Dataset Overview
| Dataset Name | General Medical Imaging Annotation Dataset |
| Total Data Volume | 108,500 cases (including CT/X-Ray/ ultrasound/MRI polymodular) |
| Model distribution | CT 42 000 / X-Ray 35 000 / Ultrasound 18,000 / MRI 13 500 |
| Annotation Type | Disease detection frames, semantic partitions, classification labels, key points |
| Covered Body Parts | Breast, abdominal, bone, breast, thyroid, skull. |
| Data Source | Multi-cooperative medical institutions + enhanced public data sets (authorized compliance) |
| Format Output | DICOM / NIfTI / PNG / COCO JSON / YOLO TXT |
| Quality Control Standards | Double physician label + Director clearance, classification accuracy & g; 98%, split Dice & g; 0.88 |
Detailed Description of Data Dimensions
| Field Name | Data Type | Value Range / Description |
|---|---|---|
| case_id | String(32) | ID for de-identification cases |
| modality | Enum | CT | XRAY | US | MRI |
| body_part | Enum | CHEST | ABDOMEN | BONE | BREAST | THYROID | BRAIN |
| finding_label | Enum(Multi) | NORMAL & NODULE |
| bbox | Array[4] | [x_center, y_center, width, high] |
| segmentation | Array | COCO format polygon or RLE encoding |
| classification | Object | {primary, secondary, confidence} |
| image_size | Array[2] | [height_px, width_px] |
| slice_index | Integer | Slices numbering in CT/MRI sequence |
| acquisition_date | Date | Date of inspection (years only) |
| device_manufacturer | String | GE / Siemens / Philips / Canon / Mindray |
De-identified JSON Data Sample
Annotation Specifications and Quality Control
Annotation Workflow
Use the "pre-mark + manual correction + cross-check" model. First, the AI model (YOLO/MedSAM) generates the initial label, then it is independently revised by two physicians, and finally the director decides the difference. The process is 60% more efficient than a purely manual label, while maintaining a high accuracy rate.
Quality control indicators
98.7%
ClassificationAccuracy
0.89
Split Dice
0.91
Test mAP@0.5
<2.1%
AI Training Use Cases
Universal screening of the stoves
Applicable model: YOLOv8/v9/DTR/RT-DRR/DINO. Cross-modular, cross-sectional generic model of the test of the stove, which can serve as the base model base for the medical image AI.
Applying model: TotalSegmentator / nnU-Net / SAM-Med2D. Trained generic partition models covering multiple organs in their entire body, supporting CT/MRI dual-modular patterns.
Auto-report chest X-ray.
Applying model: R2Gen / X-REM / CheXagent (View Encoder + LLM decoder). Auto-generated model based on graphic matching of data.
Regionalization and Luxability
Applying model: DomainBed / DANN / SWAD. Training in non-sensitive domain-wide model for collecting equipment using multi-centre, multi-equipment data.
Get This Dataset
The generic medical image data set is flexiblely authorized by the mosaic/part to support the customization of the subset.
137-5502-0164 Online ConsultationRelated Datasets
Commercial Application Value
A.I.M. image SaaS
A cloud-based AI-assisted viewing service is provided to regional video centres and medical examination institutions, covering high frequency (HF) scenes such as pulmonary knots, fractures, thyroid glands.
Imagination of medical devices
Real-time AI aids embedded in the CT/DR/ ultrasound device end, NMPA AI functional registration obtained by the Auxiliary Medical Plant.
Basic model pre-training
The first is the large-scale multi-modular pre-training language for medical image base models (e.g. RadImageNet, BioMedCLIP).