Unified Framework for Registration and Segmentation of the Spinal Cord and Lesions

[naga-karthik]

Motivation: We usually have datasets with several contrasts but typically only the ground-truth for a single contrast. To get around this, we usually co-register the contrasts, copy the same label for each contrast and then train the model. This co-registration step is usually done independently requires some expertise in image registration (which cannot always be assumed from a user perspective)

Idea: Develop a unified approach that performs the co-registration between contrasts and trains a segmentation model (by concatenating the co-registered contrasts as input).

Approach:

• Step 1. Use SynthMorph approach that already works reasonably for co-registration and it is already implemented in SCT.
• Step 2. Use the synthmorph output as the inputs to the segmentation netowork. There will be two loss functions: (1) for learning the registration properly, and (2) for learning the segmentation based off of the co-registered inputs.

As this is a common problem in medical imaging, there are definitely some works that have done this. Will update the issue once when I find them.

Note: This, for now, is the high-level idea for the project. The approach will surely change in the future!

[simonqueric]

related to ivadomed/ivadomed#659