Deep Learning-Driven Pathological Prediction of Lymph Node Metastasis in Patients with Head and Neck Squamous Cell Carcinoma Using Primary Whole Slide Images

TL;DR

This study uses AI and digital pathology images to predict cancer spread to lymph nodes in head and neck cancer patients, improving pre-surgery risk assessment.

Contribution

A deep learning model and nomogram were developed to predict lymph node metastasis using whole-slide tumor images and clinical data.

Findings

The deep learning model achieved AUCs of 0.821 (internal) and 0.730 (external) for predicting lymph node metastasis.

An integrated nomogram improved performance to AUCs of 0.865 (internal) and 0.786 (external).

The model showed clinical utility in guiding treatment decisions and reducing unnecessary surgeries.

Abstract

Lymph node metastasis is one of the most important factors affecting treatment decisions and survival in patients with head and neck squamous cell carcinoma. However, accurately identifying patients at high risk before surgery remains challenging. In this study, we used digital pathology images of primary tumors and artificial intelligence to predict whether cancer had spread to cervical lymph nodes. By analyzing whole-slide images with a deep learning model and combining the results with basic clinical information, we developed a prediction tool that provides individualized risk estimates. Our model showed reliable performance in both internal and external patient cohorts and demonstrated potential clinical value for guiding neck management. This approach may help reduce unnecessary surgical procedures while ensuring timely treatment for patients at high risk of lymph node metastasis.

Background/Objectives: Accurate preoperative prediction of cervical lymph node metastasis (LNM) in head and neck squamous cell carcinoma (HNSCC) remains a major clinical challenge. This study aimed to develop a deep learning-based whole-slide image (WSI) model and an integrated nomogram to improve individualized LNM risk stratification. Methods: A total of 355 formalin-fixed paraffin-embedded (FFPE) WSIs and 282 frozen WSIs from the TCGA-HNSC cohort, along with 329 FFPE WSIs from an external institutional cohort, were retrospectively analyzed. Tumor regions were annotated and tiled into standardized patches. A dual-stage multiple instance learning framework was applied to generate WSI-level predictions. A pathological risk score (path-score) was derived and combined with clinical variables to construct a predictive nomogram. Results: The WSI-level model outperformed patch-level classifiers, with the logistic regression-based model achieving area under the curve (AUC) values of 0.821 in the internal validation cohort and 0.730 in the external cohort. The path-score was independently associated with LNM. The integrated nomogram further improved discrimination, yielding AUCs of 0.865 and 0.786 in the internal and external cohorts, respectively. Calibration and decision curve analyses demonstrated good agreement and meaningful clinical benefit. Conclusions: This deep learning-driven pathology nomogram provides a robust and clinically applicable tool for preoperative prediction of cervical lymph node metastasis in HNSCC.

Figures