Unsupervised classification of congenital inner ear malformations using DeepDiffusion for latent space representation

The identification of congenital inner ear malformations is a challenging task even for experienced clinicians. In this study, we present the first automated method for classifying congenital inner ear malformations. We generate 3D meshes of the cochlear structure in 364 normative and 107 abnormal anatomies using a segmentation model trained exclusively with normative anatomies. Given the sparsity and natural unbalance of such datasets, we use an unsupervised method for learning a feature representation of the 3D meshes using DeepDiffusion. In this approach, we use the PointNet architecture for the network-based unsupervised feature learning and combine it with the diffusion distance on a feature manifold. This unsupervised approach captures the variability of the different cochlear shapes and generates clusters in the latent space which faithfully represent the variability observed in the data. We report a mean average precision of 0.77 over the seven main pathological subgroups diagnosed by an ENT (Ear, Nose, and Throat) surgeon specialized in congenital inner ear malformations.