Artificial Intelligence-Based Breast Cancer Nodal Metastasis Detection: Insights Into the Black Box for Pathologists

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Histologic identification of lymph node metastasis is crucial for cancer staging and treatment decisions.
• Manual review of lymph node biopsies can be time-consuming and prone to errors, particularly for small tumor deposits.

Purpose of the Study:

• To evaluate the clinical application and implementation of the LYmph Node Assistant (LYNA), a deep learning-based artificial intelligence algorithm.
• To assess LYNA's performance in detecting metastatic breast cancer in sentinel lymph node biopsies.

Main Methods:

• Development and evaluation of LYNA using whole slide images from hematoxylin-eosin-stained lymph nodes (Camelyon16 dataset).
• Testing on a dataset of 399 patients (270 for development, 129 for evaluation).
• Reproducibility assessment using an independent laboratory dataset (108 slides).

Main Results:

• LYNA achieved a slide-level AUC of 99% and tumor-level sensitivity of 91% with one false positive per patient on the evaluation dataset.
• The algorithm demonstrated high performance (AUC 99.6%) on an independent dataset and was robust to common histology artifacts.
• LYNA successfully identified micrometastases on slides initially deemed "normal".

Conclusions:

• Deep learning algorithms like LYNA can achieve high sensitivity and comparable slide-level performance to pathologists in detecting lymph node metastasis.
• AI tools may enhance pathologist productivity and decrease false negatives in cancer diagnosis.
• A framework is proposed for pathologists to assess and adopt AI algorithms into their diagnostic workflow.