Deep Learning for Neuroimaging Segmentation with a Novel Data Augmentation Strategy

Abstract

Brain insults such as cerebral ischemia and intracranial hemorrhage are critical stroke conditions with high mortality rates. Currently, medical image analysis for critical stroke conditions is still largely done manually, which is time-consuming and labor-intensive. While deep learning algorithms are increasingly being applied in medical image analysis, the performance of these methods still needs substantial improvement before they can be widely used in the clinical setting. Among other challenges, the lack of sufficient labelled data is one of the key problems that has limited the progress of deep learning methods in this domain. To mitigate this bottleneck, we propose an integrated method that includes a data augmentation framework using a conditional Generative Adversarial Network (cGAN) which is followed by a supervised segmentation with a Convolutional Neural Network (CNN). The adopted cGAN generates meaningful brain images from specially altered lesion masks as a form of data augmentation to supplement the training dataset, while the CNN incorporates depth-wise-convolution based X-blocks as well as Feature Similarity Module (FSM) to ease and aid the training process, resulting in better lesion segmentation. We evaluate the proposed deep learning strategy on the Anatomical Tracings of Lesions After Stroke (ATLAS) dataset and show that this approach outperforms the current state-of-art methods in task of stroke lesion segmentation.