Pretrain model

We are building application to analyse CT scans of brain. The expected outcome is segmentation of liaison as well classification. What is the Pretrained model I can make use of for our training purpose?

Thanks for your interest in Clara Train SDK. Please note our latest release clara train V4.0 based on MONAI which uses PyTorch. Please check out the notebooks to get you started clara-train-examples/PyTorch/NoteBooks at master · NVIDIA/clara-train-examples · GitHub

I would need more information about your problem and data. If you are classifying different type of tumors you can either segment then as different labels as the brain segmentation does. (please take a look at the multi-label brats models).

Another approach would be to segment the tumor then do classification on the patches. so you would train 2 models 1 to do segmentation of the tumor then another to do the classification. In that case please take a look at any ct segmentation (liver, spleen) and the classification would be the chest xray task

you can find all our models on NGC AI Models - Computer Vision, Conversational AI, and More | NVIDIA NGC

Hope this helps

Thanks a lot @aharouni
Though I raised lot of questions in Clara forum, I got first time the response.

Where can I find multi-label brats models
To give overview of what we planned to achieve. We are planning to ascertain if there’s any brain stroke or not using CT scans. We also planned to show the liaison as highlight and classify the sub type of stroke. When come to Pretrained models, do you suggest to consider model which related to brain or any other parts of body?


All of the our models are trained on public datasets and packaged as MMARs Medical model archive. MMARs contains training scripts with all the parameters as well as the final model. you can find all our models on NGC AI Models - Computer Vision, Conversational AI, and More | NVIDIA NGC

I am not sure that we have a specific pretrained model to give you a head start. MRI have large intensity range >0 if you check the train config we do a normalization to bring the range to 0-1. However CT are more standard as the intensities are HU units. so the normalization step is to set it to a certain window level as -200 to 200 then map it to 0-1, you would need to adjust these levels.

As I mentioned before depending on the classification diversity / number of classes / number of training images you have you may either:
1- Segment each tumor with different label
2- Do one segmentation then train another model to classification
3- do a multi head CNN to produce the segmentation and the class

If the tumors are from the same family and the classification is depended on the location as sub Dural verses ventricle locations then #1 does NOT make sense. If the tumors types are hard to differentiate then #2 makes more sense as you can focus on detection/ segmentation then classification separately

Hope that helps

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.