inspiration fromVoxelMorphand its successor works, which took the problem of deformable image registration, formulated it as the loss function of a deep neural network, and then trained the neural network to solve it. “This, for me, is an excellent empirical but also elegant way to derive a solution to a problem,” he says. “You can have 20 pages of excellent theory explaining why your iterative algorithm works and converges well, but in the end, the proof of the pudding is in the validation. You don’t start from the principles of trying to solve your problem; you define what the parameters are for your problem to have been solved, and then, using gradient descent, you search for a solution.” Interestingly, Richard built the prototype for this method several years ago, but when you build a system for calculating cortical thickness, how do you validate it if there is no ground truth? Luckily, a paper by Rusak et al. last year had the answer: a synthetic phantom built using a GAN. “They generated 20 subjects with different levels of cortical thickness reduction and showed that a predecessor method to ours was very sensitive to these reductions,” he tells us. “It’s more sensitive than FreeSurfer. This work gave me the final piece of the puzzle. Now, I have a dataset coming from an independent group. It’s one thing to say my deep learning approach is close to the existing approach, but can it do the same job of resolving 8 DAILY MICCAI Wednesday Poster Presentation
RkJQdWJsaXNoZXIy NTc3NzU=