junma/CVPR-BiomedSegFM

CVPR 2025 Competition: Foundation Models for 3D Biomedical Image Segmentation

Highly recommend watching the webinar recording to learn about the task settings and baseline methods.

The dataset covers five commonly used 3D biomedical image modalities: CT, MR, PET, Ultrasound, and Microscopy. All the images are from public datasets with a License for redistribution. To reduce dataset size, all labeled slices are extracted and preprocessed into npz files. Each npz file contains

• imgs: image data; shape: (D,H,W); Intensity range: [0, 255]
• gts: ground truth; shape: (D,H,W);
• spacing

Folder structure

• 3D_train_npz_all: complete training set
• 3D_train_npz_random_10percent_16G: randomly selected 10% cases from the above training set. In other words, these cases have been included in the complete training set. Participants are allowed to use other criteria to select 10% cases as the coreset.
• 3D_val_npz: validation set
• 3D_val_gt: ground truth of validation set
• CVPR25_TextSegFMData_with_class.json: text prompt for test-guided segmentation task

Interactive 3D segmentation (Homepage)

The training npz files contain three keys: imgs, gts, and spacing. The validation (and testing) npz files don't have gts keys. We provide an optional box key in the npz file, which is defined by the middle slice 2D bounding box and the top and bottom slice (closed interval). Here is a demo to load the data:

npz = np.load(‘path to npz file’, allow_pickle=True)
print(npz.keys())
imgs = npz[‘imgs’]
gts = npz[‘gts’] # will not be in the npz for testing cases
boxes = npz[‘boxes’] # a list of bounding box prompts
print(boxes[0].keys()) # dict_keys(['z_min', 'z_max', 'z_mid', 'z_mid_x_min', 'z_mid_y_min', 'z_mid_x_max', 'z_mid_y_max'])

Remarks:

1. Box prompt is optional to use, where the corresponding DSC and NSD scores are also not used during ranking.

2. Some objects don't have box prompts, such as vessels (filename contains vessel) and multicomponent brain lesions (filename contains brats), becuase the box is not a proper prompt for such targets. The evaluation script will generate a zero-mask and then algorithms can directly start with the point prompt for segmentation.

3. The provided box prompts is designed for better efficiency for annotators, which may not cover the whole object. Here is the script to generate box prompts from ground truth.

Text-guided segmentation (Homepage)

For the training set, we provide a json file with dataset-wise prompts CVPR25_TextSegFMData_with_class.json.

In the text prompts,

• 'instance_label': 1 denotes the instance mask where each label corresponds to one instance (e.g., lesions). It is generated by tumor_instance = cc3d.connected_components(tumor_binary_mask>0)
• 'instance_label': 0 denotes the common semantic mask

For the validation (and hidden testing) set, we provided a text key for each validation npz file

npz = np.load(‘path to npz file’, allow_pickle=True)
print(npz.keys())
imgs = npz[‘imgs’]
print(npz[‘text_prompts’])

Remarks:

1. To ensure rotation consistency, all testing cases will be preprocessed to standard rotation by https://nipy.org/nibabel/reference/nibabel.funcs.html#nibabel.funcs.as_closest_canonical
2. Some datasets don't have text prompts, please simply exclude them during model training.
3. For instance labels, the evaluate metric is F1 score where the order of instance id doesn't matter.