Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Biomedical imaging analysis

Background:

• Traditional image-based survival prediction models face scalability challenges with large datasets due to patch labeling.
• Multiple Instance Learning (MIL) shows promise for histopathological image classification without annotations.

Purpose of the Study:

• To develop a scalable and interpretable framework for cancer survival prediction using whole slide images (WSIs).
• To introduce an attention-based Multiple Instance Learning approach for efficient feature learning and aggregation from WSIs.

Main Methods:

• Proposed Deep Attention Multiple Instance Survival Learning (DeepAttnMISL) framework.
• Incorporated siamese MI-FCN and attention-based MIL pooling for feature extraction and aggregation.
• Evaluated on two large cancer whole slide image datasets.

Main Results:

• DeepAttnMISL demonstrated superior effectiveness and suitability for large datasets compared to existing methods.
• The attention-based aggregation proved more flexible and adaptive.
• Achieved better interpretability in identifying key features for survival prediction.

Conclusions:

• DeepAttnMISL provides a scalable and interpretable solution for cancer survival prediction from WSIs.
• The framework aids in assessing individual patient risk, supporting personalized medicine.
• Highlights the potential of attention-based MIL in computational pathology.