Segmentation of Brain Tumors Using DeepLabv3+

Abstract

Multi-modal MRI scans are commonly used to grade brain tumors based on size and imaging appearance. As a result, imaging plays an important role in the diagnosis and treatment administered to patients. Deep learning based approaches in general, and convolutional neural networks in particular, have been utilized to achieve superior performance in the fields of object detection and image segmentation. In this paper, we propose to utilize the DeepLabv3+ network for the task of brain tumor segmentation. For this task, we build 18 different models using various combinations of the T1CE, FLAIR, T1 and T2 images to identify the whole tumor, the tumor core and the enhancing core of the brain tumor for the testing and validation data sets. We use the MICCAI BraTS training data, which consists of 285 cases, to train our network. Our method involves the segmentation of individual slices in three orientations using 18 different combinations of slices and a majority voting-based combination of the results of some of the classifiers that use the same combination of slices, but in different orientations. Finally, for each of the three regions, we train a separate model, which uses the results from the 18 classifiers as its inputs. The outputs of the 18 models are combined using bit packing to prepare the inputs to the final classifiers for the three regions. We achieve mean Dice coefficients of 0.7086, 0.7897 and 0.8755 for the enhancing tumor, the tumor core and the whole tumor regions respectively.