Table 2_Enhancing head and neck cancer detection accuracy in digitized whole-slide histology with the HNSC-classifier: a deep learning approach.xlsx

Head and neck squamous cell carcinoma (HNSCC) represents the sixth most common cancer worldwide, with pathologists routinely analyzing histological slides to diagnose cancer by evaluating cellular heterogeneity, a process that remains time-consuming and labor-intensive. Although no previous studies have systematically applied deep learning techniques to automate HNSCC TNM staging and overall stage prediction from digital histopathology slides, we developed an inception-ResNet34 convolutional neural network model (HNSC-Classifier) trained on 791 whole slide images (WSIs) from 500 HNSCC patients sourced from The Cancer Genome Atlas (TCGA) Head and Neck Squamous Cell dataset. Our pipeline was designed to distinguish cancerous from normal tissue and to predict both tumor stage and TNM classification from histological images, with the dataset split at the patient level to ensure independence between training and testing sets and performance evaluated using comprehensive metrics including receiver operating characteristic (ROC) analysis, precision, recall, F1-score, and confusion matrices. The HNSC-Classifier demonstrated exceptional performance with areas under the ROC curves (AUCs) of 0.998 for both cancer/normal classification and TNM system stage prediction at the tile level, while cross-validation showed high precision, recall, and F1 scores (>0.99) across all classification tasks. Patient-level classification achieved AUCs of 0.998 for tumor/normal discrimination and 0.992 for stage prediction, significantly outperforming existing approaches for cancer stage detection. Our deep learning approach provides pathologists with a powerful computational tool that can enhance diagnostic efficiency and accuracy in HNSCC detection and staging, with the HNSC-Classifier having potential to improve clinical workflow and patient outcomes through more timely and precise diagnoses, serving as an automated decision support system for histopathological analysis of HNSCC.

头颈部鳞状细胞癌（Head and Neck Squamous Cell Carcinoma, HNSCC）是全球第六大常见恶性肿瘤，病理医师常规通过评估细胞异质性分析组织切片以完成癌症诊断，但该过程耗时且费力。此前尚无研究将深度学习技术系统性应用于从数字化病理切片中自动化完成HNSCC的TNM分期与总体分期预测。本研究构建了一款基于Inception-ResNet34的卷积神经网络模型（HNSC-Classifier），其训练数据源自癌症基因组图谱（The Cancer Genome Atlas, TCGA）头颈部鳞状细胞数据集的791张全切片图像（Whole Slide Image, WSI），对应500名HNSCC患者。本研究的分析流程旨在区分癌组织与正常组织，并从组织病理学图像中预测肿瘤分期与TNM分类；数据集以患者为单位进行划分，以确保训练集与测试集相互独立，模型性能通过多项综合指标进行评估，包括受试者工作特征（Receiver Operating Characteristic, ROC）分析、精确率、召回率、F1分数以及混淆矩阵。HNSC-Classifier展现出优异性能，在图像块（tile）层面的癌/正常组织分类与TNM系统分期预测任务中，受试者工作特征曲线下面积（Area Under the ROC Curve, AUC）均达到0.998；交叉验证结果显示，所有分类任务的精确率、召回率与F1分数均高于0.99。患者层面的分类任务中，肿瘤/正常组织鉴别任务的AUC为0.998，分期预测任务的AUC为0.992，显著优于现有癌症分期检测方法。本深度学习方法可为病理医师提供一款强大的计算工具，能够提升HNSCC检测与分期诊断的效率与准确性；HNSC-Classifier有望通过更及时精准的诊断优化临床工作流程并改善患者预后，可作为HNSCC组织病理学分析的自动化决策支持系统。