Multiple Sclerosis (MS) is a long-term autoimmune disease that disrupts the brain and central nervous system of people resulting in vision, sensory, and motor problems, that eventually ends up in nerve deterioration or permanent physical inability if untreated. To detect MS, several screening procedures have been recommended, the most popular amongst them is Magnetic Resonance Imaging (MRI). MRI modalities assist physicians in getting accurate details of the structure and operational activity of the brain, which will be crucial for detecting and treating MS in an early stage. Manual diagnosis of MS utilizing MRI modalities is a tedious, time consuming process and is susceptible to misconception. Deep learning Based Diagnosis Systems (DBDS) employs Artificial Intelligence to diagnose MS by employing several Deep Learning techniques. In Deep Learning (DL) frameworks, feature selection and extraction are done automatically by the designed layers which can automatically learn and extract the required features from the MRI modality. In this paper, a critical study of previous research works for diagnosing MS by employing several DBDS is discussed. The several types of pre-processing methods for MRI neuroimaging modalities were discussed. The various steps involved in the processing of MRI modalities by several DBDS models for MRI segmentation and classification are analyzed. Furthermore, the major challenges and future research opportunities in engaging DL architectures for MS diagnosis from MRI modalities are discussed.