A semi-supervised multi-task learning framework for cancer classification with weak annotation in whole-slide images.

Area of Science:

• Digital Pathology and Computational Imaging
• Oncology and Cancer Diagnostics
• Machine Learning in Medical Image Analysis

Background:

• Cancer region detection (CRD) and subtyping are crucial for digital pathology, aiming to reduce pathologist workload and enhance diagnostic accuracy on whole-slide images (WSIs).
• Existing models face limitations including the need for large annotated datasets and confusion in subtyping due to the presence of non-cancerous regions.
• Previous methods of sequential CRD followed by subtyping ignore task interactions and propagate errors.

Purpose of the Study:

• To propose a novel semi-supervised multi-task learning (MTL) framework to concurrently improve cancer region detection (CRD) and subtyping.
• To address limitations of existing methods by capturing task interactions and enabling error back-propagation.
• To develop a framework that requires minimal annotations through a minimal point-based (min-point) strategy.

Main Methods:

• A semi-supervised multi-task learning (MTL) framework integrating a shared backbone feature extractor and two task-specific classifiers for CRD and subtyping.
• Implementation of a weight control mechanism to manage task interaction and error back-propagation.
• Training the framework using a minimal point-based (min-point) annotation strategy on datasets with limited annotations.

Main Results:

• The proposed MTL framework effectively captures the interaction between CRD and subtyping tasks.
• The weight control mechanism ensures proper error back-propagation, enhancing performance on both tasks.
• Extensive experiments on four large datasets demonstrated significant improvements in accuracy and generalization capabilities.

Conclusions:

• The developed semi-supervised MTL framework offers an effective solution for simultaneous CRD and subtyping in digital pathology.
• The min-point annotation strategy significantly reduces annotation effort while maintaining high performance.
• The framework shows strong potential for improving diagnostic efficiency and accuracy in cancer classification.