A Semi-Supervised Learning Framework for Classifying Colorectal Neoplasia Based on the NICE Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Medical image labeling is resource-intensive, requiring expertise and large datasets.
• Insufficient labeled data hampers supervised learning model performance, leading to underfitting.

Purpose of the Study:

• To develop a SimCLR-based semi-supervised learning framework for classifying colorectal neoplasia using the NICE classification.
• To evaluate the framework's performance against supervised transfer learning and human endoscopists.

Main Methods:

• Trained a ResNet-backboned SimCLR model using self-supervised learning on unlabeled data.
• Fine-tuned the model on a limited labeled dataset for NICE classification.
• Evaluated performance using accuracy, Matthew's correlation coefficient (MCC), and Cohen's kappa, with Grad-CAM and t-SNE for visualization.

Main Results:

• The SimCLR model achieved high accuracy (0.908), MCC (0.862), and Cohen's kappa (0.896), surpassing supervised transfer learning and junior endoscopists.
• The model's performance was comparable to senior endoscopists.
• t-SNE visualization revealed superior clustering of samples with self-supervised learning compared to supervised transfer learning.

Conclusions:

• Semi-supervised learning, particularly with SimCLR, offers a powerful solution for deep learning in medical image analysis with limited labeled data.
• This framework demonstrates potential for improving the accuracy and efficiency of colorectal neoplasia classification.
• The study highlights the advantages of self-supervised pre-training for enhancing deep learning model interpretability and performance in endoscopic imaging.