Clinical Face Phenotype Space: dysmorphology photographs to aid diagnosis of genetic syndromes

Clinical dysmorphologists require an enormous breadth of experience to correctly classify and diagnose ultra-rare diseases. Developments in computer vision research now enable computational phenotyping based on ordinary photographs. Computational phenotyping tools are poised to become a transformative technology in healthcare, bringing objective quantification and learned models to help clinicians diagnose, prioritise interventions, and monitor outcomes.

We recently introduced Clinical Face Phenotype Space (CFPS), which locates patients in the context of known syndromes, and thus can help to generate disease hypotheses. Moreover, CFPS improved the clustering of patients by phenotype by 27 fold over random chance, even when no known syndrome diagnosis exists. This holds promise as an impartial means by which to narrow the search space for suspected rare diseases, and could augment the prioritisation of testing in clinical investigations. The CFPS algorithm automatically detects faces in photographs, annotates locations of key anatomical parts, and extracts machine readable feature descriptions of the facial gestalt.

From a proof-of-principle CFPS we have further improved the face detection, feature point annotation and feature descriptors. We increase accuracies of facial phenotype representations and test how novel models of phenotype variation can aid narrowing the search space to diagnoses. More precise face detections and feature point annotations improve the fidelity of the phenotypes captured in photographs, which can be used to generate life-like representations of syndrome characteristics useful for training. Furthermore we show how the choice of feature descriptor has a large impact on CFPS modelling.

Importantly we are launching the CFPS Consortium for clinical collaborations. We have established a legal, ethical and secure framework for clinical collaborations which safeguards the interests of institutions, clinician researchers and patients. Together we will bring computational phenotyping to clinical utility.