Improving performance on LIDC-IDRI

[BorisMuzellec]

The current V-Net implementation achieves a DICE score of ~30 after 100 epochs. This seems fairly low, even though it is hard to assess due to the apparent lack of comparable works. Any suggestions on how to improve this are welcome. Here are a few that I've thought about:

• use data augmentation on patches (e.g. random rotations and shearing)
• segment the lung area and apply a lung mask prior to feeding inputs to the model
• restrict the dataset to the CT scans that were used in the LUNA16 challenge

[BorisMuzellec]

Update: I get a DICE of 41% with patch_size = (128, 128, 128), n_patches = 2 and batch_size = 1 but it's quite computationally intensive.

[Saipraneet]

This repository here might be useful: https://github.com/jaeho3690/LIDC-IDRI-Segmentation/
They seem to get ~50% dice with UNet and some additional augmentations (see notebook)