Independent assessment of a deep learning system for lymph node metastasis detection on the Augmented Reality...

Area of Science:

• Digital Pathology
• Machine Learning in Oncology
• Surgical Pathology

Background:

• Machine learning algorithms show promise for cancer identification in digitized pathology slides.
• Augmented Reality Microscopes (ARM) integrate these algorithms into real-time pathology workflows.

Purpose of the Study:

• To independently assess LYmph Node Assistant (LYNA) models, optimized for ARM, for breast cancer metastasis identification in lymph node biopsies.
• To analyze model performance across various tissue types and magnifications.

Main Methods:

• Assessment of LYNA models on 40 whole slide images at 10×, 20×, and 40× magnifications.
• Performance analysis across tissue subclasses: breast cancer, lymphocytes, histiocytes, blood, and fat.
• Review of prediction discrepancies and introduction of 'proper' vs. 'improper' ground truth for uncertain annotations.

Main Results:

• Models achieved high overall performance with AUC ~0.98, accuracy ~0.94, and sensitivity >0.88.
• Highest accuracy observed in fat and blood; lowest in histiocytes, germinal centers, and sinus.
• Models struggled with isolated tumor cells, particularly at lower magnifications.

Conclusions:

• LYNA models demonstrate strong performance for breast cancer metastasis detection when integrated with ARM.
• Performance is influenced by tissue type and magnification, highlighting areas for model improvement.
• Novel methods for analyzing complex pathology data with uncertain ground truth were introduced.