Dataset overview
BrainTumor48K combines de-identified institutional cohorts with diverse public MRI resources. The private data support model development, multi-center retrospective testing, and prospective clinical evaluation, while the public data broaden tumor coverage, add healthy controls, and support both 2D representation learning and 3D case-level training. The tables below report the institutional cohort characteristics and, for every public source, the available supervision and its role in the study.
Private data · Institutional cohorts
The private data come from Xiangya Hospital and 11 independent centers: Changde First People’s Hospital, Shanghai Tongji Hospital, the University of South China Second Hospital, Shenzhen Second People’s Hospital, the Third Xiangya Hospital, Jiangxi Provincial People’s Hospital, Chongqing Traditional Chinese Medicine Hospital, the First and Second Affiliated Hospitals of Nanchang University, Hunan Provincial Children’s Hospital, and the First Affiliated Hospital of Lanzhou University. Available T1, T1c, T2, and FLAIR MRI, demographics, radiology reports, and pathological diagnoses were de-identified. Institutional reports were terminology-standardized and reviewed by neuroradiologists before use.
| Dataset or medical center | No. of subjects | Male | Female | Age in years, mean ± SD (range) |
|---|---|---|---|---|
| Total | 10,147 | 4,921 (48%) | 5,226 (52%) | 53 ± 16 (1–86) |
| Train dataset | 4,936 | 2,571 (52%) | 2,365 (48%) | 43 ± 19 (1–86) |
| Test dataset | 5,211 | 2,350 (45%) | 2,861 (55%) | 48 ± 17 (1–86) |
| Xiangya Hospital | 8,813 | 4,256 (48%) | 4,557 (52%) | 45 ± 18 (1–86) |
| Changde First People’s Hospital | 361 | 193 (53%) | 168 (47%) | 55 ± 13 (6–86) |
| Shanghai Tongji Hospital | 104 | 54 (52%) | 50 (48%) | 54 ± 16 (5–79) |
| University of South China Second Hospital | 110 | 57 (52%) | 53 (48%) | 52 ± 15 (6–80) |
| Shenzhen Second People’s Hospital | 52 | 30 (58%) | 22 (42%) | 49 ± 15 (8–75) |
| The Third Xiangya Hospital | 103 | 49 (48%) | 54 (52%) | 52 ± 17 (7–76) |
| Jiangxi Provincial People’s Hospital | 232 | 107 (46%) | 125 (54%) | 53 ± 15 (13–82) |
| Chongqing Traditional Chinese Medicine Hospital | 42 | 17 (40%) | 25 (60%) | 58 ± 12 (26–78) |
| The First Affiliated Hospital of Nanchang University | 97 | 40 (41%) | 57 (59%) | 46 ± 17 (20–67) |
| Hunan Provincial Children’s Hospital | 103 | 56 (54%) | 47 (46%) | 28 ± 17 (1–60) |
| The First Affiliated Hospital of Lanzhou University | 81 | 31 (38%) | 50 (62%) | 52 ± 13 (8–76) |
| The Second Affiliated Hospital of Nanchang University | 49 | 31 (63%) | 18 (37%) | 54 ± 12 (13–74) |
The 4,936-case development cohort comprises 4,816 training patients and 120 validation patients from Xiangya Hospital. The 5,211-case retrospective evaluation comprises 3,877 primary-test patients and 1,334 patients from the 11 external centers.
Public data · Online brain MRI resources
The 39 online sources include BraTS23 and TCIA collections, ReMIND, UPENN-GBM, Radiopaedia, OpenNeuro, IXI, PubMed Central, ImageCLEF, and multiple Kaggle datasets. Structured repositories were converted into case-level MRI, diagnosis, metadata, and report fields. Web-derived figures were separated, filtered, and aligned with their captions. The complete source list below records data format, scale, original information, and the tasks organized for this study.
| Data type | Data resource | No. of subjects | No. of images | Original information | Organized tasks |
|---|---|---|---|---|---|
| 2D web crawling | PubMed Central | 23,849 | 116,296 | Description | 2D MRI slice–text description; tumor classification |
| 2D web crawling | Ctisus | 150 | 1,124 | Description; diagnosis | 2D MRI slice–text description; tumor classification |
| 2D web crawling | ImageCLEFmedical23 | 484 | 4,631 | Description | 2D MRI slice–text description; tumor classification |
| 2D Kaggle | Br35H | — | 2,137 | Diagnosis; segmentation | 2D MRI slice–text description; tumor classification |
| 2D Kaggle | Figshare Brain Tumor Dataset | 233 | 3,064 | Diagnosis | Tumor classification |
| 2D Kaggle | Brain Tumor Classification | — | 3,264 | Diagnosis | Tumor classification |
| 2D Kaggle | Brain Tumor MRI Images 44 Classes | — | 4,479 | Diagnosis | Tumor classification |
| 3D Tumor | LGG-1p19qDeletion | 159 | 32,829 | Diagnosis; segmentation | 3D MRI–report; tumor classification |
| 3D Tumor | BraTS23 | 3,263 | 2,156,589 | Diagnosis; segmentation | 3D MRI–report; tumor classification |
| 3D Tumor | ReMIND* | 108 | 60,963 | Diagnosis; metadata | 3D MRI–report; tumor classification |
| 3D Tumor | RHUH-GBM | 40 | 122,835 | Diagnosis | Tumor classification |
| 3D Tumor | UPENN-GBM | 630 | 622,112 | Diagnosis; segmentation | 3D MRI–report; tumor classification |
| 3D Tumor | GLIS-RT | 230 | 83,196 | Diagnosis | Tumor classification |
| 3D Tumor | QIN GBM Treatment Response | 54 | 66,988 | Diagnosis | Tumor classification |
| 3D Tumor | LUMIERE | 91 | 460,155 | Diagnosis | Tumor classification |
| 3D Tumor | OpenNeuro (Diffuse Gliomas) | 42 | 22,412 | Diagnosis | Tumor classification |
| 3D Tumor | Brain-Tumor-Progression | 20 | 46,573 | Diagnosis; segmentation | 3D MRI–report; tumor classification |
| 3D Tumor | Radiopaedia | 2,606 | 376,374 | Diagnosis; report | 3D MRI–report; tumor classification |
| 3D Tumor | Burdenko | 180 | 460,524 | Diagnosis | Tumor classification |
| 3D Tumor | RIDER | 19 | 115,100 | Diagnosis | Tumor classification |
| 3D Tumor | ACRIN-DSC-MR-Brain | 123 | 1,385,512 | Diagnosis | Tumor classification |
| 3D Tumor | ACRIN-FMISO-Brain | 50 | 508,500 | Diagnosis | Tumor classification |
| 3D Tumor | Meningioma-SEG-CLASS | 96 | 83,556 | Diagnosis; metadata | 3D MRI–report; tumor classification |
| 3D Tumor | Brain Metastasis MRI Dataset | 75 | 85,400 | Diagnosis; segmentation | 3D MRI–report; tumor classification |
| 3D Tumor | Brain-TR-GammaKnife | 47 | 8,300 | Diagnosis | Tumor classification |
| 3D Healthy | AOMIC-ID1000 | 928 | 1,101,652 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | AOMIC-PIOP1 | 216 | 251,380 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | IXI | 582 | 398,774 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | OpenNeuro (Listening Task) | 78 | 136,694 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | OpenNeuro (Healthy Adults) | 66 | 122,175 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | OpenNeuro (Visual Audiovisual Speech) | 60 | 105,160 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | OpenNeuro (Dynamic Passive Threat) | 82 | 126,299 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | OpenNeuro (UCLA Consortium) | 272 | 285,910 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | OpenNeuro (NIMH Healthy) | 157 | 461,281 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | DLBS Dataset | 315 | 99,344 | Diagnosis | Tumor or healthy classification |
| 3D Healthy | QTAB Dataset | 422 | 1,251,467 | Diagnosis | Tumor or healthy classification |
| Public test | EGD Dataset | 774 | — | Diagnosis | Tumor classification |
| Public test | Brain-Mets-Lung | 100 | — | Diagnosis | Tumor classification |
| Public test | Vestibular-Schwannoma-MC2 | 190 | — | Diagnosis | Tumor classification |
“3D Healthy” denotes healthy participants with 3D MRI volumes; “3D Tumor” denotes patients with brain tumors. ReMIND contributes 108 tumor cases after excluding six non-tumor brain disease cases. Some training data cannot be redistributed because of source-specific policies, including Radiopaedia restrictions.
BrainTumor48K dataset curation
The processed sources are assembled into complementary 2D and 3D resources rather than being reduced to a single homogeneous task. Across 47,947 individuals, BrainTumor48K contains 24,716 2D MRI cases for broad image–text representation learning and 23,231 case-level 3D MRI studies for volumetric diagnosis. Within the 3D component, 18,113 cases include both reports and pathological labels, while 5,118 provide MRI and diagnostic labels for classification.
Progressive training and diagnostic reliability
BrainVLM was developed with 35,227 online cases and 4,816 cases from Xiangya Hospital, with another 500 online and 120 institutional cases reserved for validation. Training progresses from large-scale 2D MRI–text representation learning to joint 2D–3D learning and, finally, instruction tuning on higher-quality case-level 3D MRI data linked with diagnoses and reports. After diagnostic training, consensus-driven response clustering is used to construct reliability supervision, enabling BrainVLM to associate each diagnosis with a confidence score and to provide a second candidate for review when the leading prediction is uncertain.
Data preprocessing
Structured repositories, web-derived MRI–text pairs, and institutional cohorts differ in format, annotation, and quality-control requirements. Separate processing routes standardize these heterogeneous sources into a common patient-level representation linking MRI with available diagnoses, metadata, and reports.
