ESTRO 2021 Abstract Book

S410

ESTRO 2021

Conclusion Our findings provide evidence that supports the wider clinical implementation of MR-only planning for anal and rectal cancers.T2 MR sequences (which are optimal for anal and rectal cancer target delineation) can also accurately be used for sCT generation for anal and rectal cancer sites. Changing reference image modality has minimal impact on registration accuracy. OAR dose reduction, due to the use of MR, could lead to improved patient outcomes if OAR dose reductions translate into less treatment related toxicity. OC-0525 HarMonAE: Zero-shot, unsupervised harmonisation of multi-scanner MRI for radiotherapy applications M. Nix 1 , M. Tyyger 1 , K. Fatania 2 , B. Al-Qaisieh 1 1 Leeds Cancer Centre, Medical Physics, Leeds, United Kingdom; 2 Leeds Teaching Hospitals NHS Trust, Radiology, Leeds, United Kingdom Purpose or Objective Robust deep-learning (DL) and radiomics with MR imaging for radiotherapy requires minimisation of intra- scanner variability, but generalisable harmonisation by removal of scanner-specific features is very challenging. Previous deep-learning models require paired training data (patients imaged on source and target scanners) or are specific to a single source scanner. Ideally a harmonisation model would handle images from multiple scanners, including unseen ones, without retraining (zero-shot learning). We demonstrate HarMonAE, an un-paired multi-scanner/vendor harmonisation model, on T1w brain MR, including zero-shot harmonisation of data from an unknown scanner. Hence, HarMonAE has potential to improve generalisability for any deep- learning or radiomic analyses of MR in RT. Materials and Methods A disentangled-representation autoencoder GAN (HarMonAE - fig. 1a) was trained to harmonise MR images from 5 scanners. The residual encoder learned a style-blind latent ‘content’ representation adversarially, against a scanner ID classifier. Scanner ID was explicitly provided to the decoder, allowing full reconstruction of input images, whilst the latent ‘content’ representation became scanner-independent, enabling zero-shot harmonisation. Altering the provided scanner ID to that of the target scanner (GE 3T), enabled harmonisation. Adversarial loss ensured similarity to the target contrast and encouraged sharpness. HarMonAE was trained for 30 epochs on 53986 images from 5 of the 6 scanners in the CC-359 T1w brain dataset, and tested on 11364 withheld images. All data from the final scanner (Philips 1.5T) were excluded from training for zero-shot harmonisation analysis.