Logo do repositório

Percorrer por autor "Ribeiro, Ana Filipa Ferreira"

A mostrar 1 - 1 de 1
  • Artificial intelligence applied to brain tumor classification in MRI scans
    Publication . Ribeiro, Ana Filipa Ferreira; Ribeiro, Maria Margarida do Carmo Pinto; Carreira, Fernando Paulo Neves da Fonseca Cardoso
    Abstract Gliomas are heterogeneous, difficult to delineate, and consume significant clinical time. This work prioritizes MRI preprocessing and a simple 2D U-Net, comparing pipelines that improve image quality instead of optimizing a full segmentation system under GPU/time constraints. We used multimodal BraTS 2023 data (FLAIR, T1Gd, and T2), excluding native T1 due to practical redundancy. Three methods were compared. Method 1 applied per-volume min–max normalization and geometric-center cropping. Method 2 replaced this with z-score normalization (computed only on intracranial/non-zero voxels) and centering guided by T1Gd intensity. Method 3 added bias-field correction (N4) and label-free localization based on FLAIR (Otsu + largest connected component), avoiding label leakage in testing/validation. In all cases, a simple 2D U-Net (MONAI) with three input channels and four classes was trained using Dice + cross-entropy loss and early stopping. Evaluation followed the BraTS subregions (ET, TC, WT) and standard metrics (Dice, IoU, sensitivity, precision, and specificity). Results show consistent gains for Method 3 over the others. On the test set, Method 3 achieved mean Dice of 0.626 (ET), 0.682 (TC), and 0.844 (WT), surpassing Method 2 by +0.073, +0.153, and +0.072, respectively. Improvements are most pronounced in the TC, the most challenging region. In conclusion, the intracranial normalization, bias-field correction, and FLAIR-guided centering increase input reliability and the stability of the 2D model, providing a solid foundation for future 3D extensions and external validation.
    2025-12-05Dissertação de mestrado Acesso aberto Ver mais