search

wasserth / TotalSegmentator

Tool for robust segmentation of >100 important anatomical structures in CT and MR images
Apache License 2.0
1.42k stars 237 forks source link

Poor cardiac segmentations, including highres #334

Open yakovdan opened 2 months ago

yakovdan commented 2 months ago

Hello, I'm using the latest TotalSegmentator to segment the heart in CT scans. Visually inspecting the output I can see that the output is very roughly correct but large parts of the heart remain unlabeled. I suspect there's an issue with my preprocessing.

My CT scans are composed of 512x512 slices with a pixel spacing of about 0.7mm and a slice thickness of 1.25mm. Axially, the scans are variable but include the chest and upper abdomen. Since TotalSegmentator expects 1.5mm slices, I've accounted for that by resampling. I've examined small dataset, a subset of TotalSegmentator's train dataset, and noticed that the slices are 255x255. I've also noticed that the orientation in the training dataset for Total Segmentator is such that the positive axial direction is from feet to head. I couldn't find any information on pixel spacing and any information at all regarding the heart_highres dataset, neither on github nor in the paper.

Can you please specify in detail the input format parameters that TotalSegmentator expects, both for the "total" task and for "heart_highres"? I mean input dimensions in all axes, pixel spacings, slices thickness, orientations, etc.

Thank you

wasserth commented 2 months ago

TotalSegmentator expects as input the original CT image and typically works well with any image setting. So you do not need anything specifically. You have to make sure that you never change the value range of your image before passing it to totalsegmentator. It should contain the original HU values.