Landmarking-Based Unsupervised Clustering of Human Faces Manifesting Labio-Schisis Dysmorphisms

Abstract

Ultrasound scans, Computed Axial Tomography, Magnetic Resonance Imaging are onlyfew examples of medical imaging tools boosting physicians in diagnosing a wide rangeof pathologies. Anyway, no standard methodology has been dened yet to extensivelyexploit them and current diagnoses procedures are still carried out mainly relying onphysician's experience. Although the human contribution is always fundamental, it isself-evident that an automatic procedure for image analysis would allow a more rapidand eective identication of dysmorphisms. Moving toward this purpose, in this workwe address the problem of feature extraction devoted to the detection of specic dis-eases involving facial dysmorphisms. In particular, a bounded Depth Minimum SteinerTrees (D-MST) clustering algorithm is presented for discriminating groups of individu-als relying on the manifestation/absence of the labio-schisis pathology, commonly calledcleft lip. The analysis of three-dimensional facial surfaces via Dierential Geometry isadopted to extract landmarks. The extracted geometrical information is furthermoreelaborated to feed the unsupervised clustering algorithm and produce the classication.The clustering returns the probability of being aected by the pathology, allowing physi-cians to focus their attention on risky individuals for further analysis.