J, Piantadosi S, Epstein JI. Cor-
65. Manfredi M, Costa Moretti TB, Emberton M, Villers A, relation of prostate needle biopsy and radical prostatectomy
Valerio M. MRI/TRUS fusion software-based targeted Gleason grade in academic and community settings. Am J
biopsy: the new standard of care? Minerva Urol Nefrol Surg Pathol 1997;21(5):566–576.
2015;67(3):233–246. 75. Lee G, Singanamalli A, Wang H, et al. Supervised multi-view
66. Perrin A, Venderink W, Patak MA, et al. The utility of in- canonical correlation analysis (sMVCCA): integrating histo-
bore multiparametric magnetic resonance-guided biopsy logic and proteomic features for predicting recurrent prostate
in men with negative multiparametric magnetic resonance- cancer. IEEE Trans Med Imaging 2015;34(1):284–297.
ultrasound software-based fusion targeted biopsy. Urol
Oncol 2021;39(5):297.e9–297.e16.

TM
This journal-based SA-CME activity has been approved for AMA PRA Category 1 Credit . See rsna.org/learning-center-rg.

You might also like