Deep learning-based context aggregation network for tumor diagnosis

Abstract

Craniopharyngioma is a type of benign brain tumor but has severe biological malignant behavior. Whether the craniopharyngioma invades the surrounding brain tissue has important influence on making treatment plan and the prognosis of patients, so the accurate diagnosis of craniopharyngioma is a crucial step in the treatment processing. It is important to explore some methods for judging the invasiveness of craniopharyngioma preoperatively. Therefore, we proposed a context aggregation network (CA-2D Network) based on deep learning algorithm, which can diagnose the invasiveness of craniopharyngioma by judging the characteristics of head MRI images. The proposed CA-2D Network utilizes ResNet as the backbone, and has a context modeling block and feature aggregating head to correlate features from different slices, capture context information, and aggregate features for classification. The features extracted by the CA-2D Network yield area under the curve (AUC) values of 82.59% for the test set. As demonstrated in the results, the proposed CA-2D Network is promising.