Vision Transformer Based Cross Disorder MRI Screening with Interpretable Saliency Guided Decisions
The diagnosis at an early stage for the neurological problems like Alzheimer's disease, epilepsy, and brain tumors plays a critical role in increasing the chances of recovery from these diseases. However, traditional CNN techniques used for identifying such neurological problems in MRI images suffer from the limitations of understanding the long-range dependencies and context relations among various kinds of brain diseases. This paper proposes ViT-CrossMRI, which uses the vision transformer to perform cross-disorder MRI analysis and generates saliency guided by interpretations. It makes decision-making processes more clinically meaningful and transparent. The model was tested experimentally using multi-center heterogeneous datasets, where superior results were obtained in terms of several different measures, such as accuracy (93.12%), precision (92.76%), recall (92.30%), F1-score (92.53%) and AUC-ROC (95%). Balanced sensitivities and specificities of 92.30% and 93.45%, respectively, indicate that both positive and negative cases are identified effectively. Compared to baseline models such as CNN, ResNet-50, and LSTM, this approach demonstrates superiority in relation to the general ability to use global context and attention in cross-disorder diagnosis. From the computational point of view, reasonable inference time (145 ms for each MRI scan) and low parameter number (86.2M) and memory usage (6.8 GB) can be noted.