RSNA Abdominal Trauma CT (RATIC) Dataset

• RATIC is a comprehensive, multi-level annotated dataset with 4,274 abdominal CT studies from 23 institutions across 14 countries.
• It features study-level, image-level, and voxelwise annotations for solid organs and traumatic findings, enabling detailed injury analysis.
• Benchmark studies indicate that optimized CNNs can outperform transformer models in segmentation and detection tasks on diverse clinical data.

The RSNA Abdominal Traumatic Injury CT (RATIC) dataset is a public abdominal trauma imaging resource created for the RSNA 2023 Abdominal Trauma Detection competition and intended to support machine learning research in detection, classification, grading, localization, and segmentation of traumatic findings on CT. It comprises 4,274 abdominal/pelvic CT studies and 6,481 CT image series from 23 institutions across 14 countries, with expert annotations spanning liver, spleen, kidneys, bowel, mesentery, and active extravasation. Its defining characteristic is multi-level supervision: study-level labels, image-level labels for sparse findings, and voxelwise organ segmentations on a curated subset (Rudie et al., 2024).

1. Dataset scope and cohort composition

RATIC is described as the largest publicly available collection of adult abdominal CT studies annotated for traumatic injuries (Rudie et al., 2024). The cohort includes adults aged at least 18 years, with mean age approximately 48 years and range 18–90+, and draws from 23 contributing institutions across 14 countries and 6 continents. Images are distributed in DICOM format, organ masks in NIfTI format for a subset, and tabular annotations in CSV files; the dataset is de-identified per HIPAA and global privacy standards and is available for non-commercial use via Kaggle (Rudie et al., 2024).

The dataset was assembled to reflect clinically relevant variation rather than a tightly controlled acquisition protocol. The included studies span arterial, portal venous, split bolus, and multiphasic acquisitions, with slice thickness at most 5 mm and preferably thinner, and broad anatomical coverage. Quality assurance included DICOM integrity checks and cropping to the abdominal/pelvic region using DeepATLAS followed by radiologist review. Sites were encouraged to enrich for positive injuries, targeting roughly 50% positive cases and balancing across injury types, institutions, sex, and age (Rudie et al., 2024).

The released training and evaluation structure for the competition-scale dataset is 3,147 train, 404 public test, and 723 private test studies (Rudie et al., 2024). Downstream studies sometimes restate the task at the CT-series level; one label-scarcity study summarizes the RSNA Abdominal Trauma Detection dataset as 4,711 CT series, emphasizing that only 206 have segmentation masks (Chaudhary et al., 12 Mar 2026). This suggests that RATIC supports both study-centric and series-centric experimental formulations.

2. Annotation schema and curation workflow

RATIC annotations are organized at three distinct levels. At the study level, the dataset records presence and grade of injury for liver, spleen, and kidneys, and presence or absence of bowel or mesenteric injury and active extravasation. At the image level, bowel or mesenteric injury and active extravasation are labeled per image because these findings may only appear on a few slices. At the voxel level, a subset contains organ segmentation masks for liver, spleen, left kidney, right kidney, and bowel, where the bowel label includes esophagus, stomach, duodenum, and small and large bowel (Rudie et al., 2024).

For solid-organ injuries, grading follows the American Association for the Surgery of Trauma Organ Injury Scale, with grades collapsed into low and high categories:

$$\text{Injury Grade} = \begin{cases} \text{Low}, & \text{AAST grade} \in \{I, II, III\} \ \text{High}, & \text{AAST grade} \in \{IV, V\} \end{cases}$$

This grading is included for each solid-organ injury where present and is explicitly framed as clinically important for triage (Rudie et al., 2024).

Curation relied on subspecialty review. Volunteers from the American Society of Emergency Radiology and Society of Abdominal Radiology performed the annotations. Each scan with injury was reviewed by 2–3 radiologists per organ or injury category; solid-organ cases used three independent blinded annotations with majority vote and adjudication for ties, whereas bowel and mesenteric injury and extravasation used consensus approaches with two or three annotators depending on the task. Annotation was performed on md.ai, and annotators were assigned to specific organs or injuries to maximize consistency (Rudie et al., 2024).

The segmentation subset was produced by first applying nnU-Net trained on TotalSegmentator and then manually correcting the masks. That subset was enriched for challenging and high-grade injuries (Rudie et al., 2024). A plausible implication is that the voxelwise labels serve a dual role: anatomical context for injury models and a test bed for strong-supervision methods that would be infeasible to scale to the entire corpus.

3. Segmentation subset and organ-context benchmarks

Voxelwise segmentations are available for 206 series from 128 studies (Rudie et al., 2024). A dedicated benchmark for multi-organ segmentation on RATIC uses 206 annotated abdominal CT scans from 23 institutions, covering liver, spleen, left kidney, right kidney, and bowel. In that benchmark, labels were generated by nnU-Net trained on TotalSegmentator and then manually corrected by experts; scans were split into 144 training, 41 validation, and 21 test volumes (Bayer et al., 19 Mar 2026).

The benchmark emphasizes heterogeneity at every stage. Images are 512×512 per slice with 50–300 slices in the $z$-dimension, and the cohort shows considerable variation in scanner models, acquisition protocols, and voxel spacing. Preprocessing reorients images to RAS, resamples all data to isotropic $1\text{ mm}^3$ voxels, scales intensities to $[0,1]$, and trains on random $96^3$ patches with oversampling for smaller organs using the ratio $[1:1:2:2:2:1]$ for background, liver, left kidney, right kidney, spleen, and bowel (Bayer et al., 19 Mar 2026).

All evaluated models were trained under identical conditions, and performance was measured with the Dice Similarity Coefficient,

$DSC = \frac{2 |A \cap B|}{|A| + |B|},$

computed per organ and averaged across organs (Bayer et al., 19 Mar 2026).

Model	Average DSC	Convergence
UNETR	0.918	~10,000 iterations
SwinUNETR	0.829	~10,000 iterations
UNETR++	0.934	~5,000 iterations
SegResNet	0.945	~10,000 iterations

Under this controlled setup, the CNN-based SegResNet achieved the highest average DSC, 0.945, and the best result for every organ. Among transformer-based models, UNETR++ was strongest at 0.934 and converged in approximately half the iterations of the other models, while SwinUNETR underperformed substantially, particularly on the left kidney with DSC 0.698 (Bayer et al., 19 Mar 2026). The authors interpret these results as evidence that, on small- to medium-sized heterogeneous datasets such as RATIC, a well-optimized CNN can remain more data-efficient than hybrid transformer architectures (Bayer et al., 19 Mar 2026).

This benchmark also clarifies a recurrent misconception: RATIC is not primarily a dense organ-segmentation resource. The segmentation subset is high value but small relative to the full dataset, and conclusions drawn from it must be understood in that context (Rudie et al., 2024).

4. Classification and triage formulations on RATIC

The full dataset supports study-level injury triage and organ-specific classification. One hybrid framework uses the RSNA 2023 challenge data as CT volumes represented as stacks of 96 slices per scan, with image-level and organ-level annotations. Its pipeline applies 3D segmentation to create masks and crops, then uses a 2.5D 2D CNN plus RNN architecture, followed by ensemble aggregation across slice-level and patient-level predictions (Jiang et al., 2024). Reported scores were 0.6831/0.6712 for a 2D CNN baseline, 0.5123/0.5231 for segmentation using ResNet50, 0.3451/0.3451 for 3D segmentation plus ResNet3D plus UNet, and 0.3323/0.3356 for the full 3D segmentation plus 2D CNN plus RNN pipeline on the public and private sets, respectively (Jiang et al., 2024).

Another line of work treats the task as multi-organ injury classification and focuses on representation learning from unlabeled slices. The 2D-VoCo framework adapts volume contrastive learning to 2D CT slices and transfers the pretrained backbone into a CNN-LSTM classifier for kidney, liver, and spleen injury status. Using 3,147 CT studies with an 80:20 patient split and five-fold cross-validation for thresholding, the baseline without 2D-VoCo reported RSNA score $0.4133 \pm 0.0068$, mAP $59.47 \pm 0.01$, precision $60.84 \pm 0.55$, and recall $z$0. With 2D-VoCo, the corresponding values were $z$1, $z$2, $z$3, and $z$4; adding extra unlabeled data further yielded $z$5, $z$6, $z$7, and $z$8 (Chiu et al., 21 Jan 2026).

These results situate RATIC as a test bed for both end-to-end clinical triage systems and label-efficient pretraining strategies. A notable pattern across these studies is that sequence modeling across slices remains central even when the final task is study-level classification, reflecting the fact that traumatic findings may be focal, discontinuous, or only visible across a short run of adjacent images (Jiang et al., 2024, Chiu et al., 21 Jan 2026).

5. Label-efficient detection and organ localization

RATIC has also motivated methods designed explicitly for sparse supervision. A self-supervised and semi-supervised framework uses patch-based masked image modeling to pretrain a 3D U-Net encoder on 1,206 CT volumes without annotations, then applies the encoder to two downstream tasks: 3D injury detection via VDETR with Vertex Relative Position Encoding and multi-label injury classification (Chaudhary et al., 12 Mar 2026). For detection, the study uses 144 labeled training samples, 30 validation samples, 32 test samples, and 2,000 unlabeled volumes in a semi-supervised consistency-regularization setup. The reported validation [email protected] rises from 26.36% without SSL to 56.57% with SSL, while test [email protected] rises from 23.03% to 45.30%, corresponding to a 115% validation improvement and a 97% test improvement over supervised-only training (Chaudhary et al., 12 Mar 2026).

The same study reports that, for multi-label injury classification, a frozen pretrained encoder combined with a linear probe and 2,244 labeled samples achieves 94.07% test accuracy across seven injury categories; with only 144 labeled samples and fine-tuning plus heavy augmentation, test accuracy is approximately 77.7% (Chaudhary et al., 12 Mar 2026). In this formulation, label scarcity is the central methodological constraint rather than a secondary nuisance variable.

Organ localization provides another intermediate task on the segmentation subset. CT-3GDINO adapts a Grounding-DINO-style detector to fixed abdominal organs using frozen pseudo-text class tokens, a Swin3D backbone, bidirectional feature enhancement, pseudo-text-guided query selection, and a cross-modality decoder. Trained and evaluated on 193 matched RSNA/RATIC CT volumes with segmentation-derived bounding boxes, the best multi-scale model trained from scratch achieves overall top-1 class-wise mAP 0.5830 across 3D IoU thresholds 0.1 to 0.7, with AP 0.9649 at IoU 0.1 and 0.1552 at IoU 0.7 (Chen et al., 25 Jun 2026). The strong loose-threshold performance and limited strict-threshold performance indicate that coarse organ localization is already reliable, whereas precise box alignment remains difficult (Chen et al., 25 Jun 2026).

A related precursor, though not trained on RATIC itself, is the Multi-Scale Attentional Network for segmentation of active arterial bleeding in abdominopelvic trauma CT after pelvic fractures. MSAN uses a ResNet-based encoder with ASPP, multi-scale training and inference, a non-local attention module, and multi-view fusion across axial, coronal, and sagittal planes. On 65 pelvic trauma CT cases, with 45 for training and 20 for evaluation, the best ResNet101-MSAN-3-scale variant achieved 59.89% DSC, exceeding Yu et al. at 52.12% and 3D U-Net at 40.81% (Zhou et al., 2019). Because RATIC includes active extravasation labels at the image level, this earlier hemorrhage-localization literature is directly relevant to the broader RATIC research agenda.

6. Evaluation caveats, limitations, and interpretive issues

RATIC’s breadth does not eliminate annotation and deployment caveats. The dataset paper explicitly notes inter-rater variability, exclusion of delayed phase imaging, and the fact that not all trauma types such as fractures and hematomas are exhaustively annotated (Rudie et al., 2024). These limitations matter because negative cases for one target can still contain other abnormal findings, creating compound distribution shift during evaluation.

This issue is analyzed directly in a bowel-injury study comparing two foundation models, MedCLIP and RadDINO, with three task-specific approaches on the multi-institutional RSNA dataset. Training used 3,147 patients with bowel injury prevalence 2.3%; evaluation used an enriched 100-patient test set. Foundation models achieved AUC 0.64–0.68 versus 0.58–0.64 for task-specific models and had higher sensitivity, 79–91% versus 41–74%, but lower specificity, 33–50% versus 50–88%. When specificity was stratified by negative-class composition, all models had high specificity in patients without abdominal pathology, 84–100%, but specificity declined sharply when solid-organ injuries were present: 50–51 percentage-point drops for foundation models versus 12–41 percentage-point drops for task-specific models (Raythatha et al., 10 Feb 2026).

The significance of that result is methodological as much as clinical. It shows that prevalence imbalance alone does not explain false-positive behavior; negative-class heterogeneity is itself a primary source of failure (Raythatha et al., 10 Feb 2026). For RATIC, this means that strong discrimination metrics on enriched or simplified test sets do not automatically imply robust clinical triage performance under confounding pathology.

A second interpretive issue concerns architectural generalization claims. On the segmentation subset, transformer-based designs did not uniformly outperform CNNs, and on bowel-injury detection, foundation models matched discrimination but not specificity under confounding conditions (Bayer et al., 19 Mar 2026, Raythatha et al., 10 Feb 2026). Together, these findings suggest that RATIC is less a benchmark for confirming a single model family than a benchmark for stress-testing how inductive bias, supervision level, and dataset heterogeneity interact in trauma CT.

RATIC’s broader importance follows from that role. It combines multi-institutional acquisition diversity, expert consensus labels, sparse lesion-level supervision, and a small but carefully curated voxelwise subset in a single public resource (Rudie et al., 2024). This combination makes it unusually suitable for comparative work spanning organ segmentation, organ localization, injury grading, image-level detection of subtle findings such as bowel injury and active extravasation, and label-efficient representation learning.