Optimizing strategies for face classification in the detection of acromegaly

Introduction: It has been shown that face classification software might help distinguishing between subjects with and without acromegaly on regular photographs. In this project we investigated several aspects that will be necessary and helpful to bring this recognition method closer to clinical application.

Methods: Face classification was based on nodes placed on frontal and side photographs of individuals and analysis the underlying texture and geometric functions. In the first step we analysed whether omission of nodes considered less relevant changed classification rates in the original database on 57 acromegalics and 60 controls. In a second step we analysed how a complete new set of nodes (referring to the most common morphological changes in face) will affect the classification rate.

In a third step, we analysed whether classification was improved in an external data set consisting of 13 acromegalics and 45 controls for both steps.

Results: Correct classification rates in the original database were 79% with all nodes 78% if irrelevant points were omitted and 80% using the new set of nodes.

Using the same approach, in the validation set, correct classification rate were 88% with all nodes (92% and 87% of acromegalics and controls, respectively) 94% (100% and 93% of acromegalics and controls, respectively) after omission of irrelevant nodes and 95% (100% and 94% of acromegalics and controls, respectively) with the new set of nodes

Conclusions: Reduction of nodes associated with unwanted noise can improve correct classification rates in the detection of acromegaly by face classification software.