This repository contains code to generate data for the ISBI 2024 one-page-abstract poster describing the differences between geometric variability and dosimetric effects. This corresponds to the figure on the left in the image below.
The figure on the right is generated using ASTRA, as shown in the corresponding EMBC '23 paper.
TL;DR: measuring the quality of AI-driven automation in the radiotherapy treatment planning process cannot simply be limited to looking at Dice scores for auto-segmentation, or, DVH curves for deep learning dose calculations, but needs to at least be a combination of both. Current literature is awash with new deep learning network architectures promising a test set Dice of 0.9, which is great! However, do the errors that correspond to the remaining volume of 0.1 impact the dose computation adversely? Uncertainties in a preceding step may propagate into complex failure modes into the next: and a holistic quality assurance system must account for these.
For more, see LinkedIn post here.
The data needed to reproduce this is unfortunately not in the public domain - please reach out to me if you'd like to learn more.