XDecompo: Explainable Decomposition Approach in Convolutional Neural Networks for Tumour Image Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Colorectal cancer and brain tumors are leading causes of death globally.
• Accurate medical image diagnosis is crucial for effective treatment.
• Self-supervised learning shows promise for medical AI, especially with limited annotated data.

Purpose of the Study:

• To develop a robust self-supervised model, XDecompo, for improved medical image classification.
• To enhance the transferability of features from pretext tasks to downstream diagnostic tasks.
• To address challenges in medical image analysis, such as data irregularities and insufficient annotations.

Main Methods:

• Proposed XDecompo, a novel self-supervised model utilizing affinity propagation-based class decomposition.
• Implemented an explainable component to identify key image features and validate class decomposition effects.
• Evaluated model generalizability on histopathology images for colorectal cancer and brain tumors.

Main Results:

• XDecompo achieved high accuracy: 96.16% for colorectal cancer and 94.30% for brain tumors.
• Demonstrated robust performance and generalization across different medical imaging datasets.
• Validated feature transferability and representation accuracy using a post hoc explainable method.

Conclusions:

• XDecompo offers a robust and generalizable solution for medical image classification using self-supervised learning.
• Class decomposition effectively improves feature learning and diagnostic accuracy.
• The model's explainability component aids in understanding classification decisions and feature importance.