Deep learning-based multiple- sclerosis 3D lesion segmentation using small data

Multiple Sclerosis (MS) is the most common autoimmune system disorder that causes lesions in the brain. Early lesion detection and accurate lesion tracking are critical for MS diagnosis and treatment planning. Manual analysis of Magnetic Resonance Images (MRI) is the standard procedure for detecting these lesions. However, it is time-consuming, expensive, and error-prone due to the subjectivity of human perception. To overcome these obstacles, an automated system for detecting lesions in a brain scan can be used to mitigate the issues mentioned above.

Deep learning has demonstrated outstanding performance on various computer vision tasks, including image segmentation; however, deep learning methods require a large amount of annotated data for training. One of the most significant obstacles to applying deep learning techniques in the medical domain is the scarcity of annotated data.

The purpose of this study is to demonstrate techniques for segmenting MS brain lesions using small training sets. This study is significant in two ways: it proposes methods for accurate MS lesion segmentation, and it presents strategies for deep network training on small datasets. This research aims to examine and propose three methods for addressing the issue of small datasets in medical image analysis.

In the first phase, we assess the effectiveness of feature extraction. We evaluate the efficiency of models that were pre-trained on other computer vision tasks and self- supervised learning techniques for assisting neural networks to extract discriminative features from data. We propose several self-supervision tasks and compare them to pre-trained networks to illustrate this point.

The second phase investigates the efficacy of multi-task learning and how training a network with auxiliary tasks can assist in overcoming the data scarcity problem. To investigate the effectiveness of multi-task learning, we develop an auxiliary task and propose multi-task architectures.

In the third phase, our concentration is on enriching training datasets. We propose two strategies to achieve this objective: one for data augmentation and another for data elimination. The former is concerned with expanding the size of the dataset. The latter is concerned with removing unnecessary data to allow the network to concentrate on critical information.