search

duyhominhnguyen / LVM-Med

Release LMV-Med pre-trained models
176 stars 22 forks source link

Can you support computation on 3D DICE/IOU scores? #3

Closed GewelsJI closed 1 year ago

duyhominhnguyen commented 1 year ago

Hi @GewelsJI, yes, we already involve result estimation with 3D IOU scores in 3D segmentation downstream tasks.

You can run an example by python train_segmentation.py -c ./dataloader/yaml_data/mmwhs_ct_endtoend_R50.yml (MMWHS-CT dataset) or python train_segmentation.py -c ./dataloader/yaml_data/mmwhs_mr_endtoend_R50.yml (MMWHS-MRI dataset).

These scripts will call the function evaluate_3d_iou in evaluate.py. Hope it help!

GewelsJI commented 1 year ago

Thanks. But it does not support calculation of 3D dice score, could you provide some available reimplementations?

duyhominhnguyen commented 1 year ago

@GewelsJI Oke, we have updated the evaluate_3d_dice in evaluate.py for 3D Dice. You can check it now!

GewelsJI commented 1 year ago

Big thanks. Another question is how to download ~1.3 million medical images that you mentioned?

duyhominhnguyen commented 1 year ago

We summarize all used datases in the Table 13, Section F in the Appendix of LVM-Med's paper. You have to access each link and download one-by-one by yourself. Due to data privacy, we cannot host the dataset.

GewelsJI commented 1 year ago

That's good. BTW, strongly recommend you a largest colonoscopy video dataset for polyp segmentation: https://github.com/GewelsJI/VPS/tree/main

duyhominhnguyen commented 1 year ago

Great suggestions! We would consider training another version of LVM-Med (SSL) including your dataset.