ESTRO 2022 - Abstract Book

S42

Abstract book

ESTRO 2022

Deep learning contouring (DLC) networks can underperform, or even fail, when applied to images which significantly differ in appearance relative to those used for training the networks. This poor generalisation is due to image features varying in a way that the network does not expect. For instance, the bladder can vary in appearance between pelvic contrast-enhanced CT (CECT) and non-enhanced (NECT) scans. This is due to the presence of contrast agent in CECTs, the distribution of which is highly variable among cases. This means DLC networks trained on NECT-only datasets tend to underperform when applied to CECTs. The purpose of this work was to investigate whether generating synthetic training examples is feasible for improving DLC performance when applied to real data. Augmentation of a prostate NECT dataset with synthetic CECTs was investigated. Materials and Methods To perform this investigation, 2 datasets were used: 160 prostate NECTs, and a further 6 clinical excretory-phase prostate CECTs. CTs had corresponding structure sets of organs at risk (OARs), delineated by experienced dosimetrists. A generative appearance model with 8 degrees of freedom (DOFs) was inferred from real CECTs and used to simulate CECT images from NECT data. An example NECT and a corresponding generated CECT are shown in Figure 1. A synthetic CECT was then generated for each NECT, forming an augmented dataset of 320 cases. Values for the DOFs were randomly sampled from the learned distributions. Two separate CNN-based OAR DLC network models were trained, using the same architecture: • DLC : Trained using NECT data only • AugDLC : Trained using both NECTs and synthetic CECTs 80% of each dataset was randomly assigned to network training, 10% to validation, and 10% to testing. Both networks were applied to the 6 clinical CECTs to produce contours. They were also applied to 12 NECTs not used in training. Structures for 4 OARs were autocontoured: the bladder, prostate, rectum, and seminal vesicles. These were compared to the expert delineations using 3D Dice similarity coefficient (DSC) and added path length (APL). APL is correlated to the time required to edit a contour to a clinically-acceptable standard.

Results Results are presented in in Figure 2. Substantial improvements can be seen for CECT bladder contours in both measures. The augmentation also appears to aid performance on NECT bladder delineation. However, other structures show mixed results. Prostate DSCs improved for CECTs on average, whereas some seminal vesicles underperformed on the NECTs.