Hemorica: A Multi-Granularity ICH CT Dataset

• Hemorica dataset is a comprehensive non-contrast head CT benchmark designed for automated classification, detection, segmentation, and volumetric quantification across five hemorrhage subtypes.
• It offers fine-grained annotations including patient-wise labels, slice-wise labels, subtype-specific bounding boxes, 2D pixel masks, 3D voxel masks, and hemorrhage volume estimates from 372 CT studies.
• Baseline experiments demonstrate high accuracy in binary slice classification and segmentation, underscoring its value for multi-task AI in emergency neuroradiology.

Searching arXiv for papers on Hemorica and closely related dataset papers. Hemorica is a publicly available head CT dataset for intracranial hemorrhage (ICH) intended to support automated classification, detection, segmentation, and quantification across five major hemorrhage subtypes: epidural hemorrhage (EPH), subdural hemorrhage (SDH), subarachnoid hemorrhage (SAH), intraparenchymal hemorrhage (IPH), and intraventricular hemorrhage (IVH). It was introduced to address the fragmentation of public ICH resources by combining patient-wise labels, slice-wise labels, subtype-specific bounding boxes, two-dimensional pixel masks, three-dimensional voxel masks, and subtype-specific hemorrhage volume estimates within a single benchmark framework (Davoodi et al., 26 Sep 2025).

1. Dataset identity and intended scope

Hemorica was presented as a comprehensive non-contrast head CT dataset for acute hemorrhage analysis. Its stated purpose is to support a broad spectrum of AI tasks with fine-grained supervision, rather than only slice-level classification or only binary segmentation. In the dataset paper, the core claim is that Hemorica is the first public dataset with a sufficient number of CT studies and simultaneous patient-wise labels, slice-wise labels, subtype-specific detection boxes, 2D masks, and 3D voxel-level annotations for all five major hemorrhage subtypes, plus hemorrhage volume estimates (Davoodi et al., 26 Sep 2025).

The dataset is explicitly positioned against earlier public resources that are either predominantly classification-oriented or lack subtype-specific dense annotations. In that framing, CQ500 and RSNA are treated as large but weakly localized resources, whereas smaller segmentation datasets are described as lacking a unified multi-task annotation hierarchy. This positions Hemorica as a benchmark for models that must move across annotation granularities, from coarse diagnosis to lesion localization and volumetric delineation.

The medical taxonomy follows standard hemorrhage subtype definitions used in emergency neuroradiology. EPH and SDH are treated as extra-axial entities, whereas SAH, IPH, and IVH involve the brain parenchyma or ventricular system. The dataset paper also emphasizes that subtype labels are not mutually exclusive: multiple hemorrhage types may co-occur within the same patient and on the same slice, so the underlying annotation structure is multi-label even when downstream experiments collapse the task to binary hemorrhage-versus-non-hemorrhage prediction.

2. Cohort construction, imaging data, and preprocessing

The formal dataset paper reports 372 head CT examinations from Rasoul Akram Hospital, Iran, acquired retrospectively between 2012 and 2024. The release was obtained by screening an initial collection of 420 CT studies and excluding 48 because of noise, low-quality scans, post-operative radiography, and corrupted files, leaving the final 372-study cohort (Davoodi et al., 26 Sep 2025).

The imaging modality is head non-contrast CT. Raw studies were exported from PACS in DICOM format, converted to NIfTI, and anonymized. The paper does not report scanner manufacturer or model, reconstruction kernel, in-plane spacing, tube voltage, tube current, or institution count beyond the fact that the source is single-center. It also does not provide patient demographic summaries such as age or sex.

The total slice count is reported almost consistently as 12,067 slices. The number of slices per patient ranges from 16 to 56, with mean 32.2 and standard deviation 7.4. For benchmark experiments, the paper uses a patient-wise 80/20 train/test split with all slices from a patient kept in only one partition. The reported split is 298 patients / 9,638 slices for training and 74 patients / 2,429 slices for testing. At the patient level, the split contains 86 healthy and 212 hemorrhagic training cases, versus 23 healthy and 51 hemorrhagic test cases. Slice-wise, the train split contains 7,482 healthy slices and 2,156 hemorrhagic slices, whereas the test split contains 1,906 healthy slices and 523 hemorrhagic slices (Davoodi et al., 26 Sep 2025).

Preprocessing in the benchmark section is simple but not completely internally consistent. The paper states that Hounsfield Units were limited by setting window width = 40 and window length = 80, while the configuration table lists Width = 80, Length = 40. This creates an unresolved notation inconsistency. The benchmark inputs were resized to $224 \times 224$, normalized to [0,1], and the single-channel CT slice was duplicated to produce a 3-channel input compatible with ImageNet-pretrained backbones. For some descriptive visualizations, head orientation was corrected so the skull was aligned consistently for spatial heatmap aggregation.

3. Annotation hierarchy and quality-control workflow

The defining feature of Hemorica is its multi-granularity supervision. The dataset supports patient-wise classification, slice-wise classification, object detection, 2D segmentation, and 3D segmentation, all with subtype specificity. Slice-level bounding boxes are not described as manually drawn from scratch; rather, subtype masks were manually produced first, after which connected component algorithms were used to generate lesion polygons and then derive bounding boxes. The segmentation outputs include both 2D pixel-level masks and 3D voxel-level masks, and the paper further states that hemorrhage volume estimates are available for each subtype (Davoodi et al., 26 Sep 2025).

Annotation level	Content
Patient-wise	Presence or absence of each hemorrhage subtype
Slice-wise	Subtype presence on individual axial slices
Detection	Subtype-specific slice-level bounding boxes
2D segmentation	Subtype-specific pixel masks
3D segmentation	Subtype-specific voxel masks

The annotation process was carried out in 3D Slicer on full volumetric CT series. The annotating team included a neurosurgeon, a radiology specialist, and a practitioner. The neurosurgeon supervised the process; the radiology specialist and practitioner independently annotated the scans while blinded to each other’s work and without access to patient history or clinical metadata. A pilot consensus phase on 50 CT scans was used to establish guidelines and define borderline cases. After that phase, annotation proceeded in weekly batches that also revisited a subset of prior cases for ongoing intra-rater and inter-rater reliability assessment. To reduce fatigue, annotation sessions were limited to 4–5 hours. The paper states that the radiologist’s annotations were designated as the segmentation ground truth, while the practitioner’s annotations were used for annotation-quality assessment (Davoodi et al., 26 Sep 2025).

The paper characterizes this workflow as a double-reading procedure with neurosurgeon adjudication and states that it maintained low inter-rater variability. However, it does not report quantitative agreement statistics such as Cohen’s $\kappa$, Dice between readers, ICC, or adjudication rates. Likewise, although it refers to a standardized annotation protocol, the detailed operational rules are not reproduced in the manuscript.

4. Label distribution, lesion geometry, and subtype co-occurrence

At the patient level, Hemorica contains 263 hemorrhagic scans (70.7%) and 109 healthy scans (29.3%). The patient-wise subtype composition is reported as IPH 57.3%, IVH 15.1%, SAH 15.6%, SDH 8.2%, and EPH 3.8%. At the slice level, the paper reports 2,678 hemorrhagic slices (22.2%) and 9,388 healthy slices (77.8%). Because these values sum to 12,066 rather than 12,067, there is a one-slice discrepancy relative to the total dataset count (Davoodi et al., 26 Sep 2025).

The slice-wise subtype composition among hemorrhagic slices is IPH 49.7%, IVH 14.5%, SAH 23.1%, SDH 7.8%, and EPH 4.8%. Hemorrhage occurrence across axial position is also characterized: the number of positive cases by slice index ranges from 0 to 42, with a peak at the 14th slice, and the mean hemorrhage occurrence index is 15.29 with standard deviation 6.16.

Lesion-size analysis emphasizes marked heterogeneity. IPH has the largest and broadest distribution, with maximum 27,322 voxels, median 1,985, and mean 3,076 in one-slice lesion area terms. Across dataset-level comparisons, Hemorica’s mean bounding-box area is reported as 3,172 voxels, and its mean segmentation mask area as 2,885 voxels. The paper also states that IVH and SAH tend to be smaller and morphologically more challenging. Heatmap analyses are consistent with anatomical expectations: IPH and IVH are centrally concentrated, whereas SDH and SAH are more peripheral, and EPH is both peripheral and rare (Davoodi et al., 26 Sep 2025).

Subtype co-occurrence is quantified using the conditional probability

$P(m \mid n) = \frac{P(m \cap n)}{P(n)}.$

At the patient level, Hemorica shows especially strong coupling between IPH and IVH; for example, $P(\text{IPH}\mid \text{IVH}) = 0.891$. By contrast, EPH is largely isolated, with $P(\text{EPH}\mid \text{IVH}) = 0.000$ and $P(\text{EPH}\mid \text{IPH}) = 0.005$. At the slice level, overlap weakens but the IPH–IVH relationship remains notable, with $P(\text{IPH}\mid \text{IVH}) = 0.578$ and $P(\text{IVH}\mid \text{IPH}) = 0.168$. These statistics reinforce that Hemorica is a multi-label dataset rather than a single-label subtype benchmark (Davoodi et al., 26 Sep 2025).

5. Baseline experiments and demonstrated downstream use

The original dataset paper uses baseline experiments primarily as a validation of annotation quality and task feasibility. For binary slice classification, it fine-tunes ten ImageNet-1K pretrained 2D architectures, including ResNet-18/50, DenseNet-121/161, EfficientNetV2 variants, Swin Transformer V2 variants, and MobileViT XS/S. The reported training setup uses 50 epochs, batch size 16, Adam, Cross Entropy Loss, PyTorch, random seed 32, and an NVIDIA GeForce RTX 4070. Among the reported results, MobileViT-S achieves the best overall F1 = 88.9%, while the paper highlights MobileViT-XS as a lightweight model with F1 = 87.8%, AUC = 97.1%, and Specificity = 97.1%. For binary lesion segmentation, the paper benchmarks U-Net and PSPNet with multiple encoders; the strongest result is U-Net with DenseNet-161, which reaches Dice = 85.5% and IoU = 74.7% (Davoodi et al., 26 Sep 2025).

A notable limitation of the benchmark section is that, although detection is presented as a supported task at the dataset level, the paper does not report a detection baseline table. The experiments therefore validate binary slice classification and binary segmentation directly, but not lesion detection quantitatively within the original publication.

Subsequent papers use Hemorica as a downstream benchmark in ways that rely specifically on its coexistence of slice labels and pixel masks. One explainability study treats Hemorica as a binary slice-classification dataset with lesion masks for quantitative CAM evaluation and reports that AblationCAM at Stage 5 / layer [-3] of EfficientNetV2-S yields the best global pixel-level localization with Pixel Dice 0.5744 and Pixel IoU 0.4029, while HiResCAM achieves the best global bounding-box overlap with BBox Dice 0.5723 and BBox IoU 0.4009 (Rafati et al., 25 Aug 2025). A later detection study reformulates CT volumes as ordered slice sequences, uses Hemorica for YOLO-based lesion detection and ByteTrack-based slice-to-slice consistency enforcement, and reports that on independent test data a Hybrid ByteTrack strategy increases Precision from 0.703 to 0.779 and F1-score from 0.674 to 0.707 relative to the baseline 2D detector while keeping Recall near baseline (0.643 to 0.647) (Parvahan et al., 5 Jan 2026).

These follow-on studies show that Hemorica functions not only as a classification or segmentation corpus, but also as a reusable substrate for explainability benchmarking, localization from weak supervision, and object-level volumetric reasoning.

6. Relation to other hemorrhage CT datasets and unresolved issues

Within the ICH imaging landscape, Hemorica is described as occupying a middle ground between large weakly labeled corpora and smaller densely annotated datasets. The dataset paper contrasts it with CQ500, which provides patient-level labels for five hemorrhage types but originally lacks slice-wise labels, boxes, or masks; RSNA 2019, which is much larger but does not supply dense masks in the original release; PhysioNet CT-ICH, which provides binary voxel masks on a smaller cohort; and PHE-SICH-CT-IDS, which supports edema-focused segmentation and detection rather than comprehensive subtype-specific hemorrhage annotation (Davoodi et al., 26 Sep 2025).

A recurrent point of confusion is Hemorica’s reported size and provenance across papers. The formal dataset paper describes 372 examinations from a single center. By contrast, the CAM benchmark uses 327 non-contrast head CT studies/examinations and 12,067 axial slices, with 2,679 hemorrhage-positive and 9,388 hemorrhage-negative slices, and describes the resource as collected from multiple institutions (Rafati et al., 25 Aug 2025). The later CT-as-video paper likewise uses 327 patients, 12,067 axial slices, 261 training patients, and 66 test patients, but also contains an internal statement elsewhere referring to 372 patients (Parvahan et al., 5 Jan 2026). This suggests that later benchmark papers may have operated on a 327-study subset or an earlier curation, but the available texts do not fully reconcile the discrepancy.

Other limitations are more straightforward. Hemorica is explicitly a single-center retrospective dataset in its formal description, which raises generalizability concerns relative to multi-center resources. Although 372 studies are substantial for a finely annotated medical dataset, the cohort remains modest compared with very large classification datasets such as RSNA. The paper also omits important metadata that many clinical AI studies would regard as essential: scanner inventory, acquisition protocol details, patient demographics, mechanism of injury or cause, outcomes, symptom-onset timing, and comorbidities. Release logistics are similarly incomplete in the paper text: Hemorica is repeatedly described as publicly available, but no URL, DOI, license, or access protocol is given there (Davoodi et al., 26 Sep 2025).

These omissions do not negate the dataset’s importance, but they define its current documentary boundaries. Hemorica is best understood as a fine-grained, task-complete ICH CT benchmark whose main contribution is the coexistence of subtype-specific supervision at patient, slice, lesion, 2D mask, and 3D voxel levels. Its most consequential unresolved issues are not annotation richness, but cross-paper versioning ambiguity, incomplete release metadata, and limited reporting of acquisition and demographic context.