Plataforma web de anotações interativas para segmentação de neuroimagens por ressonância magnética

Abstract

The growing volume and complexity of medical imaging, especially magnetic resonance imaging, have increasingly overwhelmed healthcare professionals and highlighted the need for computer-aided diagnostic tools. In this context, this work presents the development of an interactive web platform for the segmentation of magnetic resonance neuroimages, integrating artificial intelligence techniques and accessible visualization. The tool was designed with a modular architecture composed of a Flutter Web interface, a Flask-based API, and deep learning models implemented in PyTorch. The system performs automatic segmentation using the U-Net architecture and enables dataset enhancement through mask generation using the Flood Fill algorithm, validated by a CNN. The interface provides features for image upload, region of interest marking, result visualization, and retraining of the model with customized data. Load and stress tests were conducted to assess system performance, along with both quantitative and qualitative analyses of the segmentations. The U-Net segmentation model achieved a mean Dice Score of 90.22% , while the CNN for mask validation obtained 97.49% accuracy. These results, combined with the API’s high success rate under stress, demonstrate the feasibility of applying the platform in clinical and research environments, highlighting its flexibility, adaptability, and integration with medical workflows.