Guidance Needed for Using BraTS2023 Dataset — Shape Mismatch

[AshleyAlexJacob]

Hello, and thank you for this great project. I'm currently trying to adapt the codebase for use with the BraTS 2023 dataset and have encountered several shape mismatch errors during preprocessing and training. I’ve tweaked the BraTS dataloader provided in the repo to align with the file naming conventions of BraTS 2023 and prepared the data using the standard four modalities (T1, T1CE, T2, FLAIR) along with the segmentation masks (label values 0, 1, 2, and 4). Despite this, I'm still facing shape mismatch issues—likely due to differences in expected input shapes, label formatting, or other preprocessing steps I might have missed.

Could you please provide guidance on the correct way to integrate BraTS 2023 with this repo? Specifically, it would help to clarify the expected input shape and any preprocessing requirements any metadata files in case of Brats. A dedicated guide or example for BraTS integration—especially for the 2023 version—would be greatly appreciated and valuable to the community. I’d be happy to contribute improvements once things are working on my end.

Thanks again!

[AshleyAlexJacob]

Just to add a bit more context — even when I tweak the code to match the shape requirements in train.py, I run into channel-related errors. Specifically, the SAM model expects a 3-channel input (likely assuming RGB), whereas the BraTS images are single-channel per modality. This channel mismatch leads to errors during model forward passes. Any guidance on how to correctly adapt or preprocess the input to meet SAM's expectations would be very helpful.