Linha de pesquisa: Ciência de Dados

Resumo: Brain tumors are among the most life-threatening diseases, and early detection through Magnetic Resonance Imaging (MRI) is critical for patient outcomes. This work re ports partial progress toward an attention-driven deep learning pipeline for brain tu mor recognition: a UNet-based segmentation pipeline that jointly classifies brain tu mors and produces pixel-wise tumor masks and bounding boxes in a single forward pass. The pipeline integrates Cellular Automata (CA) preprocessing, Genetic Algorithm (GA) hyperparameter optimization, and four encoder-decoder backbones spanning a range of computational cost (MobileNetV2-UNet, MobileNetV3-UNet, EfficientNet-UNet, and ResNet-UNet) that output a four-channel tumor mask covering background, glioma, meningioma, and pituitary classes; image-level classification and bounding boxes are de rived directly from this mask. On the Figshare brain tumor dataset (3,064 T1-weighted contrast-enhanced MRI slices), EfficientNet-UNet achieves the best segmentation quality among the four backbones (73.8% mean Dice, 59.0% mean Intersection over Union (IoU) over the tumor classes) with an 85.2% F1-score for segmentation-derived classification. A controlled CA/GA ablation shows that the two components do not combine addi tively: the untuned baseline attains the highest classification F1, while the combined CA+GA configuration yields the best Dice and IoU, the metric this pipeline prioritizes. Computational-cost profiling, reported in Floating-Point Operations (FLOPs) to keep the comparison independent of hardware and system load, further shows MobileNetV3-UNet is the lightest backbone (7.3GFLOPs) and ResNet-UNet the heaviest by roughly 43 times (314.2GFLOPs), a distinct efficiency trade-off. These results establish a validated seg mentation and interpretability baseline for the originally proposed attention-driven DETR detector, with explicit spatial, channel, and hybrid attention evaluated via Common Ob jects in Context (COCO)-style metrics, which remains ongoing work.

Link para a defesa: https://teams.microsoft.com/dl/launcher/launcher.html?url=%2F_%23%2Fmeet...