A novel multi-feature learning model for disease diagnosis using face skin images

Area of Science:

• Dermatology
• Medical Imaging
• Machine Learning

Background:

• Facial skin characteristics offer insights into underlying health conditions.
• Divergent samples, influenced by environmental or genetic factors, reduce diagnostic accuracy.
• Accurate diagnosis is crucial for managing various health conditions.

Purpose of the Study:

• To develop a novel multi-feature learning method to mitigate divergent sample influence on diagnoses.
• To enhance the accuracy of classifying facial skin samples, especially those on the boundary.
• To enable non-invasive simultaneous detection of Diabetes Mellitus (DM), Fatty Liver (FL), and Chronic Renal Failure (CRF).

Main Methods:

• Proposed a Multi-Feature Learning with Centroid Matrix (MFLCM) method.
• Introduced a novel discriminator with a centroid matrix strategy integrated into a unified model.
• Applied centroid matrix to embedding feature spaces using relaxed Hamming distance for classification.

Main Results:

• Achieved high F1 scores: 92.59% for Healthy, 83.35% for DM, 82.84% for FL, and 85.46% for CRF.
• Demonstrated superior performance compared to single-view and state-of-the-art multi-feature methods.
• Validated the method on a clinical facial skin dataset.

Conclusions:

• MFLCM effectively mitigates the impact of divergent samples in facial skin analysis.
• This study is the first to use facial skin images for non-invasive, simultaneous detection of DM, FL, and CRF in Han Chinese.
• The proposed method offers a promising non-invasive approach for early disease detection.