A Deep Adaptive Convolutional Network for Brain Tumor Segmentation from Multimodal MR Images

Abstract

In recent times, Convolutional Neural Networks (ConvNets) have gained immense popularity as the machine learning approach implemented in various applications including medical image analysis. In this work, we propose an automatic and adaptive convolutional neural network architecture segmenting brain tumor out of the Multimodal Magnetic Resonance Images with greater efficacy. The proposed model has the capability to choose the local as well as the global features from the image data automatically using the interacting subpaths of different lengths available in the architecture. The model finally produces the segmented part by avoiding overfitting and making use of dropouts whenever required. Extensive experiments on the popular benchmark dataset of brain tumor segmentation (MICCAI BraTS 2018) from 3D magnetic resonance (MR) images corroborated the efficacy of the proposed architecture both visually and with respect to some of the well-known evaluation metrics. The proposed framework yields an overall Dice coefficient, Mean Intersection over Union, Sensitivity and Specificity of 92%, 88%, 94% and 99% respectively which significantly exceeds the score of some other state-of-the-art methods.