Thank you and inquiry about MR preprocessing

[kulsoom-abdullah]

First, I want to express my sincere gratitude for TotalSegmentator. Your open-source contribution is invaluable to the medical imaging community.

I'm currently working on an L3 vertebra detection/selection model using TotalSegmentator as part of my workflow. For this project, I'm particularly interested in understanding the MR-specific preprocessing steps that occur before segmentation.

Could you please point me to the part of the codebase where MR preprocessing is implemented? I've looked through the nnUNet_predict_image function but haven't been able to identify the MR-specific steps.

Any guidance would be greatly appreciated. Thank you for your time and for creating this fantastic tool!

[wasserth]

The preprocessing is rather simple. TotalSegmentator resamples the images to 1.5mm isotropic resolution, then the images go to nnU-Net. nnU-Net will normalise the images, if I remember correctly to mean of 0 and stddev of 1. But this might have changed a bit. This information can be found in the nnunetv2 documentation. I hope this helps. The most relevant code start at totalsegmentator/python_api.py.

[meeselizabeth]

Hi, similar question here! First of all, thank you for your public datasets! However, I am using your MR data to fine tune my model on, but the dice score stays superlow. I assume because of the way I preprocess/transform. Are the images in the dataset already in RAS orientation and already have right spacing? I do see your scripts for alignment and resampling, but I can't find out if this is already applied or still need to happen. So, what are the steps of preprocessing / transforms that still need to be done after loading the nifti files? Hope to hear from you soon!

[wasserth]

If you can share one of your datasets with me I can try to reproduce the issue. Otherwise it is difficult to say what the problem might be.

[meeselizabeth]

Hi, thx for your response. I downloaded this dataset: https://zenodo.org/records/11367005. And then this is what I do:

class TotalSegDataset3DMRI:
    def __init__(self, data_dir):
        self.data_dir = data_dir
        self.subjects = [subject for subject in os.listdir(self.data_dir) if os.path.isdir(os.path.join(self.data_dir, subject))]

    def __len__(self):
        return len(self.subjects)

    def __getitem__(self, idx):
        subject_id = self.subjects[idx]
        subject_dir = os.path.join(self.data_dir, subject_id)
        mri_path = os.path.join(subject_dir, "mri.nii.gz")

        # Load combined segmentation map
        combined_label_path = os.path.join(subject_dir, "combined_mask.nii.gz")

        data_dict = {'image': mri_path, 'label': combined_label_path}

        return data_dict

train_transform_mri = Compose([
    LoadImaged(keys=["image", "label"], reader='NibabelReader'),
    EnsureChannelFirstd(keys=["image", "label"]),
    Orientationd(keys=["image", "label"], axcodes="RAS"),
    Spacingd(keys=["image", "label"], pixdim=(1.5, 1.5, 1.5), mode=("bilinear", "nearest")),
    ScaleIntensityd(keys=["image"], minv=0.0, maxv=1.0),  
    Resized(keys=["image", "label"], spatial_size=(128, 128, 128)),
    # SpatialPadd(keys=["image", "label"], spatial_size=(128, 128, 128)),
    # RandCropByPosNegLabeld(keys=["image", "label"], spatial_size=(128, 128, 128), label_key='label', pos=0.5, neg=0.5),
    RandRotate90d(keys=["image", "label"], prob=0.10, max_k=3),
    RandShiftIntensityd(keys=["image"], offsets=0.10, prob=0.20),
    ToTensord(keys=["image", "label"]),
])

val_transform_mri = Compose([
    LoadImaged(keys=["image", "label"], reader='NibabelReader'),
    EnsureChannelFirstd(keys=["image", "label"]),
    Orientationd(keys=["image", "label"], axcodes="RAS"), 
    Spacingd(keys=["image", "label"], pixdim=(1.5, 1.5, 1.5), mode=("bilinear", "nearest")),
    ScaleIntensityd(keys=["image"], minv=0.0, maxv=1.0),
    Resized(keys=["image", "label"], spatial_size=(128, 128, 128)),
    ToTensord(keys=["image", "label"]),
])

The segmentations I use are the organs from class_map_parts_mr.

[wasserth]

For this dataset it should work well. You do not need any of this code. Just get one of the nifti files and then run TotalSegmentator -i your_downloaded_file.nii.gz -o segmentations -ta total_mr

[meeselizabeth]

Thank you! I already have the data stored as subjects (s0001 etc) with their corresponding mr.nii.gz and then their segmentations folders. I created a function to instead of having every segmentation in a separate file (like aorta.nii.gz), have all segmentations in one map, stored these in combined_mask.nii.gz for each subject. So do I still need to do TotalSegmentator -i your_downloaded_file.nii.gz -o segmentations -ta total_mr ? My use case is tuning my pre-trained segresnet on this data.