Breast cancer continues to be a major health challenge worldwide, necessitating advanced diagnostic techniques for early and accurate detection. Despite advancements in using deep learning to analyze medical images, its dependence on extensive labeled datasets restricts its practical use, particularly in environments with limited annotated data. This study proposes a Self-Supervised Dual-Pathway Learning Method to address this limitation. The framework integrates Contrast-Limited Adaptive Histogram Equalization (CLAHE) for image preprocessing, Self-Supervised Learning (SSL) using SimCLR for feature extraction, and Transfer Learning for model adaptation. EfficientNetB0, DenseNet121, ResNet18, and MobileNetV2 are applied to both mammographic and thermal breast images to enhance classification accuracy, robustness, and interpretability. By leveraging self-supervised pretraining, the model effectively learns feature representations from unlabeled data, improving generalization across imaging conditions. Experimental evaluations demonstrate that EfficientNetB0 achieves 98% accuracy, outperforming conventional supervised models. The method ensures high sensitivity and specificity, enabling reliable differentiation between malignant and benign cases while reducing dependency on labeled datasets. Incorporating thermal imaging enhances diagnostic reliability by capturing physiological variations associated with malignancies. This approach holds significant promise for clinical implementation, particularly in resource-limited settings where expert-annotated data is scarce. Future research will focus on exploring Transformer-based architectures and multi-modal imaging techniques to refine feature extraction and diagnostic precision further. The proposed strategy offers a scalable and efficient solution to breast cancer detection, addressing key limitations in medical imaging while advancing the role of self-supervised learning in healthcare AI.