Semi-supervised lung adenocarcinoma histopathology image classification based on multi-teacher knowledge distillation

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Accurate classification of lung tumor growth patterns is crucial for treatment but challenging with limited labeled whole slide images (WSIs).
• High demand for labeled data in medical image analysis hinders the application of deep learning (DL).

Purpose of the Study:

• To develop a semi-supervised learning (SSL) scheme using a patch-based DL framework for high-precision classification of seven lung tumor growth patterns.
• To enhance generalization ability and reduce reliance on extensive labeled data in WSI analysis.

Main Methods:

• Implemented an SSL approach with a dynamic confidence threshold mechanism to manage pseudo-label quality and quantity.
• Utilized multi-teacher knowledge distillation (MTKD) to transfer knowledge from multiple teacher models, protecting student models from poor predictions.
• Trained and evaluated the framework on 150 WSIs representing seven growth patterns.

Main Results:

• The SSL framework achieved high accuracy in classifying lung tumor growth patterns in histopathology images.
• Performance was comparable to fully supervised models and human pathologists.
• Demonstrated superior generalizability on a public dataset compared to previous studies.

Conclusions:

• SSL approaches can achieve expert-level performance in medical image analysis, offering efficient and cost-effective solutions.
• Dynamic confidence thresholding and MTKD represent significant advancements for DL in complex medical image analysis.
• This work paves the way for faster, more accurate diagnoses and improved patient outcomes.