Classification of brain lesions from MRI images using a novel neural network

There is considerable interest in using convolutional neural networks for computer aided medical diagnostics. In this paper, we present some recent results on brain lesion segmentation. While algorithms have been developed to automate the process, the results often lack accuracy in tracing the lesions. Here, we present our results on the publically available ATLAS R1.1 dataset using deep CNN architectures. We utilize a deep U-net architecture with 3D convolutions, consisting of an encoder and a corresponding decoder along with dense residual connections between them, leveraging both low and high level features from the encoder at each step. We also analyze the use of Inception, Residual and Inception-Res blocks for the encoder. The architecture also makes use of an attention gate for the decoder part to suppress irrelevant parts of the incoming input while highlighting the salient features. The dataset had 239 3D MRI images of dimensions (197 x 233 x 189) with their corresponding segmentation maps. We used a train-validation-test split of 7:2:1, employing mean normalization, histogram equalization and suitable data augmentation techniques. We used SGD optimizer with focal Tversky loss function. The Dice score and AUC were used as metrics. We were able to achieve a Dice coefficient of 0.53 with an AUC around 0.95 with a single end-to-end model, unlike previous models that either focus only on one part of the brain or use other metadata, and matched their performance in every aspect. We further discuss the possibilities of improving the results through denoising and data cleansing using standard machine learning and computer vision methodologies.