3305 - Physician Assessment and Clinical Evaluation of the Suitability of Deep Learning Based Auto Segmentation Contours for Organs at Risk Delineation in Radiation Oncology

S. M. Way¹, S. Varadhan², M. Khattab³, C. Gits¹, J. Borkenhagen⁴, and R. J. Hayward⁴; ¹Minneapolis Radiation Oncology PA, Minneapolis, MN, ²Edina High School, Edina, MN, ³Minneapolis Radiation Oncology, Edina, MN, ⁴Minneapolis Radiation Oncology, Minneapolis, MN

Purpose/Objective(s): Modern inverse radiation treatment planning requires accurate delineation of organs at risk (OAR) which can be an extremely time consuming process and subject to inter and intra observer variability. The use of deep learning based auto segmentation (DLAS) contours hold great promise in dramatically reducing the time involved in contouring and improve contour standardization. We report physician assessment and clinical applicability of DLAS contours from two commercial vendor platforms for three different patient anatomies. Materials/

Methods: Head and Neck, Prostate and Lung cancer patients that received intensity modulated image guided radiotherapy were randomly chosen for this study. DLAS contours for several OARs were done on both MIM Contour Protégé AI^R and Radformation AutoContour^R commercial platforms. The accuracy of DLAS contours were compared to the “ground truth” consensus OAR contours manually drawn and reviewed by three board certified radiation oncologists. We evaluated Dice Similarity Coefficient, Jaccard Index and Hausdorff distance between the ground truth and each of the DLAS contours from both commercial platforms.
Results: For the Head and Neck anatomy both the commercial platforms produced acceptable DLAS contours (Dice and Jaccard index > 0.8) with the exception of optic chiasm (DSC= 0.36 ± 0.15) and brachial plexus (DSC= 0.42 ± 0.12) contours which were clinically unacceptable. The DLAS RTOG Level I -V Lymph node delineation in head and neck although requiring minor edits when compared to physician volumes was found to be clinically acceptable and presented a huge savings in time in the clinical workflow. For prostate anatomy, AI generated OAR contours from both platforms were clinically acceptable with the exception of bowel (DSC = 0.74 ± 0.21) and sigmoid colon (DSC = 0.6 ± 0.23) across both the platforms. In thorax anatomy, both the commercial platforms provided clinically acceptable OAR contours with the exception of proximal bronchial tree. (DSC = 0.34 ± 0.08). The RTOG definition of proximal bronchial tree to include most inferior 2 cm of distal trachea and the proximal airways on both sides was not always followed in AI models that contributed to poor agreement with the physician reference contours.
Conclusion: To a great extent, DLAS can accurately delineate OARs for the Head and Neck, Prostate and Thorax anatomies studied, translating into enormous time savings in the RT process. These AI models are continuously refined by the vendors, highlighting the need for the physician “ground truth contours” to serve as an effective quality assurance check for verifying the accuracy of future DLAS contour models. Visual inspection is recommended before clinical use.