Artificial Intelligence (AI) in Medicine

Abstract

Congenital syndromes with subtle changes in maxillofacial morphology can pose significant diagnostic challenges, wherein artificial intelligence holds great promise in aiding diagnosis through shape analysis. We applied the recently proposed Swap Disentangled Variational Autoencoder (SD-VAE) in diagnosis of Beckwith-Wiedemann Syndrome (BWS). The SD-VAE model was trained on a dataset primarily comprised of 3D stereophotogrammetry (3D SPG) scans gathered using a 3dMD Head System (3dMD LLC). It was also trained on CT scans when available. A total of 72 syndromic scans were used belonging to 56 different BWS patients. Scans of head shapes were pre-processed to create meshes in dense anatomical correspondence with consistent topology. This was achieved through non-rigid iterative closest point (NICP) registration, guided by 68 facial landmarks and additional Gaussian Processes, if needed. To make the SD-VAE outputs more interpretable, latent vectors with high dimensionality were embedded into a twodimensional space via t-distributed Stochastic Neighbour Embedding (t-SNE). Furthermore, latent vectors were classified with Quadratic Discriminant Analysis (QDA). The model demonstrated a perfect diagnostic accuracy for BWS on the test set, with the most characteristic regions being the chin, cheeks, zygoma, eyes, jaw, and supraorbital region. This paper demonstrates how SD-VAE can be applied to 3D head meshes, to quantify the characteristic features of BWS. We distinguished BWS-specific features from those of the general population with high diagnostic accuracy. This makes SD-VAE a promising tool for aiding the referral and diagnosis of BWS in the future.

Biography

TBA