Comparison of four convolutional neural networks for histopathological diagnosis of salivary gland carcinomas

Abstract

Background: Malignant salivary gland tumors represent a particular diagnostic challenge due to the large number of histopathological entities, their rare occurrence, and the diverse clinical and histological presentations. The aim of this work is to investigate and compare convolutional neural networks (CNNs) as a diagnostic tool for histological diagnosis of salivary gland cancer. Methods: From salivary gland cancer preparations of 68 patients, 118 histological slides were digitized at high resolution. These virtual sections were then divided into small image sections, and the resultant 83,819 images were sorted into four categories: background, connective tissue, non-neoplastic salivary gland tissue, and salivary gland cancer tissue. The latter category grouped the entities adenoid cystic carcinoma, adenocarcinoma (not otherwise specified), acinar cell carcinoma, basal cell carcinoma, mucoepidermoid carcinoma, and myoepithelial carcinoma. The categorized images were then processed in a training, validation, and test run by the ImageNet pretrained CNN frameworks (Inception ResNet v2, Inception v3, ResNet152, Xception) in different pixel sizes. Results: Accuracy values ranged from 18.8% to 84.7% across all network architectures and pixel sizes, with the Inception v3 network achieving the highest value at 500 × 500 pixels. The recall values/sensitivity reached up to 85% for different pixel sizes (Inception v3 network at 1000 × 1000 pixels). The minimum F1 score achieved was 0.07 for the Inception ResNet v2 and the Inception v3 at 100 × 100 pixels each, the maximum F1 score achieved was 0.72 for the Xception at 1000 × 1000 pixels. Inception v3 was the network with the shortest training times, and was superior to all other networks at any pixel size. Conclusion: The current work was able to demonstrate the applicability of CNNs for histopathological analysis of salivary gland tumors for the first time and provide a comparison of the performance of different network architectures. The results indicate a clear potential benefit for future applications.