Disparity Autoencoders for Multi-class Brain Tumor Segmentation

Abstract

AbstractMulti-class brain tumor segmentation is important for predicting the aggressiveness and treatment response of gliomas. It has various applications including diagnosis, monitoring, and treatment planning of gliomas. The purpose of this work was to develop a fully automated deep learning framework for multi-class brain tumor segmentation. Brain tumor cases with multi-parametric MR Images from the RSNA-ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge 2021 were used. Six Disparity Autoencoders (DAE) were developed including 2 DAEs to segment the whole-tumor (WT), 2 DAEs to segment the tumor-core (TC) and 2 DAEs to segment the enhancing-tumor (ET). The output segmentations of a particular label from their respective DAEs were ensembled and post-processed. The DAEs were tested on the BraTS2021 validation dataset. The networks achieved average dice-scores of 0.90, 0.80 and 0.79 for WT, TC and ET respectively on the validation dataset and 0.89, 0.82, 0.81 for WT, TC and ET respectively on the test dataset. This framework could be implemented as a robust tool to assist clinicians in primary brain tumor management and follow-up.KeywordsBrain tumor segmentationGlioma segmentationDeep learningBraTSAutoencodersMRIImaging features