LiteMIL: a computationally efficient cross-attention multiple instance learning for cancer subtyping on whole-slide...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Accurate cancer subtyping is crucial for precision medicine.
• Gigapixel whole-slide images (WSIs) present computational challenges for analysis.
• Existing transformer-based multiple instance learning (MIL) methods are computationally intensive, limiting clinical use.

Purpose of the Study:

• Introduce LiteMIL, a computationally efficient cross-attention MIL method for WSI classification.
• Optimize MIL for WSI classification to overcome computational demands.
• Enable clinical deployment of advanced MIL techniques for cancer subtyping.

Main Methods:

• Developed LiteMIL, a cross-attention MIL using a single learnable query for bag-level aggregation.
• Evaluated LiteMIL against five baseline methods on four TCGA datasets (breast, kidney, lung, TUPAC16).
• Utilized nested cross-validation with patient-level splitting and conducted systematic ablation studies.

Main Results:

• LiteMIL achieved competitive accuracy (83.5%), matching TransMIL.
• Demonstrated significant efficiency gains: 4.8x fewer parameters, 2.9x faster inference, and 6.7x lower GPU memory usage.
• Showcased task-dependent performance, with single-query optimality for focused attention and multi-query benefits for heterogeneous tasks.

Conclusions:

• LiteMIL offers a resource-efficient solution for WSI classification, suitable for consumer GPUs.
• The cross-attention architecture balances performance with computational efficiency for clinical integration.
• Task-dependent design insights guide practical implementation for diverse WSI classification scenarios.