ESTRO 37 Abstract book

S1171

ESTRO 37

Results DIRLAB data-based extreme phase registration (inspiration to expiration phase) shows lowest TRE values for ANTS [(2.4 ± 1.3] mm], VarReg [(2.5 ± 1.3) mm] and Elastix [(2.6 ± 1.2) mm]. However, high(est) DIR accuracy not necessarily translated into high(est) correspondence modeling accuracy as ANTS registration yields highest mean TRE (3.0 mm) in model-based motion field estimation. Correspondence model formation and model- based 4D dose simulation for patient 4D CT data shows at least for lung metastases high accordance for the different DIR algorithm, irrespective of DIR accuracy differences. In contrast, results of 4D dose simulation for the investigated liver metastases are more diverging and vary for the different DIR approaches. However, large negative ΔD 95% values for 4 out of 5 algorithms were successfully connected to positive local recurrence for 2 liver metastases. Conclusion Especially the diverging results of 4D dose simulation in the liver raise doubts regarding reliability of DIR application in low contrast areas. Although, 4D dose accumulation for lung metastases shows promising results, current open source DIR frameworks should not be considered ready for "plug-and-play" use for 4D dose accumulation. EP-2123 Clinical evaluation of an auto-segmentation toolbox for breast CTV R. Simões 1 , R. Rozendaal 1 , J. Trinks 1 , R. Kalisvaart 1 , U. Van der Heide 1 , P. Remeijer 1 1 Netherlands Cancer Institute, Radiation Oncology, Amsterdam, The Netherlands Purpose or Objective Mirada Medical’s atlas-based auto-segmentation toolbox (Workflow Box™) is used to obtain breast CTV contours, which are manually corrected (if necessary) by the clinicians before being sent to the Treatment Planning System. In this work, our aim is to build an evaluation framework to assess the quality of the automatic contours with respect to the manual adjustments made on them, by analyzing both geometrical and dosimetric differences. Material and Methods 20 automatically generated contours of the left breast were available, as well as the manual adjustments made by the clinicians. The automatic contours were obtained with Mirada’s Workflow Box™, using 10 left breast atlases that had previously been selected and verified by a clinician. Within this toolbox, Deformable Image Registration is performed using an adaptation of Lucas- Kanade optic flow and the labels are fused by majority voting. Standard performance evaluation metrics were used to geometrically compare the contours: Dice coefficient, 95 th percentile Hausdorff distance (95%HD) and centroid distance (CD). A visual assessment of the errors was performed and displayed by means of projections onto a mean breast shape. The treatment plans were automatically generated using an in-house built framework (FAST), which has been clinically validated. Two treatment plans were generated per patient based on the auto-contour and the adjusted contour and the resulting dose distributions were compared. The dosimetric differences were quantified using indicators from γ analysis (3%/3mm): γ pass rate (γPR) and γ 95% percentile (γ95). Results The results for the geometric metrics are the following (mean ± standard deviation): Dice score 0.95±0.04, 95% Hausdorff distance (10.85±7.65)mm, centroid distance (4.50±4.35)mm. The differences are most often found in the most posterior and cranial parts of the breast (Fig. 1). We observe that larger breasts tend to be under- segmented and smaller breasts are often over-

segmented. Finally, we obtain the following values for the gamma indicators: γPR (95.25±4.86)%, γ95 (0.85±0.53). The γ95 indicator correlates well with the geometric metrics (Fig. 2), except for a few outliers.

Figure 1 - Average surface distances (in mm) projected onto the mean breast shape (cranial, posterior and frontal views).

Figure 2 - γ95 vs. the geometric metrics; R: Pearson correlation coefficient, p: p-value. Conclusion Mirada’s Workflow Box™ provides accurate CTV segmentations. The geometric differences occur in the same locations as the inter-observer variability reported in the literature. A visual analysis of the differences indicates that a larger variety of breast sizes should be included in the set of atlases used. The dosimetric impact of the manual adjustments is overall limited with most γ95 values lying below 1, indicating potential for an automated workflow that would require the manual revision of the contours in only a few cases. The proposed evaluation framework can be readily applied to other treatment sites. EP-2124 Time-saving evaluation of deep learning contouring of thoracic organs at risk T. Lustberg 1 , J. Van der Stoep 1 , D. Peressutti 2 , P. Aljabar 2 , W. Van Elmpt 1 , J. Van Soest 1 , M. Gooding 2 , A. Dekker 1 1 MAASTRO Clinic, Department of Radiation Oncology MAASTRO- GROW – School for Oncology and Developmental Biology- Maastricht University Medical Centre, Maastricht, The Netherlands 2 Mirada Medical Ltd, Science and Medical Technology, Oxford, United Kingdom Purpose or Objective Accurate contouring of organs-at-risk (OARs) in radiotherapy treatment planning is recommended for a successful outcome, but is often a time-consuming process. Auto-contouring methods have been proposed to improve consistency and reduce contouring time. This study reports on time savings generated by two automatic contouring methods: an atlas-based automatic segmentation (ABAS) (WorkflowBox 1.4, Mirada Medical Ltd, Oxford, UK) and a deep learning contouring (DLC) (WorkflowBox 2.0alpha, Mirada Medical Ltd, Oxford, UK). Their performance was evaluated on thoracic OARs for treatment of lung cancer patients. Material and Methods The automatic methods were evaluated on 20 stage I-III NSCLC patient CT images. The DLC system was trained on 450 lung cases collected at the same clinic. For the ABAS system 10 stage I NSCLC patients were carefully contoured according to the institutional guidelines and employed as atlases. All cases were acquired at a single institution. Evaluation was performed on the lungs, spinal cord, esophagus, heart and mediastinum envelope.