Professional Documents
Culture Documents
Xin Yu1 , Bin Lou1 , Bibo Shi1 , David Winkel1,2 , Nacim Arrahmane1 , Mamadou Diallo1 , Tongbai Meng1 ,
Heinrich von Busch3 , Robert Grimm3 , Berthold Kiefer3 , Dorin Comaniciu1 , Ali Kamen1 ,
ProstateAI Clinical Collaborators∗
1
Digital Technology and Innovation Division, Siemens Healthineers, Princeton, NJ, USA
2
Universitätsspital Basel, Basel, Switzerland
3
Diagnostic Imaging, Siemens Healthineers, Erlangen, Germany
ABSTRACT the mp-MRI approach [3]. Many attempts have been made
to help radiologists in detecting and classifying PCa lesion
Prostate cancer (PCa) is the most prevalent and one of the
using mp-MRI. Litjens et al. [4] and Cao et al. [5] indicated
leading causes of cancer death among men. Multi-parametric
that computer-aided detection (CAD) system using either
MRI (mp-MRI) is a prominent diagnostic scan, which could
mp-MRI or bp-MRI scans can achieve comparable detection
help in avoiding unnecessary biopsies for men screened for
sensitivity to a radiologist. However, as compared to radiolo-
PCa. Artificial intelligence (AI) systems could help radiolo-
gists’ performance, the CAD systems usually have relatively
gists to be more accurate and consistent in diagnosing clini-
lower specificity. This can in turn lead to overdiagnosis or
cally significant cancer from mp-MRI scans. Lack of speci-
overtreatment, which should be avoided as much as possible.
ficity has been identified recently as one of weak points of
Therefore, false positive reduction (FPR) is considered as an
such assistance systems. In this paper, we propose a novel
essential part of any assistance system. Furthermore, there
false positive reduction network to be added to the overall de-
are significant intensity pattern similarities between cancer-
tection system to further analyze lesion candidates. The new
ous and benign tissues within prostate gland as expressed
network utilizes multiscale 2D image stacks of these candi-
in various contrasts of MRI scans. This together with size
dates to discriminate between true and false positive detec-
variations among cancerous lesions makes the FPR for PCa
tions. We trained and validated our network on a dataset with
lesions particularly challenging.
2170 cases from seven different institutions and tested it on
a separate independent dataset with 243 cases. With the pro- Inspired by recent articles published aiming at lung nod-
posed model, we achieved area under curve (AUC) of 0.876 ule FPR [6, 7, 8], we propose a novel approach for utilizing
on discriminating between true and false positive detected le- multiscale images to learn better image features to effectively
sions and improved the AUC from 0.825 to 0.867 on overall remove false positives in PCa detection. The FPR network
identification of clinically significant cases. utilizes the results of an up-stream detection network. These
Index Terms— Deep learning, prostate cancer, false pos- detected lesions could be seen as candidates that need to be
itive reduction, mp-MRI further analyzed and separated into true and false lesions. The
detection network uses Prostate Imaging Reporting and Data
System (PI-RADS) scores, which is the standard for report-
1. INTRODUCTION ing prostate MRI findings [9], as ground truth. In this system,
a lesion is regarded as clinically significant if its PI-RADS
Prostate Cancer (PCa) is one of the most prevalent cancers
score is equal or greater than 3 [10]. The strategies used in
in 2019 among males in the United States. It is estimated
our proposed FPR network are summarized as follows: (1)
that over 3.6 million men have a history of PCa and 174,650
utilizing 2.5D inputs to incorporate more out-of-plane contex-
cases will be newly diagnosed in 2019 [1]. In recent years,
tual information; (2) extracting image features from different
multi-parametric MRI (mp-MRI) has shown its utility as a
fields of view (i.e. multiscale images); (3) devising a fusion
non-invasive imaging tool for detection, localization, and
module to assign optimal weights for different scales based on
classification of PCa [2], and there is growing evidence that
their contributions to the final classification; and (4) applying
biparametric MRI (bp-MRI) protocols, consisting of only
a multi-task loss function to extract the most discriminative
T2-weighted (T2w) imaging and diffusion weighted imaging
set of features. The proposed model was validated on an in-
(DWI), offer similar diagnostic performance compared to
dependent dataset with 243 patients and showed a significant
∗A list of members and affiliations appears at the end of the paper. improvement in terms of removing false positive lesions.
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:23:30 UTC from IEEE Xplore. Restrictions apply.
2. PROPOSED FRAMEWORK resolution in the through-plane direction, we only imple-
mented this multiscale strategy in-plane. To fit various lesion
Our overall system for detecting clinically significant PCa le- sizes, for three different scales we used dimensions of 32×32,
sions has two stages: (1) PCa detection using a detection net- 48×48 and 64×64 respectively and resized them all to 48×48
work, and (2) FPR using the detection results from the up- as FPR network input.
stream detection network. In the first stage, we aim to achieve
a higher sensitivity whereas in the second stage aim to min-
2.3. FPR Network Architecture
imize the false detection rate with a minimal impact to the
sensitivity. The whole pipeline is depicted in Fig.1. The FPR network started with a feature encoding module
where three different cropped fields of view were fed into
2.1. Prostate Cancer Detection 3 groups of residual blocks independently before the fusion
step. Each group consisted of 3 consecutive residual blocks
The detection network had a UNet architecture with 2D resid- of basic architecture [12]. The filter number for the first block
ual blocks [11, 12]. The Res-UNet was designed to have 5 was 16, with the size doubling for the following blocks. After
down-sampling and 5 up-sampling residual blocks. For each feature concatenation, we added a Squeeze-and-Excitation
residual block, the “bottleneck” architecture [12] with a stack (SE) block [13] to allocate different weights to each channel.
of 3 layers was adopted. There were 16 filters in the first Another residual block was adopted after the SE block for
residual block with doubling filter sizes for every block. further feature fusion. The output was flattened using global
Our detection network was evaluated using 3D volumes. average pooling and ultimately fed into 2 fully connected
Each patient imaging study was passed through the detec- layers with a size of 128 and 1 respectively to achieve the
tion network separately and the output was a heatmap volume final classification. The network structure is shown in Fig.2.
where the lesion candidates or regions were designated with Images coming from different fields of view ended up having
non-zero values. The heatmap was then thresholded to gen- different weights and hence importance. For example, the
erate a set of 3D connected components. True positive (TP) smaller scale images best demonstrated local lesion features
detection was identified if the detection was within annotated whereas the larger scale images emphasized the difference
lesion boundary or less than 5mm away from the lesion cen- between the lesion and its surroundings. This parallel design
ter. Otherwise, the detection was considered as a false pos- enables the network to learn more detailed set of features
itive (FP). Lesions without matched detection were regarded from different scales. Adding SE block helped to adjust the
as false negatives (FNs). The detected components were used weight for each channel resulting in different contributions of
as inputs for the FPR network. TP and FP detections were re- scales to the final classification.
spectively used as positive and negative samples for training
down-stream FPR network. After computing the connected
component peaks, we cropped small patches from the origi- 2.4. Loss Function for FPR Network
nal set of scans, and used them as inputs for the training of We used binary cross-entropy loss (BCEL) as our main loss
FPR network. function to train the network. Our goal was to minimize the
FP while having a minimum impact to the overall detection
2.2. 2.5D Multiscale Data sensitivity. To achieve this, we used a weight of α to have
larger penalty on misclassified positive samples:
Our FPR network was a 2.5D network by design. For each
2D input slice of Is , we also used its neighboring slices of LBCE = αyi log(pi ) + (1 − yi ) log(1 − pi ) (1)
Is−1 , Is+1 as additional channels. For this reason, we referred
to it as 2.5D input. The advantage of the 2.5D input is that where pi ∈ [0, 1] is the predicted lesion probability and y
a) in the case of true positive lesions, we expect the pattern ∈ {0, 1} is the ground truth label. In addition to BCEL, cen-
extends to the neighboring slices to some extent and based ter loss (CL) [14] was employed as a secondary term to en-
on the consistency across images the network could identify hance the discriminative power between classes, while min-
them as true positives; and b) since the detection network is imizing the distances between samples from the same class.
operating only on 2D, there is a chance for the FPR network We assumed that the intra-class variation should be smaller
to focus on inconsistencies and lack of coherent signatures for TP samples as compared to FP ones, therefore we set
across the slices and, by taking the context into account, de- higher weights for true positive samples:
tect and eliminate false positives. !
m m
In additional to the contextual information, we used a 1 X X
LC = β yi kxi − c1 k22 + (1 − yi ) kxi − c0 k22 (2)
multiscale strategy to better capture the discriminative pat- 2 i=1 i=1
terns. This strategy was employed by providing several
images of different fields of view for the same region to where xi ∈ Rd is the deep feature and c0 , c1 ∈ Rd de-
the network. Since the bp-MRI images have relatively low note two class centers of deep features. The total loss is a
1356
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:23:30 UTC from IEEE Xplore. Restrictions apply.
Detection 32x32 48x48 64x64 TP detection
FP detection
Crop multiscale patches
Detection FPR
Mask Network Network
B-2000
ADC T2w
Heatmap Ground Truth
Fig. 1. Depiction of the overall processing pipeline consisting of Detection Network and FPR Network.
32x32x9
SE Block 3.2. Model Training
Res Blocks
FC layers
TP lesion?
We trained and fixed the detection network before conducting
48x48x9
FP lesion?
CE Loss
experiments on the FPR network. Four types of image con-
Res Blocks Res Block
deep trasts were used as inputs: T2w, DWI ADC, DWI B-2000 MR
feature
64x64x9 images and a binary prostate region mask providing anatom-
Res Blocks Center Loss ical information. The prostate mask was generated based on
T2w volume using a learning-based method as presented in
[16, 17, 18]. The overall model was trained for 200 epochs
Fig. 2. Architecture for the false positive reduction network. with lesion masks as training labels. Detection model selec-
tion was done based on the highest dice coefficient on the
validation set.
combination of BCEL and CL during the training, Ltotal = Experiments for the FPR network was conducted using
LBCE + λLC , where λ is a hyper-parameter that was tuned to the patches generated from the outputs of the detection net-
control the balance between two loss functions. work. 80% of the detections were used for training and the
other 20% for validation. The network was trained using
ADAM as the optimizer for 100 epochs, with L2 regulariza-
3. EXPERIMENTAL RESULTS tion of 10−4 . Rotation range [-45°, 45°], shift range [-5, 5],
and vertical flips were adopted for data augmentation. Each
3.1. MRI Data and Pre-processing input sample had 9 channels including all the sequences from
the previous and next slices while the ground truth label was
Datasets from seven institutions with 2170 cases in total only from the middle slices. The number of samples from
were used for the analysis. 1736 cases were used for train- positive and negative classes were balanced within each batch
ing (80%) and 434 cases for validation (20%). We used during the training. In our experiments, we set the dimension
another 243 cases from ProstateX challenge public dataset reduction ratio to 8 when SE block was included in the model.
[4] as the test dataset to evaluate the performance of various The weight for CL was set to λ = 0.2, and weights of the pos-
models. All images and corresponding clinical reports were itive samples in BCEL and CL computation were assigned to
carefully reviewed by a radiologist with four years of experi- α = 3 and β = 2 respectively.
ence in radiology and subspecialty training in prostate MRI
examinations. Lesion contours were manually re-annotated
3.3. FPR Results
in 3D using an internally developed annotation tool based
on the original clinical reports. All images were registered Our baseline was a network with 3 residual blocks followed
to the T2w images and resampled to a voxel spacing of by 2 fully connected layers with size 256 and 1 respectively.
0.5mm×0.5mm×3mm, with an image size of 240×240×30 Inputs for the baseline were 2D data without 2.5D contextual
[15]. T2w images were linearly normalized to [0, 1] using or multiscale images.
0.05 and 99.5 percentiles of the image’s intensity histogram We conducted an ablation study to verify the impact of
as the lower and upper thresholds. Since the actual ADC 2.5D data, multiscale, SE block and center loss. Area under
value is highly relevant to clinical significance of lesions [9], curve (AUC) was evaluated on 2D sample level, 3D compo-
we normalized ADC intensity [0, 3000] to [0, 1] by a constant nent level and patient case level. The results are shown in
value. The DWI B-2000 images were first normalized to the TABLE 1. Component level suspicion scores were calculated
the median intensity in the prostate region of the correspond- by averaging the suspicion score of each 2D slice within the
ing DWI B-50 images, and then normalized by a constant stack. Component level AUC reflects the overall FPR net-
value to map the range of intensities into [0, 1]. work performance on discriminating true and false positive
1357
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:23:30 UTC from IEEE Xplore. Restrictions apply.
(a) Lesion level FROC (b) Case level ROC
1.0 1.0
Table 1. Evaluation of the effectiveness of 2.5D multiscale
data, SE block, center loss and a comparison with other mul- 0.9
0.8
Sensi!vity
Sensi!vity
Settings Samples Components Cases 0.6
1358
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:23:30 UTC from IEEE Xplore. Restrictions apply.
cer treatment and survivorship statistics, 2019,” CA: A Cancer [14] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao, “A
Journal for Clinicians, vol. 69, no. 5, pp. 363–385, 2019. discriminative feature learning approach for deep face recog-
[2] Romaric Loffroy, Olivier Chevallier, Morgan Moulin, Sylvain nition,” in European conference on computer vision. Springer,
Favelier, Pierre-Yves Genson, Pierre Pottecher, Gilles Cre- 2016, pp. 499–515.
hange, Alexandre Cochet, and Luc Cormier, “Current role of [15] Atilla P Kiraly, Clement Abi Nader, Ahmet Tuysuzoglu,
multiparametric magnetic resonance imaging for prostate can- Robert Grimm, Berthold Kiefer, Noha El-Zehiry, and Ali Ka-
cer,” Quantitative imaging in medicine and surgery, vol. 5, no. men, “Deep convolutional encoder-decoders for prostate can-
5, pp. 754, 2015. cer detection and classification,” in International Conference
[3] Christiane K Kuhl, Robin Bruhn, Nils Krämer, Sven Nebelung, on Medical Image Computing and Computer-Assisted Inter-
Axel Heidenreich, and Simone Schrading, “Abbreviated bi- vention. Springer, 2017, pp. 489–497.
parametric prostate mr imaging in men with elevated prostate- [16] Dong Yang, Daguang Xu, S Kevin Zhou, Bogdan Georgescu,
specific antigen,” Radiology, vol. 285, no. 2, pp. 493–505, Mingqing Chen, Sasa Grbic, Dimitris Metaxas, and Dorin Co-
2017. maniciu, “Automatic liver segmentation using an adversar-
[4] Geert Litjens, Oscar Debats, Jelle Barentsz, Nico Karssemei- ial image-to-image network,” in International Conference on
jer, and Henkjan Huisman, “Computer-aided detection of Medical Image Computing and Computer-Assisted Interven-
prostate cancer in mri,” IEEE transactions on medical imaging, tion. Springer, 2017, pp. 507–515.
vol. 33, no. 5, pp. 1083–1092, 2014. [17] Haozhe Jia, Yong Xia, Yang Song, Donghao Zhang, Heng
[5] Ruiming Cao, Amirhossein Mohammadian Bajgiran, Huang, Yanning Zhang, and Weidong Cai, “3d apa-net:
Sohrab Afshari Mirak, Sepideh Shakeri, Xinran Zhong, 3d adversarial pyramid anisotropic convolutional network for
Dieter Enzmann, Steven Raman, and Kyunghyun Sung, “Joint prostate segmentation in mr images,” IEEE Transactions on
prostate cancer detection and gleason score prediction in Medical Imaging, pp. 1–1, 2019.
mp-mri via focalnet,” IEEE transactions on medical imaging, [18] Donghao Zhang, Yang Song, Dongnan Liu, Haozhe Jia, Siqi
2019. Liu, Yong Xia, Heng Huang, and Weidong Cai, “Panoptic
[6] Bum-Chae Kim, Jee Seok Yoon, Jun-Sik Choi, and Heung-Il segmentation with an end-to-end cell r-cnn for pathology im-
Suk, “Multi-scale gradual integration cnn for false positive age analysis,” in International Conference on Medical Im-
reduction in pulmonary nodule detection,” Neural Networks, age Computing and Computer-Assisted Intervention. Springer,
vol. 115, pp. 1–10, 2019. 2018, pp. 237–244.
[7] Qi Dou, Hao Chen, Lequan Yu, Jing Qin, and Pheng-Ann
Heng, “Multilevel contextual 3-d cnns for false positive reduc-
tion in pulmonary nodule detection,” IEEE Transactions on
Biomedical Engineering, vol. 64, no. 7, pp. 1558–1567, 2016.
[8] Zhancheng Zhang, Xinyi Li, Qingjun You, and Xiaoqing Luo,
“Multicontext 3d residual cnn for false positive reduction of
pulmonary nodule detection,” International Journal of Imag-
ing Systems and Technology, vol. 29, no. 1, pp. 42–49, 2019.
[9] Andrei S Purysko, Andrew B Rosenkrantz, Jelle O Barentsz,
Jeffrey C Weinreb, and Katarzyna J Macura, “Pi-rads version
2: a pictorial update,” Radiographics, vol. 36, no. 5, pp. 1354–
1372, 2016.
[10] Jeffrey C Weinreb, Jelle O Barentsz, Peter L Choyke, Fran-
cois Cornud, Masoom A Haider, Katarzyna J Macura, Daniel
Margolis, Mitchell D Schnall, Faina Shtern, Clare M Tempany,
et al., “Pi-rads prostate imaging–reporting and data system:
2015, version 2,” European urology, vol. 69, no. 1, pp. 16–40,
2016.
[11] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net:
Convolutional networks for biomedical image segmentation,”
in International Conference on Medical image computing and
computer-assisted intervention. Springer, 2015, pp. 234–241.
[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun,
“Deep residual learning for image recognition,” IEEE Confer-
ence on Computer Vision and Pattern Recognition (CVPR), pp.
770–778, 2015.
[13] Jie Hu, Li Shen, and Gang Sun, “Squeeze-and-excitation net-
works,” in Proceedings of the IEEE conference on computer
vision and pattern recognition, 2018, pp. 7132–7141.
1359
Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on May 27,2020 at 11:23:30 UTC from IEEE Xplore. Restrictions apply.