UNSUPERVISED DEEP LEARNING FOR DETECTION OF BRAIN DISEASE IN MR IMAGING

• Published in: Clinical neuroradiology (Munich) Vol. 29; no. S1; p. S14
• Main Authors: Eisawy, Rami, Moosbauer, Julia, Finck, Tom, Zimmer, Claus, Menze, Bjorn, Pfister, Franz M.J, Wiestler, Benedikt
• Format: Journal Article
• Language: English
• Published: Springer 01.09.2019
• Subjects: Algorithms; Machine learning; Magnetic resonance imaging; Medical research; Medicine, Experimental; Physiological aspects
• ISSN: 1869-1439

Background: Manual detection and interpretation of suspicious findings in radiological exams is a slow and lengthy process, requiring the highest level of attention and expertise. Introducing an automatic approach to distinguish abnormal from normal anatomy and physiology has the potential to speed up the diagnostic process and avoid errors by serving as a second read. Recently, there have been numerous attempts to solve this problem with the use of supervised machine learning. However, this poses three major limitations: Costliness of data annotation, risk of annotation errors being reproduced by the algorithm and lack of training data for rare diseases. Methods: To avoid those limitations, an unsupervised, progressively growing adversarial autoencoder model is presented. The model is exclusively trained on images from a healthy cohort, thus learning the normal anatomy and signal of the brain. Subsequently, pathologies are automatically detected as deviations from this normality. We evaluate the model on datasets containing a variety of MS lesions, WMH, stroke, and GBM across the FLAIR MR sequence. Results: The evaluation shows the effectiveness of the method and its ability to highlight even small abnormalities in brain MRI exams, yielding state-of-the-art results in terms of overlap error and distance-based measures (Average DSC = 0.614 [+ or -] 0.135) in diseases the network has never seen before. Discussion: Qualitative results effectively show the power of the proposed autoencoder model. Reconstructed samples show high detail while simultaneously not reconstructing the pathological areas, facilitating a segmentation pipeline. Additionally, no modifications to the model are needed to allow for extensive generalisability to various applications such as detecting other white matter hyperintensities, tumors or inflammation, or even other modalities (e. g. CT) or parts of the body. Conclusion: To further evaluate the model and achieved results thus far, additional experiments on a larger dataset consisting of various different pathologies are scheduled. Clinical studies proving model efficacy is subject to ongoing research and intended to substantiate the use of the model as a clinical tool.