ESTRO 2021 Abstract Book

S767

ESTRO 2021

Poster discussions: Poster discussion 34: Deep-learning for auto-contouring

PD-0926 Large scale analysis of the clinical implementation of deep learning contouring in the thorax region F. Vaassen 1 , R. Canters 1 , I. Lubken 2 , J. Mannens 2 , W. van Elmpt 1 1 Maastro Clinic, Physics Innovation, Maastricht, The Netherlands; 2 Maastro Clinic, Clinic, Maastricht, The Netherlands Purpose or Objective The quality of automatic contouring has been widely studied on the geometrical level by comparing manual delineations and user-adjustments, but mostly this is done in a limited patient cohort. This study aimed to evaluate the extend of manual adjustments following auto-contouring of organs-at-risk (OARs) in the thorax region in clinical practice for a large patient group. Materials and Methods In the period March 2020 - December 2020, in total 322 lung cancer (LC) and 295 breast cancer (BC) patients were contoured in clinical routine using deep learning contouring (DLCExpert, Mirada Medical Ltd., Oxford, United Kingdom). For the LC group, left lung, right lung, heart, esophagus, spinal cord, and mediastinum were contoured. For the BC group, left lung, right lung, heart, CTVp1, and contralateral breast were delineated, and the esophagus and thyroid when L3-L4 or parasternal lymph nodes were involved. Following the automatic contouring, user-adjustments were made where necessary to make the contours clinically acceptable. Commonly used geometrical measures, such as the volumetric Dice Similarity Coefficient (vDSC), mean slice- wise Hausdorff distance (MSHD), and surface DSC (sDSC) were calculated between automatic and user- adjusted contours. For the LC patients the added path length (APL) was quantified for all OARs (except lungs), and from this measure a predicted time gain was estimated using a linear fit relation between the APL and time-for-adjustment. Results Geometrical results of user-adjustments are shown in Table 1. Better or comparable results were found in the LC group compared to BC group, except for the heart, where better results are found for the BC group (vDSC 0.83±0.13 vs. 0.86±0.15 (p=0.005), sDSC 0.37±0.22 vs. 0.56±0.22 (p<0.001), MSHD 1.58±0.99 vs. 0.92±0.70 (p<0.001), LC vs. BC respectively). Most likely the LC underperforms due to the lower image quality of the 4DCT compared to a breath-hold 3DCT in the BC group. The left and right lung were placed outside the body contour in the LC group for 37 and 19 of patients, respectively, and in the BC group for 1 and 2 patients, respectively; these patients were excluded from further geometrical analysis. For scan ranges that were extended compared to the training data, the automatic generated CTVp1 and contralateral breast in the BC group were placed in the hip for 9 and 4 patients, respectively. On average, a predicted time-saving of 11.3±6.3 min was achieved within the LC group (see Figure 1). A time gain of 2.3±1.5 min, 2.4±2.1 min, 0.4±1.2 min, and 6.1±3.3 min was predicted for the heart, spinal cord, esophagus, and mediastinum respectively.