Diagnostic use of facial image analysis software in endocrine and genetic disorders: review, current results and future perspectives

Cushing's syndrome (CS) and acromegaly are endocrine diseases that are currently diagnosed with a delay of several years from disease onset. Novel diagnostic approaches and increased awareness among physicians are needed. Face classification technology has recently been introduced as a promising diagnostic tool for CS and acromegaly in pilot studies. It has also been used to classify various genetic syndromes using regular facial photographs. The authors provide a basic explanation of the technology, review available literature regarding its use in a medical setting, and discuss possible future developments. The method the authors have employed in previous studies uses standardized frontal and profile facial photographs for classification. Image analysis is based on applying mathematical functions evaluating geometry and image texture to a grid of nodes semi-automatically placed on relevant facial structures, yielding a binary classification result. Ongoing research focuses on improving diagnostic algorithms of this method and bringing it closer to clinical use. Regarding future perspectives, the authors propose an online interface that facilitates submission of patient data for analysis and retrieval of results as a possible model for clinical application.

Automatic face classification of Cushing's syndrome in women - a novel screening approach

OBJECTIVE

Cushing's syndrome causes considerable harm to the body if left untreated, yet often remains undiagnosed for prolonged periods of time. In this study we aimed to test whether face classification software might help in discriminating patients with Cushing's syndrome from healthy controls.

DESIGN

Diagnostic study.

PATIENTS

Using a regular digital camera, we took frontal and profile pictures of 20 female patients with Cushing's syndrome and 40 sex- and age-matched controls.

MEASUREMENTS

Semi-automatic analysis of the pictures was performed by comparing texture and geometry within a grid of nodes placed on the pictures. The leave-one-out cross-validation method was employed to classify subjects by the software.

RESULTS

The software correctly classified 85.0% of patients and 95.0% of controls, resulting in a total classification accuracy of 91.7%.

CONCLUSIONS

In this preliminary analysis we found a good classification accuracy of Cushing's syndrome by face classification software. Testing accuracy is comparable to that of currently employed screening tests.

Efficient evaluation of serial sections by iterative Gabor matching

Evaluation of electron microscopic images of serial sections is a time-consuming process requiring a high level of expertise. Here we present an algorithm to ease and accelerate this process. It is a modification of an algorithm successfully used in computer vision for object recognition. However, rather than recognising individual structures, we estimate the spatial mapping of a whole section onto the consecutive one. This mapping is used to transfer labelled information of the very first section, e.g. a classification by a human expert of different visible structures, onto structures visible in the next section. We investigate its performance on an artificially constructed benchmark as well as on real electron microscopic samples taken in primary visual cortex and demonstrate its potential for dramatically facilitating the evaluation process of serial sections.

A Hierarchical Dynamic Link Network to Solve the Visual Correspondence Problem

Conventional neural networks try to solve the problem of object recognition in a single step by building a stimulus — response system that codes its result as cell activities. We take a different approach assuming that recognition is an active process with temporal dynamics and results in an ordered state. We present a structure of neuronal layers, interconnected by dynamic links (von der Malsburg, 1985 Berichte der Bunsengesellschaft für Physikalische Chemie89 703 – 710) that solves the correspondence problem between two images and thus constitutes an important building block for a model of recognition.

Images as well as stored models are represented as Gabor pyramids. This allows the dynamics to proceed from coarse to fine scale and reduces the sequential processing time inherent in the concept. Invariance under background changes is also made possible.

On the lowest frequency level, a single blob of activity moves across the image and model layer, respectively. Dynamic links between these layers are initialised to the (highly ambiguous) feature similarities. Links grow or decline according to a combination of feature similarity and correlated activation. This enforces correct neighbourhood relationships in addition to feature similarity. On the higher levels the established correspondences are refined by several blobs in parallel.

We present an improved version of the dynamical system proposed by Würtz [1995 Multilayer Dynamic Link Networks for Establishing Image Point Correspondences and Visual Object Recognition (Thun, Frankfurt a.M.: Harri Deutsch)] and show, with examples of human faces, that it evolves from an unordered link distribution to any ordered state where only corresponding point pairs are connected by strong links. Correspondences between sample points are population-coded by a set of neighbouring links.