Fully automated analytic system for measuring endolymphatic hydrops ratio in Ménière's disease using deep learning and MRI (Preprint)
BACKGROUND Recently, analysis of endolymphatic hydrops (EH) using inner ear magnetic resonance imaging (MRI) in Ménière's disease (MD) has been attempted in various studies. In addition, artificial intelligence (AI) has rapidly been incorporated into the medical field. In our previous study, the automated analysis algorithm of EH was completed using the convolutional neural network (CNN). However, several limitations existed, and further studies were conducted to compensate for these limitations. OBJECTIVE The aim of this study is to develop a fully automated analytic system for measuring endolymphatic hydrops, which provides enhanced analysis accuracy and clinical usability in studying Ménière's disease with MRI. METHODS We propose 3into3Inception and 3intoUNet, whose network architectures are based on Inception-v3 and U-Net, respectively. The developed networks were trained for inner ear segmentation using magnetic resonance (MR) images of 124 people and were embedded in a new automated EH analysis system, INner ear Hydrops Estimation via ARtificial InTelligence - version 2 (INHEARIT-v2). After 5-fold cross-validation, an additional test was performed using 60 new unseen MR images to evaluate the performance of our system. INHEARIT-v2 has a new functionality to automatically select representative images from a full MR stack. RESULTS The average segmentation performances of 5-fold cross-validation were measured by the intersection of union, which showed 0.743 ± 0.030 for 3into3Inception and 0.781 ± 0.030 for 3intoUNet. The automatic representative slice selection results of the INHEARIT-v2 differed only within two slices from the expert selection on an unseen dataset. Compared with the ratio measured by experienced physicians, the average interclass correlation coefficient (ICC) for all cases was 0.941; the average ICC of the vestibules was 0.968, and that of cochleae was 0.914. The time required for the fully automated system to accurately analyze the EH ratio in one patient's MRI stack was approximately 3.5 seconds. CONCLUSIONS In this study, a fully automated full-stack MR analysis system of the EH ratio was developed, named INHEARIT-v2, which showed high agreement with experts in an additional test. The system is an upgraded version of INHEARIT and provides higher segmentation performance and includes automatic representative image selection in the MR stack. The new model can help clinicians by providing an objective analysis result and reduce their workload in reading MRIs.