ESTRO 2020 Abstract Book

S185 ESTRO 2020

process will unavoidably propagate during the treatment stage and impact the clinical outcome. In this work we explore the case of the isolation forest (IF) algorithm as an ADS in the case of parotid gland delineations. Material and Methods A code was developed to predict anomalous parotid gland contours for a total of 1539 H & N patients treated at the NKI/AVL hospital since 2008. A total of 56 shape, image and distance metrics were extracted from each contour such as volume, symmetry, uniformity, entropy contrast etc. An isolation forest ADS was initially trained on 70 % of the total patient data with the addition of synthetically generated anomalies, while the remaining 30 % were left for validation. For the validation phase, 34 new anomalous structures were manually contoured that included isotropic and anisotropic expansions or contractions of 3 - 6 mm, missing or extra slices, extreme concave or prolonged duct regions. Finally previously treated clinical patients that presented higher than expected anomaly scores were evaluated and analyzed individually. Results The ADS system was able to depict more than 90 % of the anomalous structures manually contoured with a FPR at the level of 40 %. The ADS successfully identified as anomalous all uniform expansions and concave regions, as well as uni-directional expansions at the 6 mm level. From the clinical patients with the highest anomaly scores the contour anomaly was mostly associated with a malformed anatomy due to tumor intrusion, while in others due to the lack of imaging contrast. Interestingly, in several patient cases the contour area was expanded within the duct or muscle surrounding the parotid, a choice which was considered as debatable by the planner evaluating them. Conclusion Isolation forest ADS can assist in detecting outliers during the contouring phase and increase the consensus among radiotherapy planners. As we are moving into an automated era of the radiotherapy process, such systems may become an essential and powerful tool for the clinic. Future work will expand this study in other critical organs and tumor contours and enlarge the patient database. [1] F. T. Liu, K. M. Ting and Z. Zhou, "Isolation Forest," 2008 Eighth IEEE International Conference on Data Mining , Pisa, 2008, pp. 413-422

Effective Quality Assurance (QA) programmes are essential to the success of clinical trials. At present, contours delineated by clinicians in benchmark outlining cases are usually evaluated visually against a Gold Standard (GS). This can be resource intensive, time consuming and may delay trials opening. Conformity Index (CI) data for a variety of Target Volume (TV) and Organ At Risk (OAR) structures were reviewed to determine suitability for implementing Machine Learning (ML) models, aiming to automatically evaluate structures against the GS, based on respective CI data. CI data of 44 adrenal gland, 53 liver, 85 pelvic lymph node and 71 spine Stereotactic Ablative Body Radiotherapy benchmark cases, delineated by 132 clinicians from 25 centres were collected. The GS structures for each benchmark were approved independently by 2 senior oncology clinicians. Material and Methods Structures were imported into Velocity (v4.01 Varian Medical Systems) and compared the GS in axial, coronal and sagittal planes. This was performed by a blinded experienced oncology clinician, who scored each structure either pass or fail, dependent on conformity to the GS. The structures were then imported into CERR (Computational Environment for Radiotherapy Research) software to calculate conformity against the GS using the Jaccard Index, Dice Similarity Coefficient, van’t Riet Index, Hausdorff Distance, Geographical Miss Index and Disconcordance Index. The CI results and corresponding pass/fail score for each structure were imported into the Matlab Classification Learner Application in order for ML algorithms to decipher patterns between CI data (predictors) and the pass/fail score (response), building trained models that predict conformity of future structures based on their CI values. All preinstalled ML algorithms were applied to the CI data and trained models were produced. A 5-fold cross validation technique, which resamples the data to train and validate each model, was applied to determine the Predictive Accuracy (PA) of each model. The model with the highest PA was selected and exported as a script to implement in CERR. Results Figure 1 shows that different algorithms produced trained models with different PAs that vary by structure type. By separating all contours by structure type, trained models with higher PAs can be produced. Stomach produced models with the highest PA and liver GTV produced models with the lowest PA.

OC-0348 Automatic Evaluation of Contours Utilising Conformity Indices and Machine Learning S. Terparia 1 , R. Mir 2 , Y.M. Tsang 1 , R. Patel 2 , C. Clark 3 1 Mount Vernon Cancer Centre, Department of Radiotherapy, Northwood, United Kingdom ; 2 National Radiotherapy Trials Quality Assurance RTTQA Group, Mount Vernon Cancer Centre, Northwood, United Kingdom ; 3 Royal Surrey County Hospital, Department of Medical Physics, Guildford, United Kingdom

Purpose or Objective