Professional Documents
Culture Documents
Diagnostic Use of Facial Image Analysis Software in Endocrine and Genetic Disorders: Review, Current Results and Future Perspectives
Diagnostic Use of Facial Image Analysis Software in Endocrine and Genetic Disorders: Review, Current Results and Future Perspectives
Abstract
Cushing’s syndrome (CS) and acromegaly are endocrine diseases that are currently diagnosed with a delay of several years
European Journal of Endocrinology
from disease onset. Novel diagnostic approaches and increased awareness among physicians are needed. Face classification
technology has recently been introduced as a promising diagnostic tool for CS and acromegaly in pilot studies. It has also
been used to classify various genetic syndromes using regular facial photographs. The authors provide a basic explanation of
the technology, review available literature regarding its use in a medical setting, and discuss possible future developments.
The method the authors have employed in previous studies uses standardized frontal and profile facial photographs for
classification. Image analysis is based on applying mathematical functions evaluating geometry and image texture to a grid
of nodes semi-automatically placed on relevant facial structures, yielding a binary classification result. Ongoing research
focuses on improving diagnostic algorithms of this method and bringing it closer to clinical use. Regarding future
perspectives, the authors propose an online interface that facilitates submission of patient data for analysis and retrieval
of results as a possible model for clinical application.
European Journal of
Endocrinology
(2015) 173, M39–M44
Introduction
Cushing’s syndrome (CS) and acromegaly are endocrine hypothetically be used for early diagnosis. Recent studies
diseases that manifest with metabolic complications and have found that CS is currently diagnosed with a latency
typical changes to the appearance, especially charac- of about 2–6 years (1, 2). This may partly be due to an
teristic changes to facial features that are particular to overlap of clinical features with the more common
either disease. These pathognomonic signs can be metabolic syndrome. Acromegaly is still diagnosed with
detected by face classification software and could a delay of about 6 years, most likely due to the slow
www.eje-online.org Ñ 2015 European Society of Endocrinology This article is adapted from work presented at IMPROCUSH-1, 12–14 October 2014.
DOI: 10.1530/EJE-15-0429 Printed in Great Britain The meeting was supported by the European Science Foundation, Deutsche
Published by Bioscientifica Ltd. Forschungsgemeinschaft, Carl Friedrich von Siemens Stiftung, European Neuroendo-
crine Association and the DeutscheDownloaded
Gesellschaftfrom Bioscientifica.comThe
für Endokrinologie. at 03/19/2020
opinions or03:47:08AM
via free access
views expressed in this article are those of the authors, and do not necessarily reflect
the opinions or recommendations of the supporters of the symposium.
Special Section Review R P Kosilek and others Review of facial image analysis 173:4 M40
in diagnostics
www.eje-online.org
Downloaded from Bioscientifica.com at 03/19/2020 03:47:08AM
via free access
Special Section Review R P Kosilek and others Review of facial image analysis 173:4 M41
in diagnostics
subjects (24 patients, 25 controls) were included. The total Table 1 Results from Schneider et al., JCEM 2011. Classi-
classification accuracy was 85.7%, with 17 of 24 patients fication accuracy in percent by software, medical experts in
(71%) being correctly classified. The authors noted that acromegaly, and general internists. Republished with
the study sample was relatively small and also diverse in permission of Endocrine Society, from Schneider HJ, Kosilek
the patient group in terms of gender and race, while the RP, Gunther M, Roemmler J, Stalla GK, Sievers C, Reincke M,
control subjects were all white males. From our point of Schopohl J, Wurtz RP. A novel approach to the detection of
view, this can lead to a potential bias in classification acromegaly: accuracy of diagnosis by automatic face classi-
analyses. In a follow-up publication from 2011, the fication. Journal of Endocrinology and Metabolism 2011 96
accuracy of this computer system was compared to the 2074–2080.
diagnostic accuracy of ten generalist physicians, resulting
Classification Softwarea Expertsb Internistsc
in 86% vs 26% accuracy, respectively. Additionally, the
authors sorted a database of 200 normal subjects according Correct classification
to the presence of features of acromegaly using the same rates (%)
Overall 81.9 72.1 64.9
method (8). Acromegaly 71.9 63.2 42.1
Based on previous research, the authors of this review Controls 91.5 80.8 87.0
conducted a study regarding the detection of acromegaly By sex
Acromegalic women 62.1 48.3 26.4
and published the results in 2011 (4). In this study, a Control women 93.1 82.8 89.7
method consisting of Gabor wavelet transformations for Acromegalic men 82.1 78.6 53.6
feature extraction, geometric analysis, and a maximum Control men 90.0 76.7 84.4
By severity of facial
likelihood classifier was used. The overall correct classi-
European Journal of Endocrinology
features in acromegaly
fication rate was 81% in a sample of 57 patients Mild 58.3 38.9 20.8
(29 females, 28 males) and 59 control subjects (29 females, Intermediate 77.3 74.2 47.0
Severe 90.9 93.9 78.8
30 males) matched for gender and age, with 71.9% of
patients (sensitivity) and 91.5% of control subjects (speci- a
Classification by software, using a combination of the Gabor jet-based
ficity) correctly classified. Additionally, the patients were function P (normed scalar product including phase) and geometry-based
function L (edge difference length) on frontal and side views of patients
categorized into subgroups by disease severity according and controls.
b
Means of visual classification by three medical experts with extensive
to facial features of acromegaly. We also compared the experience in acromegaly.
c
accuracy of the software tool to the diagnostic accuracy of Means of visual classification by three general internists not particularly
experienced with acromegaly.
physicians assessing the images. Detailed results are shown
in Table 1. It is of note that software classification accuracy
exceeded the diagnostic performance of physicians asses-
sing the same images, especially in the category of patients 40 age- and gender-matched control subjects. The overall
with mild features of acromegaly. In a further step the classification accuracy was 91.7%, with 85% of patients
procedure was modified by reducing the number of nodes (sensitivity) and 95% of control subjects (specificity)
used for analyses and optimizing their placement. This correctly classified. The patient cohort was categorized
approach was then validated within the original data set by etiology of CS (Cushing’s disease, nZ8; adrenal CS,
and a new database of subjects, showing a moderate nZ4; iatrogenic CS, nZ8). Only patients with Cushing’s
improvement of classification rates (9). disease were not correctly classified by the software; this is
Additionally, Gencturk et al. (10) achieved an overall a finding of unclear significance. Since this was a proof-
classification accuracy of O90% using local binary of-concept study, there were three major limitations
patterns and a Manhattan classifier for the detection of (small, all-female study cohort, no matching of
acromegaly on facial photographs. patients and controls for BMI) that need to be addressed
in the future.
Our current studies aim to build on previous results
Application in CS and current research
described above by addressing limitations and implement-
The authors of this review also applied the method to the ing other measures. The recruitment targets for the next
detection of CS in a proof-of-concept study and published assessment of classification accuracy for CS are set at 50
the results in 2013 (5). For this, the experiments were patients and 100 control subjects matched for age, gender,
repeated in the original setup with a cohort of 20 female and BMI. Additionally, the placement of nodes used for
patients suffering from endogenous or iatrogenic CS and classification needs to be slightly modified to better match
www.eje-online.org
Downloaded from Bioscientifica.com at 03/19/2020 03:47:08AM
via free access
Special Section Review R P Kosilek and others Review of facial image analysis 173:4 M42
in diagnostics
facial features of CS. Clinical and biochemical parameters classifier (17). A second study from Turkey published
will be recorded alongside the software classification result data on the use of an image-processing method based on
with the goal of establishing a combined clinical predic- local binary patterns for feature extraction and changed
tion score. Manhattan distance for classification, with an overall
The increasing prevalence of obesity and its correct classification rate of O90% in a sample of 107 facial
sequelae (metabolic syndrome) continues to pose a images collected from the internet (51 subjects with Down
diagnostic challenge in clinical practice. The problem syndrome, 56 healthy individuals) (18). Zhao et al. (19)
has been extensively discussed in recent literature, such as reported that using Independent Component Analysis for
the fact that the diagnostic accuracy of various bio- landmark placement, local binary patterns for feature
chemical tests for hypercortisolism drops significantly extraction and SVMs for classification, they achieved
when applied to a sample of obese subjects (11, 12). In our 96.7% accuracy in the classification of a data set consisting
experience so far, this problem also applies to face of 130 facial photographs (50 patients with Down
classification technology. Determining and improving syndrome, 80 healthy subjects).
diagnostic accuracy within a BMI-matched sample is a Additionally, Douglas and Mutsvangwa (20) reviewed
distinct goal of current research. the use of various image analysis methods for the
identification of facial features associated with fetal
alcohol syndrome.
Application of face classification technology
in genetic disorders
Conclusions and future perspectives
In a series of publications from 2003 to 2011, a group of
European Journal of Endocrinology
authors from Germany reported on applying face classi- The use of face detection and classification technology in a
fication technology to the detection of multiple genetic medical setting appears to be a promising and emerging
syndromes. In studies published in 2003 and 2006, Loos area of research. A growing number of publications
et al., (13) and Boehringer et al. (14) reported on regarding the use of this technology as a diagnostic tool
performing face classification using only texture analysis have shown good results within certain boundaries, most
via Gabor wavelet transformations and various classifiers. importantly small study cohorts.
The data sets consisted initially of 55 patients suffering Independent of the particular method of image
from five genetic syndromes, and later of 147 patients analysis employed, experimental clinical use of such
suffering from ten genetic syndromes, achieving an overall systems requires a large training database to account for
classification accuracy of 76 and 75.7%, respectively variations in the photographs taken by external
(13, 14). In a study published in 2008, Vollmar et al. (15) physicians. This has also been noted by other authors as
included side-view photographs and geometric infor- pointed out above. Until now, most methods require
mation in the face classification analyses, resulting in manual processing of data to a certain extent. This is
improved accuracy in an even larger data set (14 genetic another important limiting factor. We think that
syndromes). In 2011, Boehringer et al. (16) applied the optimized and finally automatic selection of landmarks
previously described method in clinical practice by using a will also facilitate clinical use.
classifier trained on an existing data set to classify a new We estimate that a database consisting of 500–1000
set of patients with pictures taken in less standardized subjects should in theory be sufficient for experimental
conditions. While classification accuracy was good in clinical application as a screening tool. Possible targets for
training conditions, this study showed an unfavorable screening for acromegaly are patient populations with a
result of 21% classification accuracy. The authors attrib- higher pretest probability, such as patients presenting
uted this to large phenotypic differences in the groups of with dysgnathia at specialized centers. Regarding CS,
test subjects and suggested broadening the data set used similar applications are possible once recruitment targets
for training to improve accuracy in clinical practice (16). have been reached. For both diseases, integrating results
Apart from these reports, several studies addressed the from software face classification with clinical results could
detection of Down syndrome using facial photographs. further increase diagnostic accuracy.
A research group from Turkey achieved a maximum Regarding future perspectives, we do not expect that
classification accuracy of 97.34% when analyzing a set of this technology will ever replace biochemical testing in
30 facial images (15 Down syndrome, 15 healthy subjects) the diagnosis of acromegaly or Cushing‘s syndrome.
using Gabor wavelet transformations and a SVM Rather, it can be helpful as a screening tool in selecting
www.eje-online.org
Downloaded from Bioscientifica.com at 03/19/2020 03:47:08AM
via free access
Special Section Review R P Kosilek and others Review of facial image analysis 173:4 M43
in diagnostics
is essential for reducing the diagnostic delay mentioned in approach. Experimental and Clinical Endocrinology & Diabetes 2013 121
the introduction. Information programs might be helpful 561–564. (doi:10.1055/s-0033-1349124)
6 Günther M & Würtz RP. Face detection and recognition using
in achieving this goal.
maximum likelihood classifiers on Gabor graphs. International
Journal of Pattern Recognition and Artificial Intelligence 2009 23
433–461. (doi:10.1142/S0218001409007211)
7 Learned-Miller E, Lu Q, Paisley A, Trainer P, Blanz V, Dedden K &
Declaration of interest
Miller R. Detecting acromegaly: screening for disease with a morphable
R P Kosilek, R Frohner, R P Würtz, and C M Berr report no conflict of
model. Medical Image Computing and Computer-Assisted Intervention 2006
interest. J Schopohl received research grants, speaker fees, and travel
9 495–503.
grants from Ely Lilly, Ipsen, Novartis, and Pfizer. M Reincke received
8 Miller RE, Learned-Miller EG, Trainer P, Paisley A & Blanz V.
research grants and speaker fees from Novartis and speaker fees from
Early diagnosis of acromegaly: computers vs clinicians.
Pfizer and Ipsen. H J Schneider received research grants, speaker fees, and
Clinical Endocrinology 2011 75 226–231. (doi:10.1111/j.1365-2265.
travel grants from Pfizer and speaker fees and travel grants from Novartis.
2011.04020.x)
9 Frohner R, Wurtz RP, Kosilek RP & Schneider HJ. Optimierung der
Gesichtsklassifikation bei der Erkennung von Akromegalie. J Klin
Endokrinol Stoffw 2013 6 20–24.
Funding
10 Gencturk B, Nabiyev VV, Ustubioglu A, Ketenci S. Automated pre-
The research has been partially funded by a German Research Foundation
diagnosis of acromegaly disease using local binary patterns and its
(DFG) grant WU 314/6-2 to R P Würtz. It has also been partially funded by
variants. In Telecommunications and Signal Processing (TSP),
a research grant from Pfizer to H J Schneider. 36th International Conference on, pp 817–821, 2013. (doi:10.1109/TSP.
2013.6614052)
11 Baid SK, Rubino D, Sinaii N, Ramsey S, Frank A & Nieman LK.
Specificity of screening tests for Cushing’s syndrome in an overweight
Acknowledgements
and obese population. Journal of Endocrinology and Metabolism 2009 94
We thank Sandra Rutz for help in organization and patient recruitment.
3857–3864. (doi:10.1210/jc.2008-2766)
We also thank Gudrun Hackenberg and Kathrin H Popp for help in patient 12 Ellis E, Chin PK, Hunt PJ, Lunt H, Lewis JG & Soule SG. Is late-night
recruitment. salivary cortisol a better screening test for possible cortisol excess than
This paper forms part of a special issue of European Journal of standard screening tests in obese patients with type 2 diabetes?
Endocrinology on Cushing’s syndrome. This article is adapted from work New Zealand Medical Journal 2012 125 47–58.
presented at the IMPROCUSH-1: Improving Outcome of Cushing’s 13 Loos HS, Wieczorek D, Wurtz RP, von der Malsburg C & Horsthemke B.
Syndrome symposium, 12–14 October 2014. The meeting was supported Computer-based recognition of dysmorphic faces. European
by the European Science Foundation, Deutsche Forschungsgemeinschaft, Journal of Human Genetics 2003 11 555–560. (doi:10.1038/sj.ejhg.
Carl Friedrich von Siemens Stiftung, European Neuroendocrine Association 5200997)
and the Deutsche Gesellschaft für Endokrinologie. The opinions or views 14 Boehringer S, Vollmar T, Tasse C, Wurtz RP, Gillessen-Kaesbach G,
expressed in this special issue are those of the authors, and do not Horsthemke B & Wieczorek D. Syndrome identification based on
necessarily reflect the opinions or recommendations of the European 2D analysis software. European Journal of Human Genetics 2006 14
Science Foundation, Deutsche Forschungsgemeinschaft, Carl Friedrich von 1082–1089. (doi:10.1038/sj.ejhg.5201673)
Siemens Stiftung, European Neuroendocrine Association and the Deutsche 15 Vollmar T, Maus B, Wurtz RP, Gillessen-Kaesbach G, Horsthemke B,
Gesellschaft fur Endokrinologie. Wieczorek D & Boehringer S. Impact of geometry and viewing angle on
www.eje-online.org
Downloaded from Bioscientifica.com at 03/19/2020 03:47:08AM
via free access
Special Section Review R P Kosilek and others Review of facial image analysis 173:4 M44
in diagnostics
classification accuracy of 2D based analysis of dysmorphic faces. 18 Burçin K & Vasif NV. Down syndrome recognition using local binary
European Journal of Medical Genetics 2008 51 44–53. (doi:10.1016/j.ejmg. patterns and statistical evaluation of the system. Expert Systems with
2007.10.002) Applications 2011 38 8690–8695. (doi:10.1016/j.eswa.2011.01.076)
16 Boehringer S, Guenther M, Sinigerova S, Wurtz RP, Horsthemke B & 19 Zhao Q, Okada K, Rosenbaum K, Kehoe L, Zand DJ, Sze R, Summar M &
Wieczorek D. Automated syndrome detection in a set of clinical facial Linguraru MG. Digital facial dysmorphology for genetic screening:
photographs. American Journal of Medical Genetics. Part A 2011 155A Hierarchical constrained local model using ICA. Medical Image Analysis
2161–2169. (doi:10.1002/ajmg.a.34157) 2014 18 699–710. (doi:10.1016/j.media.2014.04.002)
17 Saraydemir S, Taspinar N, Erogul O, Kayserili H & Dinckan N. 20 Douglas TS & Mutsvangwa TE. A review of facial image analysis for
Down syndrome diagnosis based on Gabor Wavelet Transform. delineation of the facial phenotype associated with fetal alcohol
Journal of Medical Systems 2012 36 3205–3213. (doi:10.1007/s10916- syndrome. American Journal of Medical Genetics. Part A 2010 152a
011-9811-1) 528–536. (doi:10.1002/ajmg.a.33276)
www.eje-online.org
Downloaded from Bioscientifica.com at 03/19/2020 03:47:08AM
via free access