BrainViT: Multi-Label Classification of Brain Pathologies using Vision Transformers (ViT) — 82a — Sainath Vaddi, Sony Reddy Gurram, and Neerajdattu Dudam, with advisors: Dr. Rodrigue Rizk, and Dr. KC Santosh

Accurate classification of brain pathologies is essential for diagnosing neurological diseases,  particularly when co-occurring disorders present subtle distinctions. Multi-label categorization  from medical imaging holds promises for facilitating precise diagnoses in such complex  scenarios. However, accurately identifying brain abnormalities remains a critical challenge. In  this work, we propose BrainViT, a novel solution to address these challenges through  simultaneous multi-label classification of brain pathology using Vision Transformers (ViT)  which identifies various brain abnormalities accurately from medical images. The proposed  model uses a multi-label vision transformer that exploits the advantages of the self-attention  mechanism, eliminating convolution operations commonly found in traditional deep-learning  models for disease detection. The model’s capability to simultaneously identify various brain  pathologies in a single pass distinguishes it from conventional methods, providing a more  holistic understanding of complex clinical scenarios. To enhance the robustness and  generalizability of our model, we leverage a diverse set of carefully chosen datasets  representing various brain pathologies. The datasets that we employed include brain tumors named pituitary, glioma, meningioma, and healthy brains. This selection covers a  comprehensive range of brain abnormalities, ensuring a realistic foundation for model training.  Experimental results demonstrate that our proposed model achieves a successful  classification of brain diseases with a confidence level of 97.6% in distinguishing and  categorizing complex brain pathologies from medical images. This represents a significant  advancement in the multi-label classification of brain pathologies and contributes to the  improvement of accurate and efficient diagnosis in the field of neuroimaging, offering a  valuable tool for clinical applications in neurology and radiology.

University of South Dakota
Dr. Rodrigue Rizk, and Dr. KC Santosh