Target Lesions and Dataset Coverage

The target lesions for our study are as follows: 1. Pericardial effusion 2. Pleural effusion 3. Consolidation 4. Ground-glass opacity 5. Lung nodule

However, among the publicly available datasets, only two sets of labels are currently provided: ts_lung_nodules and ts_pleural_pericard_effusion.

Could anyone advise on where to obtain the remaining labels, particularly for consolidation and ground-glass opacity?

From the task description I think the premise of the challenge is that there are not gold-standard labels to train on. At the top of the 'info' page it says: "localize five key thoracic pathologies in 3‑D chest CT volumes—without any voxel‑level labels during training."

The README in the segmentations folder says it is just the outputs from the general segmentation model TotalSegmentator, and that they have not been verified to be correct: https://huggingface.co/datasets/ibrahimhamamci/CT-RATE/blob/main/dataset/ts_seg/README.md I would guess that the TotalSegmentator model wasn't trained to segment consolidation or ground-glass opacity so there are no pseudo-ground truths for those classes.

Hi both,

The task is unsupervised localization training. You can also use external resources to include supervised segmentation labels if they are open source. We currently do not have localization labels for the entire training dataset, but we have labels for our hidden set. I hope this clarifies it.