Million-Scale General Medical Imaging Dataset

One-million-degree medical image data sets covering CT, MRI, X-Ray, ultrasound four-module, support multi-task AI model training in disease stove detection, partitioning, classification and reporting

CT Imaging MRI Image X-Ray Image Ultrasound Imaging Million-degree video slices

1,000,000+

Total number of image slices

Imaging Modality

Covered Body Parts

20+

Disease-marked category

Image Modularity and Coverage

CT Imaging

Chest CT. Abdominal CT Head CT Bones CT.

MRI Image

Brain MRI Spinal MRI Joint MRI Abdominal MRI

X-Ray Image

Breast X-ray Four limbs x-ray Spinal X-ray Dental Panorama

Ultrasound Imaging

Abdominal ultrasound. thyroid ultrasound Breast ultrasound Heart ultrasound

Imaging Data Specifications

Data Format	DICOM / NIFTI / PNG, complete metadata (scan parameters, layers, re-establishment, plant, etc.)
Matrix Resolution	CT/MRI ≥ 512×512；X-Ray ≥ 2048×2048；Ultrasound ≥ 800×600
Annotation Type	Bounding Box, Segmentation Mask, Catalogue Tags, Key Points
Annotation Format	COCO JSON / YOLO TXT / DICOM RT Struct / NIfTI Label Map
Device Distribution	Multi-manufacturer equipment such as GE / Siemens / Philips / Canon / Mindray to ensure modelability
Quality Requirements	No serious motion and metal forgery, double physician note + Director clearance

Punctuation and quality control process

Annotation Workflow	AI prescripts 2 physicians independently amends
ClassificationAccuracy	≥ 98%
Dice-Segment coefficient	≥ 0.88
Test mAP@0.5	≥ 0.91
	< 2.1%

AI Training Application Scenarios

Universal screening of the stoves

The generic model of cross-modular, cross-sectional diagnostics can serve as the base of the medical image AI base model.

The NT/MRI two-modular model is supported by a generic partition model that covers the entire body.

Automatic Imaging Report Generation

The automatic generation model of radiometric reports based on graphic matching of data is applied to visual encoder +LM decoder structures.

Regionalization and Luxability

Training in non-sensitive area-wide model for collecting equipment using multi-centre, multi-equipment data to improve the reliability of clinical deployments.

Data Security and De-identification

All data has been thoroughly de-identified, with all information that could directly or indirectly identify an individual removed; only parameters and labels relevant to medical research are retained。

De-identified Fields	All personal identifiers such as name, ID card number, phone number, address and hospitalization number
Reserved Fields	DICOM scan parameters, image pixel data, disease stove signs, pathological diagnostic information