Click here to


Are you sure ?

Yes, do it No, cancel

Evaluation of An Artificial Intelligence (AI) Based Auto Contouring Workflow Box for Head and Neck Radiotherapy

L Zhuang1*, M Pankuch2, E Bowers1, M Posner1, (1) Northwestern Medicine Lake Forest Hospital, Lake Forest, IL, (2) Northwestern Medicine Chicago Proton Center, Warrenville, IL


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: evaluate the quality and efficiency of the Head and Neck (H&N) model for an AI based auto contouring tool.

Methods: total of 22 patients treated in 2019 at a local community hospital were included in this study. Planning CT images were sent directly to the AI based contouring tool, Mirada workflow box (MWB, Mirada Medical Ltd., Oxford, OX), to create auto contours which were automatically transferred to the treatment planning system. Organ at risks (OARs) evaluated in this study include brainstem, mandible, left parotid, right parotid, and spinal cord. These OARs were also contoured manually, either from scratch or through modification of the MWB contour by experienced dosimetrists. 3D Dice Similarity coefficients (DSC) were calculated for each organ between the manual and auto contours. Auto contour time efficiencies were also evaluated.

Results: DSC is 0.73±0.09 for brainstem, 0.86±0.09 for mandible, 0.75±0.13 for Left parotid, 0.69±0.14 for Right parotid and 0.61±0.11 for spinal cord. The mandible achieved the highest DSC, likely due to that it appears high contrast on CT. Spinal cord achieved the lowest DSC mainly because of contouring range (slice) differences between MWB and the planning expert. The H&N model of MWB is AI based which was built upon data from 698 H&N patients and can generate up to 22 structures for a typical H&N patient within 5 minutes. For mandible and spinal cord OARs, contour modification can usually be performed within 3 minutes, while it may take 5-10 minutes on average for parotid and brainstem contour modification.

Conclusion: H&N model of Mirada Workflow Box archives good agreement with expert contours for OARs, especially structures with high image contrast such as the mandible. Future work includes revising the model by including institutional data and evaluating both quality and contouring efficiency improvements.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email