POTENTIAL AND LIMITATIONS OF NNU-NET FOR INTRACRANIAL HEMORRHAGE SEGMENTATION ON CT SCANS

Segmentation of Intracranial Hemorrhages in Computed Tomography: Performance and Challenges of nnU-Net Architecture Variants

Resumo

The segmentation of intracranial hemorrhages in computed tomography scans constitutes a critical challenge in radiology, given the high morbidity and mortality associated with these conditions and the complexity of manual image interpretation. This study aimed to evaluate the performance of deep learning-based architectures, with a focus on nnU-Net applied to the segmentation of hemorrhagic lesions.

For this purpose, four public datasets (BHSD, MICCAI, Murtadha, and CQ500) were integrated, totaling 388 exams. In the CQ500 dataset, additional masks were created by medical students to address the absence of annotations. Preprocessing included 3D volumetric conversion and mask binarization, ensuring uniformity across the different datasets. Training was conducted for 250 epochs with 5-fold cross-validation, using the standardized nnU-Net pipeline in its default, medium, large, and extra-large variants. The evaluation was based on established metrics, such as the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), precision, sensitivity, and specificity, allowing for a comprehensive analysis of segmentation quality.

The results revealed modest differences among the variants, highlighting the Large configuration, which achieved a Dice score of 0.3940 and an IoU of 0.3189, associated with the highest precision (0.601). Meanwhile, the Default variant showed the highest sensitivity (0.711), indicating a better capability for detecting positive voxels, albeit with lower precision (0.544). These findings reflect the balance between architectural complexity and computational cost, consistent with recent studies that indicate incremental gains in more robust U-Net variants.

However, when comparing the results with the literature, it is noted that international studies have reported superior metrics (Dice > 0.90), which highlights the limitations of the present study. These can be attributed to the small size of the dataset, the variability among the datasets, and the absence of advanced strategies such as class balancing or attention mechanisms. Nevertheless, the study contributes by demonstrating the feasibility of applying nnU-Net in heterogeneous and realistic scenarios, highlighting the importance of a joint analysis of sensitivity and precision when choosing the most appropriate configuration for the clinical context.

In conclusion, although the obtained results are below the state of the art, the work reinforces the relevance of automated pipelines for intracranial hemorrhage segmentation. It also points to future prospects, including the expansion of datasets, the adoption of hybrid architectures with attention mechanisms, and the multimodal integration of clinical and imaging data, aiming for greater generalization and applicability in radiological practice.