Tongue image dataset for Tri-Dhat classification in traditional Thai medicine

Traditional Thai medicine (TTM) is an increasingly popular treatment option. Tongue diagnosis is a highly efficient method for determining overall health, as practiced by TTM practitioners. However, the diagnosis naturally varies depending on the practitioner's expertise. In this work, we propose tongue image analysis with raw pixels using artificial intelligence (AI) to support TTM diagnoses. The target classification of Tri-Dhat consists of three classes: Vata, Pitta, and Kapha. We utilized our own organized, genuine datasets collected from our university TTM hospital. Class balancing and data augmentation were conducted, and we present analysis approaches and experimental designs. Transfer learning techniques for various pretrained deep learning models were developed. We used two-tailed paired t-tests and single-factor ANOVA for performance comparisons. Our work demonstrated that the DenseNet121 and Xception models provided the most significant results with cropped image datasets, including DSLR-taken and mobile-taken images. Notably, model ensemble evaluations yielded the highest average predictions, achieving a precision of 0.94, an F1 score of 0.96, an accuracy of 0.96, a sensitivity of 0.96, and a specificity of 0.97, supported by a p-value of 0.0003 from ANOVA. We suggest that our methods could be effectively deployed in real-world scenarios to aid TTM practitioners in their diagnoses.