ESTRO 2021 Abstract Book

S1399

ESTRO 2021

2 Newcastle

Purpose or Objective A key step in radiotherapy planning is delineation of tumour volume and Organs At Risk (OARs). Currently, this is undertaken manually by clinicians, which is laborious, time consuming and prone to inter-observer variability. The improved soft-tissue contrast of Magnetic Resonance (MR) images has shown to reduce manual delineation variability, and so may also benefit automatic methods. This study aimed to evaluate the feasibility of using convolutional neural networks to automate organ delineation of prostate radiotherapy planning MR images. Materials and Methods MR images from 49 patients treated with MR-only radiotherapy for prostate cancer were included. These were randomly split into a training (n=32), validation (n=7) and test (n=10) sets. The clinical manual delineations of the bladder, rectum, prostate, and prostate and Seminal Vesicles (SV) were included. A single fully-convolutional 2D DenseNet model was trained from 2D MRI slices (in all three anatomical planes) to simultaneously segment each of the four structures. Training incorporated several techniques and recent innovations, including image augmentation and use of focal cross-entropy loss. The trained model was evaluated using the test set, where 2D slice segmentations generated by the model in each plane were combined to form 3D delineations of each structure. The model delineations were compared to clinical manual delineations using the Dice metric and the mean expansion of the automatic contour required to cover 95% of the clinical contour. In addition two clinical delineators independently and blindly rated both the model and clinical contours or each test patient on a 4- point scale (delete and re-delineate, major modification, minor modification, accept). Results Quantitative results showed high agreement between model and clinical delineations (table 1). 10/10 of bladder and rectum model delineations and 6/10 of prostate and prostate+SVs delineations were rated as acceptable/minor modification by both assessors, while every single delineation was marked as acceptable/minor modification by at least one of the assessors. Clinical delineations were rated higher than model delineations, although not all clinical delineations were acceptable and one assessor rated one of the clinical delineations as major modification, highlighting the importance of peer review.

Conclusion Of the 40 total delineations produced by the MR-based neural network model, all were assessed as acceptable/minor modification by at least one of the assessors, with 32 so assessed by both assessors, indicating that the model can reliably produce delineations that can act as a useful starting point for clinicians (particularly for OARs). This suggests that MR-based automated organ delineation within the MR-only prostate radiotherapy pathway is feasible, subject to review and correction by an appropriate clinician.

PO-1676 Validating CBCT to CT registration QC using an AI generated dataset. J. Sage 1 , P. Looney 1 , R. Chuter 2 , G. Price 3 , D. Boukerroui 1 , D. Balfour 1 , P. Whitehurst 2 , M.J. Gooding 1 1 Mirada Medical Ltd, Science, Oxford, United Kingdom; 2 The Christie NHS Foundation Trust, Christie Medical