Interpretable self-supervised contrastive learning for colorectal cancer histopathology: GRADCAM visualization

Accurate colorectal cancer diagnosis from histopathological images is crucial for effective treatment. Therefore, it is of interest to describe a novel framework that combines self-supervised contrastive learning (SSCL) with Grad-CAM-based interpretability for classifying hyperplastic polyp (HP) and sessile serrated adenoma (SSA). A ResNet50 encoder is first pre-trained using SSCL to learn rich feature representations from unlabeled images, minimizing the need for manual annotations which are then fine-tuned in a supervised setting, achieving a classification accuracy of 85.86%. Grad-CAM is used to generate visual explanations, highlighting critical regions influencing the model's decisions. This interpretable, data-efficient approach outperforms conventional CNN methods, offering improved diagnostic accuracy and enhanced trust in automated pathology.