Texture‐based, automatic contour validation for online adaptive replanning: A feasibility study on abdominal organs. Issue 9 (26th July 2019)

• Record Type: Journal Article
• Authors: Zhang, Ying; Plautz, Tia E.; Hao, Yao; Kinchen, Catherine; Li, X. Allen
• Is Part Of: Medical physics. Volume 46:Issue 9(2019)
• Page Start: 4010
• Page End: 4020
• Publication Date: 2019-07-26
• Subjects: contour‐quality validation -- decision tree model -- online adaptive replanning -- quantitative imaging -- texture analysis; Medical physics -- Periodicals; Medical physics; Geneeskunde; Natuurkunde; Toepassingen; Biophysics; Periodicals; Periodicals; Electronic journals; 610.153
• DOI: 10.1002/mp.13697
• Languages: English
• ISSNs: 0094-2405

Abstract

Abstract : Purpose: Evaluation of contour accuracy in radiation therapy planning requires manual interaction and is one of the most limiting bottlenecks for online replanning. This study aims to develop an automatic approach to rapidly evaluate contour quality based on image texture features to facilitate the routine practice of online adaptive replanning (OLAR). Method: Fifty‐five pancreas cancer patients were selected from a clinical database of patients treated at our institution from 2011 to 2018. For each patient, the pancreas head and duodenum were contoured in five images (one fraction per week) resulting in a total of 275 CT image sets with corresponding ground‐truth contours. A second set of inaccurate contours was generated using deformable‐image‐registration‐based contour propagation. Three subregions, core, inner shell and outer shell, were generated from the contour of each organ. Texture features were extracted from each subregion and descriptive features of each subregion were identified using the image set with corresponding ground‐truth contours. A three‐level decision tree model was constructed based on texture constraints empirically determined for the three subregions. The two datasets containing ground truth and inaccurate contours were merged. Randomized threefold cross‐validation was performed and repeated three times. Results: The first level of the decision tree utilizes textures derived from principal component analysis of a subset of extracted features from the core subregion (five PCs for pancreas head, seven PCs for duodenum). The second and third levels of the decision tree use gray‐level co‐occurrence matrix (GLCM)‐based cluster prominence to reject inaccurate contours. The trained model identifies accurate and inaccurate contours with an average sensitivity/specificity of 85%/91% for the pancreas head and 92%/92% for the duodenum contours. The false‐positive rate is 9% and 8% for pancreas head and duodenum, respectively. The execution time is less than 15 s using a standard desktop computer. Conclusion: Quantitative image features can be used to develop a model to rapidly validate the quality of an organ contour. Our model accurately classifies unseen contours as accurate or inaccurate with high sensitivity and specificity. As auto‐segmentation continues to improve in quality and accuracy, this method may be integrated into a fully automatic pipeline for auto‐segmentation, contour‐quality evaluation and contour correction, which would replace the time‐consuming manual review process, thereby facilitating the more routine practice of OLAR.