Skip to content

Latest commit

 

History

History
71 lines (60 loc) · 4.09 KB

common_code_questions.md

File metadata and controls

71 lines (60 loc) · 4.09 KB
Raw

FAQ

How to process noisy label problem in CT dataset?

In some datasets, the annotation only includes organs and does not include tumors. In these cases, the tumors are improperly labeled as organs. To address this issue, we treat tumors as organs as well (for example, the liver tumor is also part of liver) and use the one vs. all manner for prediction. Each channel only predicts whether the voxels belong to the corresponding organ or tumor (voxels of liver tumors will be predicted as liver, as well as liver tumors). Only the labeled ground truth are used to back-propagate and update the network. This manner avoids the noisy label (false negative ground truth) problem for tumor segmentation.

How to train new dataset with new organ?

  1. Set the following index for new organ. (e.g. 33 for vermiform appendix)
  2. Check if there are any organs that are not divided into left and right in the dataset. (e.g. kidney, lung, etc.) The RL_Splitd in label_transfer.py is used to processed this case.
  3. Set up a new transfer list for new dataset in TEMPLATE (line 58 in label_transfer.py). (If a new dataset with Intestine labeled as 1 and vermiform appendix labeled as 2, we set the transfer list as [19, 33])
  4. Run the program label_transfer.py to get new post-processing labels.
  5. Rename the directory for new dataset with number prefix, e.g. 13. (This is used to identify which dataset the data comes from)
  6. Generate txt file with postfix xxx_train.txt, xxx_val.txt and xxx_test.txt for new dataset and add the name to --dataset_list argument in train.py validation.py test.py.

The current Template is shown below

Index Organ
1 Spleen
2 Right Kidney
3 Left Kidney
4 Gall Bladder
5 Esophagus
6 Liver
7 Stomach
8 Aorta
9 Postcava
10 Portal Vein and Splenic Vein
11 Pancreas
12 Right Adrenal Gland
13 Left Adrenal Gland
14 Duodenum
15 Hepatic Vessel
16 Right Lung
17 Left Lung
18 Colon
19 Intestine
20 Rectum
21 Bladder
22 Prostate
23 Head of Femur Left
24 Head of Femur Right
25 Celiac Truck
26 Kidney Tumor
27 Liver Tumor
28 Pancreas Tumor
29 Hepatic Vessel Tumor
30 Lung Tumor
31 Colon Tumor
32 Kidney Cyst

How to generate specific prediction for different dataset?

  1. Ensure that the dataset is properly loaded and preprocessed so that it can be used by the model.
  2. Use MERGE_MAPPING_v1 in utils/utils.py to control the label mapping from universal model to specific dataset.
  3. In the tuple, the first item is the index of the label in the universal model's template, and the second item is the index of the label in the specific dataset. For example, if the label for "liver" is indexed as 1 and the label for "liver tumor" is indexed as 2 in the LiTS dataset, the mapping tuple would be (6,1), (27,2).
  4. Finally, you may want to post-process the final results as needed, as per your specific requirements.

How to inference pseudo label without postlabel?

We add pred_pseudo.py file, where you can save the generated pseudo label without post label. python -W ignore test.py --resume MODEL.pth --data_root_path YOUR_DATA_DIR

How to customize CLIP embedding for your own dataset?

The clip embedding is generated with text encode in CLIP. We offer example in pretrained_weights/clip_embedding.py and you should revise ORGAN_NAME according to your dataset.