Automated Detection and Quantitative Assessment of Extranodal Extension in Lymph Node Breast Tumors Using Dual-Tier Deep Learning and Density-Based Contour Algorithms

Accurate detection and quantification of extranodal extension (ENE) in lymph node metastases is crucial for effective tumor staging and prognosis. In this paper, we propose a novel dual-tier deep learning-based computational method to automate the assessment of ENE in breast cancer lymph node sections. The method integrates an improved deep learning namely DenseNet-DMSCAMP that was designed for histological cell recognition, to classify cellular regions at the histological level into multiple classes—fat-only (adipose tissue), fat-and-skin (adipose and epithelial cells), skin-only (epithelial cells), and skin-and-immune (immune cells)—and applies a density-based active contour tracing (DB-ACT) algorithm, akin to DBScan, for lymph node boundary detection. Additionally, protrusion measurement algorithms are used to quantify the extent of tumor cell invasion beyond the lymph node capsule. The approach addresses the limitations of manual inspection, which can lead to misdiagnosis due to the visual complexity of whole-slide images (WSIs). Our method was validated on the Camelyon17 Grand Challenge dataset, which comprises H&E stained whole-slide images of lymph node sections. Experimental results showed a discrepancy of 3.9% to 11.7% in estimating ENE length and a position accuracy for breached capsule gaps ranging from 0.27% to 8.6%, demonstrating performance comparable to experienced histologists. These findings highlight the potential of our automated ENE evaluation system to improve diagnostic precision and alleviate the labor-intensive task of manual inspection in clinical settings.