MICCAI 2021 Daily - Tuesday

When the time does come for treatment, every radiation oncologist or medical physician will have their own way of planning in terms of how they arrange the seeds, with the ultimate goal being to radiate the prostate and nothing else . “ You only have some points where you can put the seeds, so it becomes an optimization problem where you find the optimal solution that radiates the gland, ” Tajwar tells us. “ There are so many rules. You want to minimize the needles so it’s less invasive. You want to use less seeds. You want to avoid hot spots where the seeds are too close together. Most of the techniques in the literature consider how to solve this optimization problem, but neglect to take into account all these rules, so the physicians don’t use it because with their years of experience, they have their own preferences. How can we capture their planning style and formulate that mathematically to generate high-quality plans? That’s the main challenge. ” The team worked with BC Cancer - Vancouver Centre and had a huge database of information from physicians. They propose to use machine learning to not only learn how to make the plans, but also incorporate some of the rules. “ I use generative adversarial networks so it’s not like fully supervised learning, ” Tajwar explains. “ It’s more like the discriminator is learning for each case and each plan whether the physician will accept it. The discriminator is enforcing the generator that makes plans that are similar to the database, so in the end we get plans that our physicians are comfortable with. ” In reality, with this kind of treatment, patients end up with more than one plan because they come in for their scan, the physician makes a plan offline based on that volume, then next time the patient comes in they scan again, 5 DAILY MICCAI Tuesday Tajwar Abrar Aleef