Fusion color space analysis of tongue image characteristics in different stages of metastatic colorectal cancer and three-class model construction.

OBJECTIVE: Color spaces were used to analyze the color characteristics of tongue image in different stages of metastatic colorectal cancer (mCRC), and a three-class model was constructed to predict the different tongue characteristics during the mCRC progression. METHODS: Tongue images of CRC patients were collected and classified into non-metastatic, oligometastatic, and extensive metastatic groups according to tumor metastasis. The color components in HSB, Lab, and YCrCb color spaces were measured from tongue images and the differences in color characteristics were quantitatively analyzed. Based on the modified YOLO and combined with the expert model and multilayer perceptron (MLP) fusion framework, the three classification models of tongue images were constructed. RESULTS: Most of the quantitative results showed the quantitative results for oligometastasis were the smallest values among the three metastatic stages. The S* and b* values at the edge of the tongue and the b* values in the middle of the tongue showed the quantitative results for oligometastasis were the highest values in the three stages. Besides, Cr* values at the edge of the tongue and S* values in the middle of the tongue showed a decreasing trend. In addition, all expert models constructed in the study demonstrated stable convergence on the training corpus within 100 epochs. The overall accuracy of the fused MLP classifier for tongue characteristics classification was 83.3%. CONCLUSIONS: In this study, tongue feature variations across different stages of CRC were successfully analyzed. A well-performing three-classification tongue image analysis model based on the YOLO architecture was initially constructed.