Deep learning approach to identify histological features associated with lymph node metastasis following primary tumor excision in patients with tongue squamous cell carcinoma.

Abstract

Objective: To assess whether a semi-automated deep learning (DL) detector that quantifies poorly differentiated nests on hematoxylin-eosin (HE) sections is associated with cervical lymph node (LN) metastasis in tongue squamous cell carcinoma (SCC), and to explore postoperative risk stratification in clinically node-negative early-stage disease.

Study design: Retrospective single-center study of 115 tongue SCC patients (1998-2016) with ≥5-year follow-up. A Faster region-based convolutional neural network detector quantified poorly differentiated nests at the invasive front. Mean nest counts were compared between LN-positive and LN-negative cases and evaluated by receiver operating characteristic (ROC) analysis. The ROC cut-off was explored in an independent cohort of 20 cT1-T2 cN0 cases without elective neck dissection.

Results: LN-positive cases had higher poorly differentiated nest counts than LN-negative cases. The mean count yielded an area under the curve of 0.67 for discriminating cervical LN metastasis confirmed at initial treatment or during follow-up. In the independent cohort, the cut-off (≥3.6 nests per case) showed 72.7% sensitivity and 55.6% specificity, with higher sensitivity but lower specificity than Yamamoto-Kohama mode of invasion.

Conclusions: DL-based nest quantification on routine HE sections may aid postoperative risk stratification for cervical LN metastasis in tongue SCC.