Deep learning-based histopathological classification and subclassification of benign and malignant salivary gland tumors

Salivary gland tumors are rare and histopathologically diverse, presenting a diagnostic challenge. Accurate diagnosis is essential for optimal treatment planning. This study evaluates deep learning models in distinguishing benign from malignant salivary gland tumors using hematoxylin and eosin-stained histopathological images. 184 patients with the diagnosis of a salivary gland tumor (131 benign, 53 malignant) and six patients with normal salivary gland tissue were included. In total 335 histological sections were digitized. Deep learning models - including VGG19, ResNet50, Inception-ResNet-v2, Xception, ConvNeXt, and a Vision Transformer - were trained to classify benign tumors (Warthin tumor, pleomorphic adenoma) from malignant tumors (mucoepidermoid carcinoma, acinic cell carcinoma, adenoid cystic carcinoma, squamous cell carcinoma). The models’ performance in sub classifying benign tumors and differentiating them from normal tissue was also examined, as well as classification among the malignant tumors and normal tissue. All models showed strong performance in classifying benign versus malignant tumors, with the Xception model achieving 100% balanced accuracy. Corresponding confusion matrices indicated nearly error-free differentiation. Subclassification of the four malignant tumors against each other and normal tissue yielded balanced accuracies between 56.45% and 71.51%. Class activation maps revealed the models focused on stromal components in benign tumors and on cell nuclei and cytoplasm in malignant tumors. Deep learning models show great potential for the automated classification of salivary gland tumors as benign or malignant. Our findings suggest that deep learning models could be a valuable tool in clinical practice, particularly in the differentiation between benign and malignant lesions. However, limitations in malignant subtyping and the restricted number of malignant cases warrant further validation in larger, multi-institutional datasets.