Computer Vision News - September 2022

37 Deformable Image Registration Uncertainty... anatomy of the patient is each day, ” Andreas explains. “ We re-optimize their treatment plan based on this. Usually, the planning process takes around a week and involves many doctors, but now, when we scan the patient, we want to treat them as soon as possible so that their anatomy on the scan is as close as possible to the anatomy we’re treating. ” To achieve this speed, the steps that usually happen manually, such as tissue segmentation , must be performed automatically. Another factor to consider is dose accumulation. After a treatment, it is essential to know each organ’s dose as it indicates the level of organ damage. “ We have 30 different anatomies on 30 different days, and each has a dose map, ” Andreas tells us. “ Wewant to know the total dose at the end, but you can’t just sum the doses because the anatomies are slightly different. This is where image registration comes in. We take one scan as a reference, and then the daily scans are all deformed or registered to that reference. We apply the same transformation to the doses, so we deform the doses to the reference, and then we can sum them up. ” Deformable image registration is not an easy problem . It is called ill-posed in the community, meaning there are more degrees of freedom than constraints. When you run two different algorithms, Proton radiotherapy uses protons instead of photons, delivering the dose more precisely to the tumor and sparing more of the organs at risk around it. However, the depth the protons can travel depends on the amount of tissue in the beam path. A patient’s anatomy can change throughout the treatment. For instance, people generally lose weight. With a treatment plan based on a CT scan taken at the beginning of the process, this change can cause the protons to end up in a different place than planned , degrading their precision. “ In adaptive radiotherapy, we take a daily scan so that we know what the exact