Answers are grounded in this paper and cite their sources.
Background This study aimed to assess the performance of a neural network in segmenting acute intracerebral hemorrhage (ICH) compared to the ABC/2 method, both benchmarked against reference standard (RS), and to identify imaging predictors associated with performance discrepancies. Methods A retrospective analysis was conducted using an independent test set of patients with acute ICH diagnosed on non-contrast CT from three academic hospitals in Germany and Italy (01/2016–06/2022). A pre-trained and validated nnU-Net model was employed for segmentation. Both nnU-Net and ABC/2 method segmentations were benchmarked against RS using Bland–Altman plots. Associations between volume deviations and 14 imaging markers were assessed using the Mann-Whitney U test, and inter-marker correlations with Chi-square tests. Marker reading reliability was evaluated using Cohen's kappa. Results Among 188 cases, nnU-Net's volume measurements closely matched manually segmented RS volumes, while the ABC/2 method tended to overestimate, particularly for larger hemorrhages. The median relative volume deviation was -1.7% (IQR -9.5, 6.5) for nnU-Net and 20.9% (IQR 8.1, 32.4) for ABC/2. ABC/2 discrepancies were greater in images with irregular hematoma boundaries, such as with the irregular shape (median 23.8%, IQR 15.2, 38.1 vs. 15.2%, IQR 2.0, 24.0; p<0.001) and satellite sign (median 23.4%, IQR 16.1, 37.0 vs. 15.9%, IQR 3.1, 24.2; p<0.001). The swirl sign also affected ABC/2 accuracy but was not independent of irregular shape (Chi-square 35.21, p<0.001). In contrast, the automated method was not significantly affected by shape or density markers, though intraventricular hemorrhage (IVH) occasionally led to overestimations. Conclusions Automated deep learning techniques surpass the traditional ABC/2 method for accurately measuring ICH volumes. While the ABC/2 method overestimates volumes in irregular hematomas, deep learning methods occasionally overestimate when IVH is present.
No takes yet. Share an insight, caveat, or question.
Nawabi et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: