Using face classification for detecting acromegaly in a web-based environment

Introduction:

It has been shown that face classification software can help distinguish between subjects with and without acromegaly on regular photographs. In this project we have shown that this method can be provided to the medical public in a first clinical application.

Methods:

Face classification was based on nodes placed on frontal and side photographs of individuals and analyses the underlying texture and geometry. In the first step we analysed how to optimize the method to get sufficient results in both specificity and sensitivity. We optimized the training database by expanding the original database from 116 (57 acromegalics and 59 controls) to 479 (112 acromegalics and 367 controls) based on defined quality criteria and we developed a new classification graph considering typical morphological changes in acromegaly. In a second step we connected this method with a web based application to be usable for interested medical staff.

Results:

Correct classification rates in the original database were 81% with the original classification graph (71% and 89% of acromegalics and controls, respectively). Using the same approach, in the extended and evaluated database with 367 controls and 112 acromegalics in combination with a new classification graph, the correct classification rates were 92.1% (88.2% and 93.2% of acromegalics and controls, respectively). Interested medical staff can use these method to detect acromegaly by using the web-based platform https://vidisto.mki.klinikum.uni-muenchen.de

Conclusions:

Optimizing and extending the training database by defined quality criteria and modifying graphs adapted to acromegaly can improve correct classification rates in the detection of acromegaly by face classification software. Combined with a web-based application this method can be used by medical staff in a medical environment, this study was supported with an unrestricted grant from Pfizer