CT-RATE: Chest CT with Radiology Reports

• CT-RATE is a public chest CT dataset pairing 3D non-contrast scans with detailed radiology reports to facilitate comprehensive 3D multimodal learning.
• The dataset comprises over 25,000 CT studies and 50,000 volumes with 18 report-derived abnormality labels, scaled via both automated and manual curation.
• It underpins developments in contrastive pretraining, zero-shot abnormality detection, report generation, and label-quality benchmarking in radiological AI.

CT-RATE is a large-scale public chest CT dataset that pairs volumetric non-contrast chest CT images with radiology reports and multi-abnormality labels. It was introduced as the first dataset that pairs 3D medical images with corresponding textual reports, and later work described it, to the authors’ knowledge, as the largest publicly available chest CT dataset with paired reports (Hamamci et al., 2024, Yamagishi et al., 21 Jun 2026). CT-RATE has become a reference substrate for 3D contrastive pretraining, zero-shot abnormality detection, case retrieval, report generation, question answering, region-grounded multimodal supervision, and post hoc label-quality analysis.

1. Dataset formation and scope

In its original release, CT-RATE comprises 25,692 non-contrast chest CT studies reconstructed into 50,188 volumes from 21,304 unique patients, spanning May 2015 to January 2023 (Hamamci et al., 2024). The dataset retains multiple clinical reconstructions for the same study, including sharper kernels favored for lung parenchyma and smoother kernels for mediastinal assessment, thereby preserving real-world variation in slice thickness, spatial resolution, and noise characteristics. The number of slices per volume ranges from 100 to 600, with mean 304.7 and mode 255. Scanners span three manufacturers: Philips 61.5%, Siemens 30.1%, and PNMS (Philips-Neusoft) 8.4%; in-plane resolutions are 512×512 for 65.4% of scans, 768×768 for 4.2%, and 1024×1024 for 30.4%.

The cohort includes adults aged 18–102 years, with mean age 48.8 and mode 40; 41.6% are female and 58.4% male. Inclusion was restricted to non-contrast chest CT volumes, while contrast-injected studies and scans of other body parts were excluded. All DICOM metadata were scrubbed of personal identifiers, orientations were standardized, and reports were anonymized with regular expressions and translated from Turkish to English using the Google Translate API, followed by manual correction and re-review by bilingual final-year medical students. Only the English versions are released (Hamamci et al., 2024).

Radiology reports are organized into clinical information, technique, findings, and impression. For multimodal pretraining, the findings and impression sections were used because that combination yielded the best empirical performance. In the original preprocessing pipeline, volumes were resampled to uniform voxel spacing of 0.75 mm in-plane and 1.5 mm through-plane, center-cropped or padded to 480×480×240, converted to Hounsfield Units, clipped to $[-1000, 200]$, and normalized to $[-1,1]$ (Hamamci et al., 2024). The patient-level split assigns 20,000 patients to training and 1,304 to internal validation; in one derivative description, this corresponds to 24,128 training volumes and 1,564 validation volumes (Zhang et al., 2024).

2. Report-derived supervision and abnormality taxonomy

CT-RATE provides 18 abnormality labels derived from the accompanying reports. In the original construction, 1,000 reports were manually annotated across 18 abnormalities, a RadBERT-RoBERTa-4m model was fine-tuned on 800 training and 200 validation reports, and the trained model was then applied to the remaining 24,692 reports to generate multi-label supervision (Hamamci et al., 2024). Label harmonization in that pipeline included merging left and right mucoid impactions, combining density increase and ground-glass with lung opacity, and aggregating lung and fissural nodules.

The 18 predefined abnormality categories used across later CT-RATE studies are:

• Atelectasis
• Arterial wall calcification
• Bronchiectasis
• Cardiomegaly
• Consolidation
• Coronary artery wall calcification
• Emphysema
• Hiatal hernia
• Interlobular septal thickening
• Lung nodule
• Lung opacity
• Lymphadenopathy
• Medical material
• Mosaic attenuation pattern
• Peribronchial thickening
• Pericardial effusion
• Pleural effusion
• Pulmonary fibrotic sequela (Yamagishi et al., 21 Jun 2026)

Subsequent work repeatedly reused this taxonomy as the clinical target space for CT-RATE-based evaluation. For example, report-generation studies grouped the labels into lung, heart, aorta, esophagus, and other organs for metric reporting, while classification studies used them as study-level triage targets (Liang et al., 16 Mar 2026, Dahal et al., 19 Jan 2026). This suggests that CT-RATE’s report-derived supervision functions both as a weakly supervised learning signal and as a common evaluation ontology for chest CT foundation models.

3. Foundational role in 3D multimodal learning

CT-RATE was introduced to enable contrastive language-image pretraining directly in 3D. In the original study, CT-CLIP combined a 3D vision encoder derived from CT-ViT with a CXR-BERT text encoder, aligned paired volume and report embeddings with a CLIP-style InfoNCE objective, and used findings plus impression as text input (Hamamci et al., 2024). The abnormality detection protocol used positive and negative prompts for each pathology and computed a two-way softmax over cosine similarities, thereby preserving open-vocabulary inference.

On internal CT-RATE validation and on the external RAD-ChestCT cohort, zero-shot CT-CLIP surpassed the fully supervised CT-Net baseline across all key metrics: mean AUROC improved by 0.099 internally and 0.082 externally, mean F1 by 0.061 and 0.052, mean accuracy by 0.070 and 0.052, and mean precision by 0.065 and 0.047. Fine-tuning variants further improved performance, but CT-LiPro narrowed the model to predefined labels and harmed retrieval utility relative to the zero-shot model (Hamamci et al., 2024). In case retrieval, CT-CLIP and CT-VocabFine achieved roughly six-fold higher MAP@1 than random for volume-to-volume retrieval on internal validation, while report-to-volume retrieval exceeded fifteen-fold random Recall@K. The same study reported monotonic gains when training on larger fractions of CT-RATE, from 9.8% through 100%, reinforcing the dataset-scaling argument for 3D medical foundation models (Hamamci et al., 2024).

Later report-generation work used CT-RATE to diagnose a representational bottleneck in 3D CT embeddings. One study reported that CT-CLIP’s 512-dimensional average-pooled image embeddings had $\mathrm{dim90}=2$ and participation ratio 1.4, while organ encoders had $\mathrm{dim90}=4$–9 with participation ratio approximately 2.6–6.3, and argued that this dimensional concentration limited both retrieval and clinically faithful report generation on CT-RATE (Liang et al., 16 Mar 2026). A plausible implication is that CT-RATE has not only functioned as a training corpus, but also as an empirical stress test for the adequacy of 3D representation learning in radiology.

4. Label reliability and LLM-assisted cleaning

Because CT-RATE labels are report-derived, later work evaluated whether LLMs could identify label-report discordance at scale. After report-level deduplication to collapse multiple reconstructed volumes to one report per examination, 24,446 unique radiology reports were identified; 12 reports were excluded from the primary GPT-5.4 analysis because of Microsoft Azure AI Foundry content-safety filtering, leaving 24,434 reports and 439,812 label instances across the 18 predefined categories (Yamagishi et al., 21 Jun 2026). GPT-5.4 generated binary labels from the findings and impression sections using structured JSON output, with present assigned when a finding was clearly described as present, absent when explicitly negated or unmentioned, and uncertain mentions treated as present.

Across the full dataset, overall agreement between GPT-5.4-derived labels and the original CT-RATE labels was 96.4%, with Cohen’s $\kappa = 0.884$ (Yamagishi et al., 21 Jun 2026). Agreement varied substantially by category. Hiatal hernia showed the highest agreement and kappa, with 99.6% agreement and $\kappa = 0.986$, whereas lymphadenopathy showed the lowest agreement and kappa, with 79.4% agreement and $\kappa = 0.309$. In general discordance review excluding lymphadenopathy, radiologist adjudication supported GPT-5.4-derived labels in 72 of 97 adjudicable instances (74.2%). In targeted lymphadenopathy review, adjudication supported GPT-5.4-derived labels in 91 of 99 instances (91.9%). Against a radiologist-annotated 100-report reference set comprising 1,800 label instances, three-model majority voting across GPT-5.4, GPT-5.4 mini, and DeepSeek V3.2 achieved the highest label-macro-averaged performance, with F1 96.0% and Cohen’s $\kappa = 0.951$; CT-RATE’s original labels scored F1 93.8% and $\kappa = 0.923$ (Yamagishi et al., 21 Jun 2026).

These findings localized lymphadenopathy as the principal outlier. The study attributes much of that discordance to ambiguous mapping between report language and structured labels, including references to small or “millimetric” nodes, abbreviations such as “LAP,” and descriptions that require interpretation of size thresholds and clinical context. This suggests that, within CT-RATE, not all report-derived labels are equally well defined operationally, even when the source reports are clinically valid.

5. Derivative datasets, benchmarks, and downstream systems

CT-RATE has become the base layer of a broader ecosystem of chest CT resources and task-specific benchmarks. Later studies used it to add grounded segmentation supervision, evaluate organ-aware classification, construct fine-grained report-evaluation benchmarks, and train report-generation or question-answering systems (Zhang et al., 2024, Dahal et al., 19 Jan 2026, Yuan et al., 27 Apr 2026, Mao et al., 22 May 2025, Liang et al., 16 Mar 2026, Kalisch et al., 7 Aug 2025).

Resource or system	CT-RATE role	Selected reported figures
RadGenome-Chest CT (Zhang et al., 2024)	Extension of CT-RATE with region-grounded supervision	197 segmentation categories, 665K grounded report sentences, 1.3M grounded VQA pairs
ORACLE-CT (Dahal et al., 19 Jan 2026)	Study-level chest CT triage corpus	43,738 train, 3,409 validation, 3,040 test studies; AUROC 85.74 for best supervised GAP, 0.86 with masked attention
CT-Agent (Mao et al., 22 May 2025)	Report-generation substrate and CTQA backbone	47,149 train and 3,039 test studies in that setup; CE F1 0.420
AdaRAG-CT (Liang et al., 16 Mar 2026)	End-to-end benchmark for clinically faithful report generation	Clinical F1 0.480 on CT-RATE validation, versus 0.420 for CT-Agent
CT-FineBench (Yuan et al., 27 Apr 2026)	Source of chest CT fine-grained QA evaluation	1,564 reports, 24,148 QA pairs, 94 unique attributes, 15.4 QA pairs per report
CT-GRAPH (Kalisch et al., 7 Aug 2025)	Report-generation benchmark with anatomy-guided graph modeling	22,778 train and 1,505 test samples in that study; CE F1 0.296, a +7.9% absolute gain over the strongest baseline

RadGenome-Chest CT extends CT-RATE by adding 197 organ-level segmentation masks, 665,000 grounded report sentences, and 1.3 million grounded VQA pairs, all on the same non-contrast 3D chest CT volumes and reports (Zhang et al., 2024). Its grounding pipeline used SAT for universal 3D segmentation, GPT-4 to seed anatomical-region annotations for 2,500 validation reports, a GPT-2 sentence-to-region model that reached 94.56% accuracy on validation, and manual verification for all grounded reports and VQA pairs in the validation split. In effect, this converts CT-RATE from a global image-report corpus into a region-grounded multimodal dataset.

CT-RATE has also been used as a report-generation benchmark that emphasizes pathology coverage over lexical overlap. CT-Agent treated it as the primary large-scale corpus for full-volume chest CT report generation, reporting BLEU-4 0.231, ROUGE-L 0.490, METEOR 0.425, and CE F1 0.420 (Mao et al., 22 May 2025). AdaRAG-CT later argued that static 3D CT embeddings impose a representational bottleneck, and on CT-RATE validation improved Clinical F1 from 0.420 for CT-Agent to 0.480 (Liang et al., 16 Mar 2026). CT-GRAPH instead imposed an anatomy-guided hierarchical graph over organ features and reported CE F1 0.296 with a +7.9% absolute improvement over the strongest baseline in its evaluation protocol (Kalisch et al., 7 Aug 2025). These numbers are not directly interchangeable because the studies use different splits, feature backbones, and report-generation setups, but they demonstrate the dataset’s centrality for benchmarking clinically oriented text generation from 3D CT.

Finally, CT-FineBench used CT-RATE and Merlin to construct a fine-grained diagnostic fidelity benchmark. For the CT-RATE portion, the benchmark used the 1,564-report test split to derive 24,148 question-answer pairs spanning 94 unique attributes, with an average of 15.4 QA pairs per report and 5.2 attributes per finding (Yuan et al., 27 Apr 2026). CT-FineBench reported stronger correlation with expert judgment than BLEU, ROUGE-L, BERTScore, RadGraph, RaTEScore, and GREEN on CT-RATE report-generation outputs, thereby shifting evaluation from coarse overlap to attribute-level factual consistency.

6. Limitations, usage considerations, and availability

Several limitations recur across CT-RATE studies. First, both the original labels and the cleaned LLM-derived labels are report-derived rather than image-level ground truth, so errors or ambiguities in the report propagate into downstream supervision (Yamagishi et al., 21 Jun 2026). Second, the dataset is restricted to non-contrast chest CT from a single institution, and later derivative resources explicitly note that contrast-dependent findings are out of scope (Hamamci et al., 2024, Zhang et al., 2024). Third, the translation of reports from Turkish to English introduces another processing layer, although the release pipeline included manual correction and re-review for coherence and residual protected health information (Hamamci et al., 2024). Fourth, region-grounded extensions inherit segmentation noise: RadGenome-Chest CT applied SAT across the dataset without case-by-case manual correction in training data, and acknowledged that segmentation quality may vary by anatomy, image quality, and reconstruction (Zhang et al., 2024).

The label-cleaning study adds more specific cautions. Its discordance review sets were enriched for disagreement and therefore should not be interpreted as overall accuracy estimates; results may depend on model version, prompt design, and label definitions; LLM-generated labels can introduce bias if used as reference standards without human verification; and uncertainty mentions were treated as present in the primary extraction (Yamagishi et al., 21 Jun 2026). The same study therefore recommends preferring multi-LLM majority-vote labels when available, exercising particular caution with lymphadenopathy, and validating high-stakes applications with a radiologist-annotated subset or image-level review.

CT-RATE is publicly accessible at https://huggingface.co/datasets/ibrahimhamamci/CT-RATE (Hamamci et al., 2024). Later work states that the dataset is released under a CC BY-NC-SA license to permit noncommercial research and derived label release, and that refined labels generated by multi-LLM majority voting, together with code for cohort construction, LLM-based extraction, and downstream analyses, will be made publicly available under the same noncommercial framework (Yamagishi et al., 21 Jun 2026). In practice, CT-RATE therefore occupies a dual position: it is both a foundational multimodal corpus for 3D chest CT and an evolving benchmark whose supervision quality, anatomical grounding, and evaluation methodology continue to be refined.