Tongue image analysis in traditional Chinese medicine involves determining the health status of the human body by observing the color, shape and coating of the tongue. Tooth marks are commonly used in clinics to determine whether a patient has a spleen deficiency constitution and are an important observation index in traditional Chinese medicine (TCM) tongue diagnosis. However, there is usually a serious imbalance in the available tongue data, and the unavoidable subjectivity of the labeling process makes the labeling results uncertain, which further affects the accuracy of tongue analysis. To achieve accurate segmentation of tooth marks, this paper produces a semantic segmentation dataset of tooth marks, which provides a detailed annotation of the target edges of tooth marks in the tongue image, which provides a corresponding research basis for the subsequent semantic segmentation of tooth marks. To address the imbalance of data, we proposed a tongue image semantic segmentation model based on cluster-based adaptive data conditioning. First, the data were clustered using K-means to divide the data into different groups. Second, based on the hybrid Transformer Medical Segmentation Network (MT-UNet) of the UNet configuration, we changed its data delivery method, and in the process of network training, we adjusted the data inputs of each cluster dynamically based on the score of the validation set. The ratio dynamically adjusts the number of data inputs for each cluster, which balances the imbalance of the data distribution and simultaneously finds a unified standard for semantic segmentation learning of cluster data with different criteria.