A multi-resolution model for histopathology image classification and localization with multiple instance learning.

Area of Science:

• Digital pathology
• Computational image analysis
• Machine learning in histopathology

Background:

• Whole slide imaging (WSI) generates large datasets, offering potential for computational analysis to aid pathologists.
• Previous WSI analysis often requires detailed region annotations, limiting scalability for large datasets.
• Developing automated tools is crucial for reducing workload and improving diagnostic consistency in pathology.

Purpose of the Study:

• To develop an end-to-end computational model for prostate cancer grading using whole slide images.
• To overcome the limitations of fine-grained annotations by utilizing slide-level labels.
• To leverage saliency maps for detecting suspicious regions within whole slide images.

Main Methods:

• Proposed a multi-resolution multiple instance learning (MIL) model.
• Integrated saliency maps to guide the detection of critical regions for grade prediction.
• Trained the model using only slide-level labels, eliminating the need for pixel- or region-level annotations.

Main Results:

• Achieved 92.7% accuracy and 81.8% Cohen's Kappa for predicting benign, low-grade, and high-grade prostate cancer.
• Demonstrated high performance in differentiating malignant and benign slides with an AUROC of 98.2% and AP of 97.4%.
• Validated cancer detection on an external dataset with an AUROC of 99.4% and AP of 99.8%.

Conclusions:

• The developed MIL model effectively predicts prostate cancer grade from whole slide images using slide-level labels.
• This approach significantly reduces the annotation burden and shows high accuracy in clinical settings.
• The model demonstrates strong potential for improving diagnostic efficiency and accuracy in digital pathology.