Authors - Zala Bhargavi Harshadbhai, Priyank D. Doshi Abstract - Brain tumor classification using MRI is very important for early diagnosis. While convolutional neural networks (CNNs) showed strong performance in medical image analysis, but transformer-based architectures have recently gained popularity because of their ability to model long-range spatial dependencies through self-attention mechanisms. Our work lines up two such models - Vision Transformer and Swin Transformer to see how each handles tumor spot-ting in brain MRIs from the BRISC2025 collection. Same training setup applied to both keep things balanced and evaluated on the official test split for ensuring fairness. The official test set showed that both ViT (99.17 ± 0.26%) and Swin (99.27 ± 0.13%) have nearly identical predictive performance. Despite similar outcomes, their inner workings differ sharply behind the scenes. Swin Trans-former have approximately 40% and inference cost by nearly 50% compared to ViT while maintaining similar accuracy. The study provides insights into the performance and efficiency of trade-offs between global and hierarchical trans-former architectures in medical imaging applications.
Open Zoom