TDT-MIL: a framework with a dual-channel spatial positional encoder for weakly-supervised whole slide image...

Area of Science:

• Computational Pathology
• Digital Health
• Machine Learning in Medicine

Background:

• Whole slide image (WSI) classification is crucial for digital pathology.
• Traditional multiple instance learning (MIL) often overlooks spatial relationships between positive tissues in WSIs.
• Accurate classification is challenging due to the small percentage of positive tissues within billions of pixels.

Purpose of the Study:

• To propose a novel framework, TDT-MIL, for weakly-supervised WSI classification.
• To address the overlooked spatial positional relationships crucial for accurate WSI analysis.
• To develop a model capable of handling imbalanced WSI classification tasks effectively.

Main Methods:

• Utilized a serial connection of a convolutional neural network and transformer for feature extraction.
• Introduced a dual-channel spatial positional encoder (DCSPE) to capture local and global positional information.
• Incorporated a convolutional triple-attention (CTA) module to enhance inter-channel information mining.

Main Results:

• Achieved high classification accuracy and AUC on CAMELYON16 and TCGA-NSCLC datasets (up to 91.54%, 94.96%, 90.21%, 94.36%).
• Outperformed state-of-the-art baseline methods in WSI classification.
• Demonstrated satisfactory performance on imbalanced WSI classification tasks.

Conclusions:

• TDT-MIL effectively mines spatial positional and inter-channel information for pathological semantics in WSIs.
• The proposed framework offers an interpretable yet ingenious solution for WSI classification.
• TDT-MIL shows significant potential for advancing weakly-supervised learning in digital pathology.