MAMILNet: advancing precision oncology with multi-scale attentional multi-instance learning for whole slide image...

Area of Science:

• Digital Pathology
• Artificial Intelligence in Oncology
• Deep Learning for Medical Imaging

Background:

• Whole Slide Image (WSI) analysis using deep learning faces challenges in generalizability, annotation, and multi-magnification integration.
• Current methods require labor-intensive patch-level annotation and struggle with diverse cancer types.

Purpose of the Study:

• To introduce MAMILNet, a multi-scale attentional multi-instance learning framework for WSI analysis.
• To overcome limitations in model generalizability, reduce annotation burden, and integrate multi-magnification data.

Main Methods:

• MAMILNet treats whole slides as 'bags' and patches as 'instances', utilizing attention mechanisms for improved generalizability.
• A multi-scale strategy aggregates predictions across magnifications to enhance accuracy.
• The framework eliminates the need for detailed patch-level labeling.

Main Results:

• MAMILNet demonstrated strong performance across 1171 cases and various cancer types.
• Achieved an AUC of 0.8872 and Accuracy of 0.8760 for breast cancer tumor detection.
• Obtained an AUC of 0.9551 and Accuracy of 0.9095 for lung cancer typing and AUC of 0.7358/Accuracy of 0.7341 for ovarian cancer therapy response prediction.

Conclusions:

• MAMILNet shows significant potential for advancing precision medicine and personalized treatment planning in oncology.
• The framework effectively addresses key challenges in WSI analysis, promising improved patient care.
• Successful application in breast, lung, and ovarian cancers highlights its broad utility.