Identification of dysmorphic syndromes using landmark-specific local texture descriptors

The early detection of genetic disorders in infants is crucial for the timely management of patients and disease. The particular facial characteristics of patients affected by dysmorphic syndromes, which account to about half of genetic disorders, allows to identify positive cases prior to cytogenetic results, and avoid the overuse of genetic blood tests. However, the diagnostic accuracy by pediatricians is moderate. In this work, we present a general framework for the detection of genetic disorders from facial pictures, combining geometrical and texture features. Based on the 2D extension of Linear Discriminant Analysis, we propose the extraction of optimal landmark-specific Local Binary Pattern-based features. In particular, the proposed framework computes optimal local image filters and soft neighborhood weighting matrices that enhance the discriminative ability of the system. This new framework was tested on a database of 145 cases, including 73 pathological patients with 15 different genetic syndromes, obtaining a detection accuracy of 0.95.