Train a 2D detection model for spinal cord MS lesions

[plbenveniste]

Projects steps :

• Literature review of SOTA 2D detection models
• Manual segmentation of 20 subjects
• Data pre-processing (train, valid, test)
• Model building
• Training of the model
• Definition of the metrics
• Improvement strategies : data augmentation...

Multiple Sclerosis (MS) lesion datasets: Canproco (PSIR, STIR)

[naga-karthik]

Just thinking about this and it reminded of the approach I was considering initially -- have you taken a look at Mask RCNN? it's basically a segmentation + detection model combined into one. When I looked at it, there were no 3D models and the combined approach was a bit complicated (i.e. it was not easy to get the bounding box predictions, cuda-issues, etc). But maybe things might have changed now? MaskRCNN is also quite old but there might be new models that do detection + segmentation a bit better?

Secondly, when I was thinking about this nnDetection was just starting and not as developed. Maybe you can also consider looking at it too? especially now since we're all experts in training nnUnets ;)