Deep learning approaches for multiple sclerosis lesion segmentation using multi-sequence 3D MR images

Abstract

Multiple Sclerosis (MS) is a chronic inflammatory, immune-mediated, neurodegenerative, and demyelinating disease that impacts the Central Nervous System (CNS). The disease can cause permanent damage or deterioration (demyelination) to the nerves in the CNS. This damage results in the formation of lesions or plaques in the nervous system, leading to a wide range of symptoms such as problems with vision, loss of coordination, muscle weakness, and cognitive impairment. Early diagnosis and monitoring of MS are crucial since diagnosing the disease in its advanced stages can be more challenging. Therefore, effective methods for diagnosing and monitoring MS in its early stages are needed to improve patient quality of life and treatment outcomes. Magnetic Resonance Imaging (MRI) is widely used for monitoring, measuring, detecting, and characterizing MS lesions. T1-weighted (T1-w), T2-weighted (T2-w), and Fluid-Attenuated Inversion Recovery (FLAIR) sequences are commonly exploited in MS diagnosis as they provide different information about the brain tissues and the presence of lesions. Thereby, MRI is a useful tool for diagnosing and monitoring MS. Recently, Deep Learning (DL) methods have achieved remarkable results in the automated segmentation of MS lesions from MRI data, potentially improving the accuracy and efficiency of MS diagnosis and monitoring. Although automated methods for MS lesion segmentation have usually been performed on individual MRI scans, tracking lesion activity for quantifying and monitoring MS disease progression, especially detecting new lesions, has become an important biomarker in recent years. This Ph.D. thesis aims to develop novel and fully automated DL approaches for detecting and segmenting MS lesions from a single time-point brain MRI of a patient and also new MS lesions between two time points brain MRI of a patient. DL techniques simplify the feature extraction process from the given input data. Therefore, in this thesis, DL approaches were investigated and examined, then exploited to improve the segmentation and detection of MS lesions for both challenging tasks. Accordingly, a novel dense residual U-Net model that combines Attention Gate (AG), Efficient Channel Attention (ECA), and Atrous Spatial Pyramid Pooling (ASPP) is proposed to enhance the performance of the automatic MS lesion segmentation using 3D MRI sequences. Similarly, a unique pipeline with a deep neural network that combines U-Net, attention gate, and residual learning is proposed to perform better MS new lesion segmentation using baseline and follow-up 3D FLAIR MR images for lesion activity determination. In the proposed novel dense residual U-Net model, convolution layers in each block of the U-Net architecture are replaced by residual blocks and connected densely. Then, AGs are exploited to capture salient features passed through the skip connections. The ECA module is appended at the end of each residual block and each downsampling block of U-Net. Later, the bottleneck of U-Net is replaced with the ASSP module to extract multi-scale contextual information. Furthermore, 3D MR images of FLAIR, T1-w, and T2-w are exploited jointly to perform better MS lesion segmentation. The proposed model is validated on the publicly available ISBI2015 and MSSEG2016 challenge datasets. This model produced an ISBI score of 92.75, a mean Dice score of 66.88%, a mean Positive Predictive Value (PPV) of 86.50%, and a mean Lesion-Wise True Positive Rate (LTPR) of 60.64% on the ISBI2015 testing set. Also, it achieved a mean Dice score of 67.27%, a mean PPV of 65.19%, and a mean sensitivity of 74.40% on the MSSEG2016 testing set. The results show that the proposed model performs better than the results of some experts and some of the other state-of-the-art methods realized related to this particular subject. Specifically, the best Dice score and the best LTPR are obtained on the ISBI2015 testing set by using the proposed model to segment MS lesions. On the other hand, the generated model for the lesion activity determination within the proposed pipeline has a similar architecture to U-Net and is formed from residual units which facilitate the training of deep networks. Networks with fewer parameters are designed with better performance through the skip connections of U-Net and residual units, which facilitate information propagation without degradation. AGs also learn to focus on salient features of the target structures of various sizes and shapes. The MSSEG-2 challenge dataset was used for training and testing the proposed pipeline, and the results were compared with those of other proposed pipelines of the challenge and experts who participated in the same challenge. According to the results obtained from the testing set, the lesion-wise F1 and Dice scores were obtained as a mean of 48 and 44.30%. For the no-lesion cases, the number of tested and volume of tested lesions were obtained as a mean of 0.148 and 1.488, respectively. The proposed pipeline outperformed 22 proposed pipelines and ranked 8th in the challenge for the Dice and F1 scores. It was also ranked 4th and 5th for the number of tested and volume of tested lesions, respectively.