Potential and Limitations of nnU-Net for Intracranial Hemorrhage Segmentation on CT Scans

Abstract

Introduction: The segmentation of intracranial hemorrhages in computed tomography scans is a critical challenge in radiology due to high morbidity and mortality the complexity of manual interpretation. Deep learning emerges as a promising alternative, with the nnU-Net architecture standing out for its adaptability to different medical imaging domains. Objective: To evaluate the applicability of nnU-Net in the automatic segmentation of intracranial hemorrhages, analyzing the performance and limitations of its variants in heterogeneous clinical scenarios. Method: 388 CT exams from the BHSD, MICCAI, Murtadha, and CQ500 datasets were used, the latter being supplemented by 14 manual masks. Pre-processing included 3D volumetric conversion and binarization of the masks. The training was conducted over 250 epochs with 5-fold cross-validation, exploring the default, medium, large, and extra-large variants. The evaluation used Dice, IoU, precision, recall (sensitivity), and specificity metrics. Results: The variants showed similar performance. The Large configuration achieved the best overlap indices (Dice = 0.3940; IoU = 0.3189) and higher precision (0.601). The Default version, on the other hand, showed the highest recall (0.711), favoring lesion detection at the expense of precision (0.544). The findings highlight the balance between architectural robustness and computational cost. Conclusions: Although the results fall short of international benchmarks (Dice > 0.90), due to the limited size of the dataset and the absence of advanced techniques, the study confirms the viability of nnU-Net in heterogeneous scenarios. The work reinforces its potential as a radiological support tool and points towards future research with larger datasets, attention mechanisms, and hybrid architectures.