Answers are grounded in this paper and cite their sources.
Background/Objectives: Extra-axial hemorrhage classification on non-contrast head CT is strongly associated with the geometry of the hemorrhagic collection relative to the inner skull surface. Conventional convolutional neural networks (CNNs) process images in Cartesian coordinates and therefore learn these anatomical relationships implicitly. To develop and evaluate the Skull-Relative Rectification and Circular Focality (SUCP) framework, an anatomy-aligned framework for classifying epidural hemorrhage (EDH), subdural hemorrhage (SDH), and normal CT slices, with an emphasis on anatomical representation and rotational stability. Methods: A dataset of 465 axial CT slices from 93 patients was used, comprising 31 patients and five slices per patient in each class. The inner cranial boundary was extracted, and a peridural band was sampled inward along local surface normals to generate a depth-by-angle representation that distinguishes tangential spread from inward thickness. Circular convolutions modeled continuity along the skull, while concentration and thickness descriptors characterized hemorrhage focality. Performance was evaluated using three repetitions of patient-level nested five-fold cross-validation. SUCP was compared with ResNet-18, EfficientNet-B0, and polar and Cartesian control models. Positional stability was assessed following a 90° in-plane image rotation. Supplementary analyses on the first cross-validation repetition examined the band dimensions, rotation-augmented and quarter-turn-invariant baselines, documented Hounsfield-unit windows on an external cohort, and computational cost. Results: SUCP achieved a patient-level accuracy of 0.832, macro-F1 of 0.831, and AUC of 0.941. Its discrimination was comparable to ResNet-18 (macro-F1: 0.830) but lower than EfficientNet-B0 (0.869), with no statistically significant differences among the models in this sample. SUCP demonstrated superior positional stability: the full model preserved all patient-level labels after 90° in-plane rotation, compared with 50.9% for ResNet-18 and 52.7% for EfficientNet-B0. Discrimination did not depend on the number of angular samples (macro-F1 0.815–0.870 for S = 120–360) and was lowest with a shallow band of 16 depth samples (0.716). An EfficientNet-B0 trained with full-circle rotation augmentation discriminated better than SUCP (macro-F1 0.968 vs. 0.815, both refitted in the same software environment) and retained 95.7% of labels at 90°, whereas SUCP used 65 times fewer parameters and 24 times fewer operations than ResNet-18. Skull-relative control models also consistently outperformed polar and Cartesian controls. Conclusions: Anatomy-aligned skull-relative rectification provides a compact and rotation-stable representation of extra-axial hemorrhage morphology. Although SUCP did not demonstrate a significant discrimination advantage over conventional CNNs in this cohort, its improved positional stability supports the value of incorporating anatomical geometry into CT-based hemorrhage classification. Its advantages are compactness, invariance to cyclic shifts of the rectified band that holds by construction, and an interpretable angular read-out, with rather higher accuracy than a well-augmented CNN. Larger and more diverse datasets are required to determine whether this approach provides a clinically meaningful improvement in discrimination.
No takes yet. Share an insight, caveat, or question.
Yıldırım et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: