Accurate Localization of Inner Ear Regions of Interests Using Deep Reinforcement Learning

Abstract

We propose a novel method for automatic ROI extraction. The method is implemented and tested for isolating the inner ear in full head CT scans. Extracting the ROI with high precision is in this case critical for surgical insertion of cochlear implants. Different parameters, such as CT equipment, image quality, anatomical variation, and the subject's head orientation during scanning make robust ROI extraction challenging. We propose to use state-of-the-art communicative multi-agent reinforcement learning to overcome these difficulties. We specify landmarks specifically designed to robustly extract orientation parameters such that all ROIs have the same orientation and include the relevant anatomy across the dataset. 140 full head CT scans were used to develop and test the ROI extraction pipeline. We report an average overall estimated error for landmark localization of 1.07 mm. Extracted ROI presented an intersection over union of 0.84 and a Dice similarity coefficient of 0.91.