ESTRO 2022 - Abstract Book

S275

Abstract book

ESTRO 2022

1 The Clatterbridge Cancer Centre, Medical Physics, Liverpool, United Kingdom

Purpose or Objective Commercial software can be used to automatically delineate OARs with the potential for significant efficiency savings in the radiotherapy treatment planning pathway and simultaneous reduction of inter- and intra-observer variability. Vendors of commercial systems often claim superiority of their own system in comparison to competitor systems. To date there has been limited research comparing multiple systems using multiple comparison metrics and a common patient cohort. This has been addressed in this study. Materials and Methods Four different deep learning-based auto-segmentation systems, which had been independently developed for commercial use, were used to create five commonly used head and neck (H&N) OARs (brainstem, spinal cord, mandible, left and right parotid), for 30 H&N patient datasets. All systems were running their latest available software version at the time of study (June 2021 – Sep 2021). The resulting auto-segmented contours were compared to ‘gold standard’ clinical contours, created by Consultant Clinical Oncologists at our centre. All data used originated from patients entered into the PATHOS clinical trial. The associated trial protocol includes clear anatomical guidelines for OAR delineation and, in addition, trial entry involved pre-trial OAR outlining Quality Assurance, which all Oncologists were required to undertake. A sample of patient data was retrospectively reviewed during the trial, to provide further assurance around the quality of contours used. Standard similarity metrics of 3D Dice Similarity Coefficient (DSC) and Added Path Length (APL) were utilised for the study. Results Table 1 contains mean and one standard deviation data for both metrics, for all OARs and all systems tested. Values obtained for both 3D DSC and APL correlate well with other recent published studies. Performance differences between the four systems were statistically insignificant for both 3D DSC and APL metrics.

Conclusion Comparable levels of performance were observed between all four systems. This indicates that deep learning-based auto- segmentation products are developing at a similar pace in terms of the quality of contours produced. It is therefore likely to be more beneficial to consider other factors such as cost and range of contours offered when considering the evaluation of such a system for clinical use.

PD-0314 An explainable deep learning pipeline for multi-modal multi-organ medical image segmentation

E. Mylona 1 , D. Zaridis 1 , G. Grigoriadis 2 , N. Tachos 3 , D.I. Fotiadis 1

1 University of Ioannina, Department of Biomedical Research, FORTH-IMBB, Ioannina, Greece; 2 University of Ioannina, Unit of Medical Technology and Intelligent Information Systems, Materials Science and Engineering Department, Ioannina, Greece; 3 University of Ioannia, Department of Biomedical Research, FORTH-IMBB, Ioannina, Greece Purpose or Objective Accurate image segmentation is the cornerstone of medical image analysis for cancer diagnosis, monitoring, and treatment. In the field of radiation therapy, Deep Learning (DL) has emerged as the state-of-the-art method for automatic organ delineation, decreasing workload, and improving plan quality and consistency. However, the lack of knowledge and interpretation of DL models can hold back their full deployment into clinical routine practice. The aim of the study is to develop a robust and explainable DL-based segmentation pipeline, generalizable in different image acquisition techniques and clinical scenarios. Materials and Methods The following clinical scenarios were investigated: (i) segmentation of the prostate gland from T2-weighted MRI of 60 patients (543 frames), (ii) segmentation of the left ventricle of the heart from CT images of 11 patients (1856 frames), and (ii) segmentation of the adventitia and lumen areas of the coronary artery from intravascular ultrasound images (IVUS) of 42 patients (4197 frames). The workflow of the proposed DL segmentation network is shown in Figure 1. It is inspired by the state-of-the-art ResUnet++ algorithm with the difference that (i) no residual connections have been used and (ii) the addition of a squeeze and excitation module to extract interchannel information for identifying robust features. The model was trained and tested separately for each clinical scenario in 5-fold cross-validation. The segmentation performance was assessed using the Dice Score (DSC) and the Rand Error index. Finally, the Grad-CAM technique was used to generate heatmaps of feature (pixel) importance for the segmentation outcome. This is an indicator of model uncertainty that reflects the segmentation ambiguities, allowing to interpret the output of the model.