Application of visual prompts in the domain of medical image processing BRATS

Peter Bartoš

Supervisor(s): Igor Janos

Slovak Technical University

Abstract: Transformer-based architectures, originally developed for natural language processing, have recently demonstrated significant potential in image analysis, surpassing traditional convolutional neural networks in many tasks. In this work, we investigate their application to brain tumor segmentation in medical imaging, a domain where annotated data is often limited. We propose a two-stage segmentation pipeline that incorporates visual prompting, allowing the model to leverage intermediate guidance derived from initial predictions to improve final segmentation accuracy. The approach is evaluated on publicly available brain tumor datasets, demonstrating improved performance in capturing tumor boundaries and fine structures compared to baseline methods. Our results highlight the effectiveness of combining transformers with visual prompting for medical image segmentation, particularly in low-data scenarios, and suggest a promising direction for future research in automated medical image analysis.
Keywords: Computer Vision
Full text:
Year: 2026