Classification of adenoid cystic carcinoma in whole slide images by using deep learning

Abstract

Adenoid cystic carcinoma (ACC) is a malignant tumor with poor prognosis, and it is prone to relapse locally and metastasize to distant sites after repeated recurrences. In the clinical diagnosis of ACC, it is often confused with basal cell adenoma (BCA). Meanwhile, the problems of the existing methods are insufficient accuracy and time-consuming, which are not conducive to clinical applications. To tackle these issues, an improved lightweight network termed adaptive feature fusion MobileNet (AFFMNet) is proposed to classify ACC and BCA. AFFMNet applies adaptive feature fusion to preserve discriminative information in both shallow-level and deep-level, attention mechanism is explored to reveal the global information by focusing on the lesion area. AFFMNet preforms patch-level classification, and then heatmaps are generated for representing whole slide images (WSIs). After that, the WSI-level classification for ACC and BCA is achieved by feature aggregation and XGBoost classifier. Experiments are conducted on the Chongqing University Cancer Hospital Adenoid Cystic Carcinoma Dataset, it shows the remarkable performance of the proposed AFFMNet with 97.37% accuracy on patch-level and 92.91% accuracy on WSI-level, which is more accurate than the same magnitude lightweight network MnasNet, and less inference time than EfficientNet. AFFMNet is an effective tool for WSIs, and it is conducive to clinical applications.