Capability to segment lung

[varghesealex90]

Hi,

I modified the inference_examples_RGB notebook by using data examples/LIDC-IDRI-0140_143_280_CT_lung.png. Additionally, I made the text prompt to be "Lung".

image_path = 'examples/LIDC-IDRI-0140_143_280_CT_lung.png'
text_prompt = ['lung']

image, pred_mask = inference_rgb(image_path, text_prompt)

However, the result segments the nodule in the lung.

Any pointers why it works this way?

[theodore-zhao]

Hi, this is because during the training, the model did not see any labels of "lung" in CT. Therefore, it will segment the thing that is closest in semantics in the image (e.g. "lung nodule" instead of "lung" for CT). We expanded the training dataset for 3D images, and now have a new model version on the v2 branch. The new model version covers a much wider variety of anatomies in various 3D modalities. Please feel free to tyr it out!

[youluexx]

What does the number 140 mean in the name LIDC-IDRI-0140_143?