Support for DICOM images

[FeryET]

My actions before raising this issue

• [x] Read/searched the docs
• [x] Searched past issues

Expected Behaviour

I want to work with dicom files interactively (for medical images).

Current Behaviour

Currently, there is a script that converts dcm to jpg/png, which later is fed as the unlabelled data to CVAT instance. This approach is wrong, since many of the medical tasks require interactive preprocessing, such as interactive adaptive thresholding for artifact removal.

Context

I am trying to come up with a labelling system for our medical imaging analysis lab here at university, and no software really supports DCM files. I have had great experience working with CVAT in other contexts before, and seeing this software supporting DICOM natively will really make this process far easier for me. There are dedicated solutions, like VinDr Lab, that are not as stable as CVAT.

I can also help in the development process of this feature if you ever greenlight it.

[bsekachev]

Hi, @FeryET

As it was described in the corresponding article, we believe that DICOM is very important feature, but it is also is too huge (from point of view variability of DICOM files and related to them annotations scenarious). Currently for our team it would be unbearable to develop and support them. Probably internally we will wait for a huge customer with specific and clear requests before starting any activities on this. BTW, you can also help us sharing your specific use-cases (what data you are annotating, what tools would be useful, screenshots, examples, etc). With a big number of such requests we would have better vision on what is demanded in the community.

We still hope CVAT's community to be able to develop something like that, if you are ready to contribute your time, we will be happy to help. Thx for undestanding.

[FeryET]

Thank you for your detailed response.

Currently, our use case is trying to label and segment CT images of brain. As you know, these images are 3D, hence they could be likened to video frames in a way. I think writing some basic adapters for DICOM data will at least cover a lot of simpler use cases such as ours.

The main issue with the current approach is the lack of fine grained control that the labeller should have on how the dicom image is rendered on the screen. The data should be adaptively thresholded and denoised, controlled by the annotator. Each modality and each body part has its own thresholding strategy and etc.

[dudulasry]

I'd also love to see a support for DICOMs in the future. We're thinking about adopting CVAT as our annotation tool, but the lack of support for DICOMs and medical imaging is a deal breaker for us.

Nevertheless, I really love this project and I hope to contribute to it in the future.

[nmanovic]

@dudulasry , thanks for the feedback.

[Stijnp]

Interested too, my DICOMs are 2D videos by default - currently I'm converting them to a video extension but would be interested to see this implemented natively.

[thirawat69]

Thank you for your detailed response.

Currently, our use case is trying to label and segment CT images of brain. As you know, these images are 3D, hence they could be likened to video frames in a way. I think writing some basic adapters for DICOM data will at least cover a lot of simpler use cases such as ours.

The main issue with the current approach is the lack of fine grained control that the labeller should have on how the dicom image is rendered on the screen. The data should be adaptively thresholded and denoised, controlled by the annotator. Each modality and each body part has its own thresholding strategy and etc.

If you don't mind, may I ask what labeling program are you currently using for the dicom files?