A Novel Multi-Perspective Deep Learning Associated Magnetic Resonance Imaging for Multi-Parametric Disease Diagnosis

The diagnostic reading of magnetic resonance imaging (MRI) plays a vital role in identifying complex and co-morbid pathological conditions; however, it is limited by the noise of modality and variability of orientation and subjectivity of radiological interpretation. This paper presents a multi-perspective deep learning architecture combining convolutional neural networks (CNNs) and vision transformers (ViT) to classify multi-parametric MRI sequences such as T1-weighted, T2-weighted, FLAIR, DWI and ADC maps to identify multiple labels of disease. Registered volumes are preprocessed and broken down into axial, sagittal and coronal views and multi-scale 2D and 3D patch generation is used to enable hierarchical learning. Each of the inputs is assigned a diagnostic relevance dynamically through a modality attention mechanism and multi-view representations are combined with a gated fusion module. The proposed model was tested on publicly available data on Kaggle and TCIA with a diagnostic accuracy of 96.8% and performed better than the traditional CNNs, ViTs, and multimodal baselines on several disease classes. The model also showed stable sensitivity to unusual conditions like necrosis and inflammation. The study identifies the possibilities of multi-perspective learning in making medical imaging more interpretable and generalizable. The framework can be deployed on edge devices such as Jetson Xavier, and used to provide a point-of-care clinical decision support. The directions in the future include the extension of the architecture of radiogenomic fusion and pathology-specific prognostic modeling.