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

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Accurate diagnosis of colorectal polyps from histopathological images is critical for effective patient treatment and management.
• Distinguishing between hyperplastic polyps (HP) and sessile serrated adenomas (SSA) is diagnostically challenging and crucial for cancer risk stratification.

Purpose of the Study:

• To develop and evaluate a novel framework combining self-supervised contrastive learning (SSCL) with Grad-CAM interpretability for classifying HP and SSA from histopathological images.
• To enhance the data efficiency and diagnostic accuracy of automated pathology systems.

Main Methods:

• A ResNet50 encoder was pre-trained using SSCL on unlabeled histopathological images to learn robust feature representations.
• The pre-trained model was fine-tuned in a supervised setting for HP and SSA classification.
• Grad-CAM was employed to generate visual explanations, identifying image regions critical for the model's classification decisions.

Main Results:

• The proposed framework achieved a classification accuracy of 85.86% for distinguishing HP from SSA.
• The SSCL approach demonstrated superior performance compared to conventional Convolutional Neural Network (CNN) methods.
• Grad-CAM provided interpretable insights into the model's decision-making process.

Conclusions:

• The developed interpretable and data-efficient framework significantly improves diagnostic accuracy in automated colorectal polyp classification.
• Combining SSCL with Grad-CAM enhances trust in AI-driven pathology by providing visual explanations for model predictions.
• This approach offers a promising solution for more reliable and efficient histopathological analysis in clinical settings.