1 June 2023 in Computers in Biology and Medicine
6 April 2023 in Biomedical Signal Processing and Control
1 March 2026 in Knowledge-Based Systems
22 May 2026 in Expert Systems with Applications
15 September 2025 in International Journal of Modern Physics E
1 February 2026 in Digital Health
Purpose: Early diabetes screening remains constrained by the high cost and reliance on specialized equipment associated with conventional biochemical assays. Tongue imaging, as a non-invasive and convenient modality, provides diagnostically relevant cues. However, diabetes-related tongue manifestations are typically spatially sparse and exhibit low contrast, making them difficult to annotate and prone to being overlooked. This study aims to develop a self-supervised learning framework that enhances the preservation of weak pathological signals for tongue image–based diabetes screening. Methods: We propose Tongue Structure-Enhanced Learning (TSEL), a self-supervised framework that incorporates structure-constrained representation learning. Specifically, a Structural Decoupling Predictor (SDP) is introduced to assign soft structural attributes to image patches, reducing background-dominant interference. In addition, a Structure-Enhanced Consistency (SECC) mechanism is designed to align representations between complete and masked views, improving representation stability. The proposed method was evaluated on three diabetes-related tongue image datasets. Results: Experimental results demonstrate that TSEL consistently outperforms multiple self-supervised learning baselines across all datasets, with particularly notable improvements under low-label settings. The framework shows enhanced capability in capturing diagnostically relevant features compared to conventional methods. Conclusion: TSEL effectively preserves and enhances weak, spatially sparse pathological signals in tongue images through structure-constrained learning. These findings highlight the importance of incorporating structural priors in self-supervised frameworks and demonstrate the potential of TSEL for non-invasive diabetes screening.