A computation-efficient network with feature aggregation for cancer subtype classification on histopathological...

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Histopathology whole-slide images (WSI) are crucial for cancer diagnosis, offering detailed tissue information.
• Deep learning (DL) aids WSI analysis but faces challenges with image size and complexity.
• Existing multiple instance learning (MIL) methods have limitations in pseudo-label accuracy and information loss.

Purpose of the Study:

• To develop a novel DL framework for accurate and efficient WSI classification.
• To address limitations of traditional MIL methods in handling WSI data.
• To enhance interpretability of DL models in histopathology.

Main Methods:

• A framework combining a lightweight convolutional neural network (CNN)-based tile encoder (CTE) and a Transformer-based feature aggregator (TFA).
• A two-stage training strategy: CTE pre-training and TFA fine-tuning for efficiency and accuracy.
• Dynamic self-attention mechanism in TFA for feature aggregation without pseudo-labels.

Main Results:

• The proposed method achieves higher classification accuracy compared to existing MIL approaches on three cancer datasets.
• The framework demonstrates computational efficiency by reducing costs and alleviating local information loss.
• Generated saliency maps highlight relevant regions, aligning model decisions with clinical reasoning.

Conclusions:

• The novel framework effectively classifies histopathology WSIs with improved accuracy and efficiency.
• The dynamic self-attention mechanism enhances global representation by integrating local information.
• The method offers a promising, interpretable tool for digital pathology and cancer subtype classification.