A study released on Jul. 23 describes the development of a multimodal deep learning model designed to screen for metabolic dysfunction-associated fatty liver disease (MAFLD) using both tongue image features and routine clinical data.
Researchers included 477 subjects from an initial pool of 904 candidates, with participants divided into training, validation, and test sets. Each participant underwent standardized tongue imaging using International Commission on Illumination L*a*b color features and comprehensive clinical evaluation. The dual-stream model combined a ConvNeXt-Tiny network for analyzing tongue images with a multilayer perceptron processing clinical variables. These features were merged through a Dynamic Affine Feature Transformation module, and the system was trained with weighted cross-entropy loss.
Results showed that patients with MAFLD exhibited significant metabolic abnormalities compared to healthy controls. The study also found that as fibrosis advanced, there was a progressive decrease in tongue yellowness (b* value). On an independent test set of 48 subjects, the multimodal model achieved an accuracy of 97.92%, Quadratic Weighted Kappa of 0.9538, sensitivity of 96.88%, and specificity of 100%. These results surpassed those obtained by single-modality or serological models alone.
Interpretability analyses indicated that the model concentrated on clinically relevant regions of the tongue and key metabolic indicators when making assessments. According to researchers, this supports the integration of objective Traditional Chinese Medicine (TCM) tongue appearance features as part of non-invasive diagnostic strategies.
The study concludes that this clinically oriented auxiliary screening tool may offer a practical approach for non-invasive assessment of MAFLD, especially in areas where healthcare resources are limited.
