The devil is in the details: a small-lesion sensitive weakly supervised learning framework for prostate cancer...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Prostate cancer (PCa) diagnosis relies on time-consuming, subjective histopathology.
• High-resolution whole slide images (WSIs) present challenges for AI in detecting small tumors.
• Lack of fine-grained annotations hinders AI performance on difficult PCa cases.

Purpose of the Study:

• To develop an AI framework for accurate and efficient prostate cancer detection and grading.
• To improve the recognition of small and difficult-to-detect tumor regions in WSIs.
• To address limitations of existing cropping methods that can damage tumor structures.

Main Methods:

• Proposed an Intensive-Sampling Multiple Instance Learning Framework (ISMIL).
• ISMIL focuses on intensive sampling of crucial regions within WSIs.
• Utilized deep learning with a focus on multiple instance learning for improved small tumor recognition.

Main Results:

• Achieved an AUC of 0.987 on the PANDA dataset for prostate cancer detection.
• Improved recall by at least 33% with higher specificity for hard cases.
• Demonstrated comparable performance to human experts in prostate cancer grading and robustness in independent cohorts.

Conclusions:

• ISMIL effectively improves the recognition of small tumor regions in prostate cancer diagnosis.
• The framework shows high accuracy, specificity, and robustness, outperforming current methods on challenging cases.
• ISMIL is a potential tool to enhance diagnostic efficiency and accuracy for pathologists.