Implementação de um pipeline para parcelamento cerebral em imagens de ressonância magnética utilizando o framework MONAI

Abstract

In this project, a brain parcellation solution was implemented using the MONAI framework. The developed approach covers all stages of the pipeline, from preprocessing medical images to training, validating, and testing two distinct neural network architectures. The goal is to segment and label different brain regions from T1-weighted magnetic resonance images. To ensure image standardization before training, several preprocessing techniques were applied, including bias correction, intensity normalization, reorientation to the RAS system, and centering of brain structures. These steps ensure that the input data is in a consistent format, improving model generalization. The selected network for segmentation was HighResNet, chosen based on a study that demonstrated its effectiveness for brain parcellation tasks. The training was performed using a set of 1,200 labeled images, while validation and testing were conducted with 200 and 200 images, respectively, comparing this architecture with U-Net, one of the most widely used networks for 2D and 3D image segmentation. The network's performance was evaluated using the Dice coefficient, which measures the overlap between predicted segmentations and ground-truth labels. After training, the model weights were saved, and to facilitate the use of the solution, a FastAPI application integrated with Uvicorn was developed. This application enables model deployment, allowing both networks to be used as a service. The API supports image submission via POST requests and returns the volumetric image properly parcellated as a response. This approach ensures that the models can be accessed and utilized in a practical and efficient manner.