Please send the signed document to tdscabus@gmail.com to receive the data access link .¶

The folder we share contains two directory and a csv file:

• DATA directory contains all IMAGE files
• MASK directory contains all ANNOTATION files
• labels.csv contains a table in which each line represents a case, identified by its unique case ID, and specifies whether the tumor is benign (abbreviated as 'B') or malignant (abbreviated as 'M'), as well as the paths to the corresponding DATA and MASK files.

Our dataset contains 200 3D volumes with refined tumor label,  these Automated 3D Breast Ultrasound (ABUS) data are obtained from a ABUS system(Invenia ABUS, GE Healthcare) in Harbin Medical University Cancer Hospital, Harbin, China.  An experienced radiologist lables and checks those data.

The image sizes are between 843*546*270 and 865*682*354. The pixel spacing is 0.200 mm and 0.073 mm, while the spacing between slices is approximately 0.475674 mm. The images will be stored in .nrrd files. Voxel-level segmentation annotations are: 0 - Background, 1 - Tumor

The proportion of training, validation and test cases  is shown as follows:

• Training cases: 100 (The relatively large number of data were used for training a robust model).
• Opened validation cases: 30 (The relatively small number of data were used for validation of algorithm from
  different participants to verify the evaluation code by validation dataset and ensure the fairness of the challenge.
  At the same time, the relatively small number of data can avoid the disclosure of test set data distribution).
• Closed test cases: 70 (The relatively large number of data  were used for a fair final leaderboard).

All three dataset are stratified sampled from whole 200 cases.