Registration Based Data Augmentation for Multiple Sclerosis Lesion Segmentation

Abstract

Deep learning has shown outstanding performance on various computer vision tasks such as image segmentation. To take advantage of deep learning in image segmentation, one would need a huge amount of annotated data since deep learning models are data-intensive. One of the main challenges of using deep learning methods in the medical domain is the shortage of available annotated data. To tackle this problem, in this paper, we propose a registration based framework for augmenting multiple sclerosis datasets. In this framework, by registering images of two different patients, we create a new image, which smoothly adds lesions from the first patient into a brain image, structured like the second patient. Due to their nature, multiple sclerosis lesions vary in shape, size, location and number of occurrence, thus registering images of two different subjects, will create a realistic image. The proposed method is capable of introducing diversity to data distribution, which other traditional augmentation methods do not offer. To check the effectiveness of our proposed method, we compare the performance of 3D-Unet on different augmented and non-augmented datasets. Experimental results indicate that the best performance is achieved when combining both the proposed method with traditional augmentation techniques.