Abstract
We train a deep learning framework known as the PointNet image segmentation model to perform several predictive geometric operations on point cloud files of human ear impressions. These operations include predictions of ear sides, percentage of unwanted portions of the ear to delete and the rotation matrix which aligns each ear impression such that the ear canal lies along the y-axis. We find that the PointNet model manages to perform all operations with consistently high batch accuracy percentages, making it suitable for use in applications that automate the pre-processing of patient ear scans in the In-the-Canal (IC) hearing aid design process. We then supplement our computational solutions using PointNet with insights gained from studying findings from a k-means unsupervised classification heuristic which we performed on samples from our ear impression pool.