A tri-branch multi-view contrastive learning framework integrated with physiological information for tumor classification in dynamic optical breast imaging data

Contrastive learning has gained popularity due to its pretraining capabilities on large-scale datasets and extensive data augmentation methods. However, its application in breast tumors classification using Dynamic Optical Breast Imaging (DOBI) data—a functional imaging modality that assesses the metabolic characteristics of tumor neovasculature through dynamic optical signals and offers a non-invasive, radiation-free alternative—faces several challenges. Beyond common algorithmic hurdles such as limited data augmentation options and small dataset sizes, DOBI technology itself introduces significant complexities, including high intra-class variability due to physiological differences (e.g., breast density, hormonal status), potential artifacts from patient motion during dynamic acquisition, and the high-dimensional, spatio-temporal nature of the data which demands robusts feature learning. To address these challenges, this study proposes a tri-branch multi-view contrastive learning classification framework combined with physiological information (TMCLph) for the binary classification of breast tumors (benign vs. malignant). Firstly, a multi-view contrastive learning-based pretraining strategy is presented, which leverages multi-view information for data augmentation, thereby enhancing the diversity of samples in the pretraining dataset. Then, a comprehensive contrastive loss function is formulated, integrating unsupervised and supervised contrastive loss with physiological information, to constrain both inter-class and intra-class variability effectively. Finally, a tri-branch model is designed to fully exploit the spatial–temporal information of the DOBI data, aiming to distinguish malignant from benign tumors accurately. Experiments conducted on 596 breast tumor patients demonstrate that the proposed framework can achieve superior performance, with an accuracy of 0.76 and an AUC of 0.81, outperforming state-of-the-art approaches for breast tumor classification. Ablation studies further confirm the effectiveness of each component, particularly highlighting that the comprehensive contrastive loss incorporating physiological information contributes to a performance gain of approximately 0.15 in accuracy. The proposed framework effectively mitigates the challenges of small-scale data and high intra-class variability in DOBI, leading to notably improved sensitivity (0.09) and specificity (0.1). Consequently, our method signigicantly enhances the diagnostic accuracy of breast cancer using DOBI and establishes a broadly applicable paradigm for small-sample medical image analysis.