search

uthmri / Deep-Learning-MS-Segmentation

A deep learning network for brain tissue and lesion segmentation in Multiple Sclerosis
GNU General Public License v3.0
6 stars 1 forks source link

preprocessing steps and sample data #3

Open NeuroImagingWorld opened 3 years ago

NeuroImagingWorld commented 3 years ago

Thank you for the great work and for sharing the code. We would like to apply the model on our dataset and am curious to know how dependent current model is on the data acquisition and preprocessing steps.

We currently have multi-modal images acquired on Siemens Skyra 3T scanner 3D T1-weighted, 3D-Flair (1x1x1 mm3) PD-T2 images acquired at (0.34x0.34x3mm3) all co-registered to T1-w image. N4 bias correction, skull stripping and normalization steps are performed as part of preprocessing.

However segmentation quality seems to have some issues. Can you please share some sample data that we could use to see that the code is setup correctly on our end.

Thanks,

rgabr commented 3 years ago

Thank you for the great work and for sharing the code. We would like to apply the model on our dataset and am curious to know how dependent current model is on the data acquisition and preprocessing steps.

We currently have multi-modal images acquired on Siemens Skyra 3T scanner 3D T1-weighted, 3D-Flair (1x1x1 mm3) PD-T2 images acquired at (0.34x0.34x3mm3) all co-registered to T1-w image. N4 bias correction, skull stripping and normalization steps are performed as part of preprocessing.

However segmentation quality seems to have some issues. Can you please share some sample data that we could use to see that the code is setup correctly on our end.

Thanks,

Thank you for the interest in this work. The steps you describe are similar to what we do in our lab, except that we do bias correction on the skull-stripped images. We also started the pipeline with an anisotropic diffusion filter for denoising. The intensity normalization step might be a source of contrast variations and may be a factor in how good the model works, especially that your data is a mix of 2D and 3D (our training data were all 2D). I'm afraid we do not currently have permission from the clinical trial PI to share the MR images. You can perhaps refine the model by retraining on few examples of labelled data from your dataset.