Vision Transformer-Based Facial Infrared Feature Mining with Biochemical Indicator Fusion for Liver Cirrhosis Staging Diagnosis
Jiahui Li, Xiaoqiong Li, Boyang Li, Kun Wang, Tingting He, Xinyu Zhang, Ruilin Wang · 2025
Liver cirrhosis significantly increases mortality risk in hepatocellular carcinoma (HCC) patients, and its staging (compensated/decompensated) serves as a critical prognostic indicator. Although blood biochemical parameters reflect hepatic injury, their diagnostic utility is limited by susceptibility to interference and inability to function as standalone diagnostic tools. In the field of Traditional Chinese Medicine (TCM) facial diagnosis, the liver is considered closely related to blood regulation, where hepatic dysfunction leads to impaired blood circulation, consequently affecting facial blood flow and temperature. Infrared thermography captures facial infrared radiation characteristics to construct temperature feature models, and prior studies have explored its feasibility and effectiveness in disease diagnosis tasks such as coronary artery disease (CAD). Therefore, integrating infrared features with biochemical parameters to develop a bimodal model may overcome the limitations of unimodal data. Based on a clinical cohort of 164 liver disease cases (compensated: 107, decompensated: 57; active HCC: 74, cured: 90), this study investigates two core questions: 1) how to extract liver diseasespecific low-dimensional features from high-dimensional infrared images; 2) how to validate the additive effect of bimodal fusion on cirrhosis staging diagnosis. We designed a Bimodal DataBased Liver Cirrhosis Classifier (BDLCC): 1) the MediaPipe facial recognition algorithm automatically constructs standardized ROIs; 2) Vision Transformer (ViT) achieves cross-regional feature fusion through global attention mechanisms and projects data into liver disease-relevant low-dimensional subspaces; 3) dimensionality-reduced infrared features are combined with biochemical parameters to build a random forest classifier, evaluated via five-fold cross-validation. Experimental results demonstrate that the bimodal model achieved 52.7 % accuracy in cirrhosis staging, reflecting a 5.4 % improvement over the unimodal blood biochemical model and a 4.1 % increase compared to the unimodal infrared imaging model. Regarding the misdiagnosis rate of misclassifying compensated cirrhosis as decompensated, the bimodal model reduced this to 8.1 %, representing a 21.6 % reduction from the blood biochemical unimodal model and an 8.1 % decrease from the infrared unimodal model. The bimodal model outperformed both unimodal models in the binary classification task of cirrhosis staging, confirming the complementary value of infrared imaging features and traditional blood biochemical parameters. The captured thermal anomaly signals provide an objective quantitative basis for validating Traditional Chinese Medicine (TCM) facial diagnostic theories.