ESTRO 2021 Abstract Book

S584

ESTRO 2021

used to achieve CBCT-based dose calculation for breast cancer adaptive radiotherapy.

PD-0753 Unsupervised AutoConfidence estimation for deep-learning synthetic-CT in MR-only liver radiotherapy M. Nix 1 , D. Bird 1 , M. Tyyger 1 , B. Al-Qaisieh 1 1 Leeds Cancer Centre, Medical Physics, Leeds, United Kingdom Purpose or Objective Per-patient estimates of AI prediction confidence are needed to enable good AI-assisted clinical decision making in RT. Deep-learning synthetic CT (sCT) generation requires spatially resolved confidence maps to identify error regions. AutoConfidence (AC) produces such maps but requires paired ground-truth CT (supervised learning). However, most clinically relevant scenarios require unsupervised learning due to unavailability of paired data. Here, we demonstrate that paired ground truth is unnecessary in general and AutoConfidence can be co-learnt with an unsupervised cycleGAN, as used clinically for sCT generation. We report unsupervised AutoConfidence (uAC) derived confidence maps for liver sCT, improving clinical confidence in the MR-only RT pathway. Materials and Methods Unsupervised AutoConfidence (uAC) extends the cycleGAN architecture with shallow U-net AutoConfidence discriminators to produce voxel-resolution confidence maps (fig. 1). These discriminators assess the likelihood of the prediction, given the input image. During training, it is necessary to learn on ‘real’ examples of CT and their associated MR. As this paired data does not exist, the ‘real’ examples are evaluated against the predictions of the reverse half of the network. This initially poor surrogate for ground-truth improves during training, allowing uAC to learn in a ‘self-supervised’ fashion. We trained uAC using unregistered T2w MR and planning CT data from 36 liver RT patients and validated on 10 further cases.

Results In the unpaired scenario, AutoConfidence is still able to learn to detect regions of the predicted image which are low quality or erroneous. Commonly highlighted areas are bone, which is known to exhibit low density on cycleGAN derived sCT. Other errors detected include missing bones, missing arm anatomy which has been ‘imagined’ by the GAN, bone appearing spuriously in the abdomen and poorly defined mediastinal structures. Examples of these sCT failure modes and their detection by AutoConfidence are shown in fig. 1. Whilst absolute qualitative validation is not possible without ground truth CT, the fraction of sCT slices with > 2% error (uAC score > 0.5) by volume was observed to decrease during cycleGAN training, from 73% after 5 epochs to 17% after 20 epochs and 6% after 100 epochs, indicating correlation of uAC score with sCT quality. Visual comparison of MR, sCT and AutoConfidence overlays shows that detected errors correspond to clear anatomical differences between MR and sCT.