Benchmarking multiple instance learning architectures from patches to pathology for prostate cancer detection and...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Prostate cancer diagnosis and grading are crucial but traditionally time-consuming and variable.
• Current AI systems require extensive pixel-level annotations, limiting clinical utility.
• Automated systems need to be scalable, interpretable, and accurate for widespread adoption.

Purpose of the Study:

• To develop and evaluate a weakly supervised deep learning framework for automated prostate cancer diagnosis and International Society of Urological Pathology (ISUP) grading.
• To minimize annotation requirements and enhance interpretability in AI-driven histopathology.
• To conduct a large-scale benchmarking study of multiple instance learning (MIL) architectures and configurations.

Main Methods:

• Benchmarking six MIL architectures, three feature encoders, and four patch extraction techniques on the PANDA dataset (10,616 whole slide images).
• Utilizing distributed cloud computing to process over 31 million tissue patches.
• Implementing attention mechanisms and Grad-CAM for interpretability.

Main Results:

• The optimal configuration (UNI2 encoder with ILRA-MIL) achieved 78.75% accuracy and 90.12% quadratic weighted kappa (QWK).
• Overlapping smaller patches and domain-specific foundation models yielded superior performance.
• The approach demonstrated scalability, feasibility, and interpretability.

Conclusions:

• Weakly supervised deep learning offers a scalable and interpretable solution for prostate cancer diagnosis and grading.
• This framework approaches expert pathologist diagnostic capability, addressing limitations of traditional methods.
• The study provides the first large-scale comparison of weakly supervised MIL methods for this application.