ESTRO 2021 Abstract Book

S442

ESTRO 2021

Belgium; 4 University Hospital RWTH Aachen University, Nuclear Medicine and Comprehensive diagnostic center Aachen, Aachen, Germany; 5 University Hospital Zürich and University of Zürich, Department of Radiation Oncology, Zürich, Switzerland; 6 Maastricht University Medical Centre+, Surgery, Maastricht, The Netherlands; 7 Cliniques Universitaires Saint-Luc, Radiation Oncology, Brussels, Belgium; 8 Affiliated Zhongshan Hospital of Dalian University, Radiology, Dalian, China; 9 Radboud University Medical Center, Radiation Oncology, Nijmegen, The Netherlands; 10 Maastricht University Medical Centre+, Radiology and Nuclear Medicine, Maastricht, The Netherlands; 11 Maastricht University Medical Center+, Pulmonary Diseases, Maastricht, The Netherlands; 12 University of California San Francisco, Radiation Oncology, San Francisco, USA Purpose or Objective Localizing and delineating tumors is essential for radiotherapy planning and quantitative imaging workflows. However, manual contouring is highly laborious and time consuming, and prone to variability and poor reproducibility. To address these issues we created a fully automated pipeline for detecting and segmenting lung tumors on CT images to facilitate adaptive re-planning and automated RECIST. Additionally, we have evaluated the prognostic power of produced segmentations. Materials and Methods Multi-centric CT images from 1328 NSCLC patients with appropriate tumour contours were used to train, test, and validate our detection and segmentation method. A three-step approach was developed, consisting of data pre-processing, lung isolation and tumor segmentation (Figure 1). A pre-processing algorithm was developed to standardize images within a heterogeneous dataset with regard to hardware and acquisition. A 2D U-net type convolutional neural network architecture with volumetric post-processing was trained on 999 CT scans. Quantitative performance was evaluated on the test (93 patients) and validation (236 patiens) sets using Dice similarity coefficient (DSC), Jaccard index (J), and 95th Hausdorff distance (H95th). A registered “in-silico” clinical trial was performed to evaluate following endpoints: comparison of auto vs manual contouring time, contours reproducibility and the preference score. To obtain a preference score we enrolled 40 participants, and developed a qualitative assessment tool, which allows for scoring the segmentations while recording the assessor’s expertise level. Prognostic power of segmentations was assessed using measurements (RECIST and tumor volume) extracted from automatically generated and manual contours for the datasets with the available survival data.

Results Proposed method achieved a median slice-wise detection accuracy of 0.93 (IQR=0.08) and ROC AUC = 0.89 (0.89- 0.90) and segmentations achieved a DSC = 0.77 (IQR = 0.23), J =0.62 (IQR =0.29) and H95th= 10 mm on the validation dataset. On average, the participants preferred the automatic segmentation above the expert's contour in 55% (IQR=12%) of the cases (Fig. 2a). Qualitative preference scores across the groups are shown on the Figure 2b. The mean segmentation time for the automated method was 2.77 sec/patient (SD = 0.44), whereas the mean time for manual segmentation was 172.19 sec/patient (SD = 158.98). The AI contours were 100% reproducible whereas the average DSC of experts intervariability was 0.83 (IQR = 0.12).