Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Clinical Implementation of AI Generated Contouring - Assessment of Accuracy of AI Generated Organ Contours Using Two Commercial Systems

M Zaman*, C Kota, Yale New Haven Health, New Haven, CT


PO-GePV-M-210 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To evaluate Artificial Intelligence generated contours for clinical implementation by comparing AI generated contours with clinically utilized contours and a head to head comparison of AI engines from multiple vendors to determine their clinical usefulness.

Methods: A total of 25 recently treated patients, five each from site specific to H&N, Breast, Lung, Abdomen and Prostate were compared with AI generated contours. Two AI engines from MIM-Vista and Radformation were used to generate contours. AI-generated contours were evaluated qualitatively comparing slice-by-slice to the clinical contours and quantitatively in terms of Mean Distance-to-Agreement (MDA) and Dice Similarity Coefficient (DSC). Contours were assigned scores based on their qualitative and quantitative agreement with clinical contours.

Results: Both MIM-Vista and Radformation performances were similar within statistical uncertainty. Qualitative evaluation reveals that most of the structures are clinically acceptable. For H&N, both performed well for most of the structures with median MDA<2mm and DSC>0.8 for each individual structures except for smaller structures such as brachial plexuses and optic pathway. For Breast, both engines performed very well, Median MDA<1.5mm and DSC>0.9 for each individual structures. For lung, both of the AI engines performed equally well, Median MDA<2mm and DSC>0.9 for most structures although esophagus and cord median DSC were between 0.8 and 0.9. For abdominal plans, Radformation performed better than MiMVista, especially on kidneys. For prostate, except for seminal vesicles and penile bulb, both performed equally well with median MDA<2mm and DSC>0.85 for each structures.

Conclusion: This study finds that AI engines generate clinically useful contours. Both MIM-Vista and Radformation performed equally well although MIM-Vista performs slightly poorly in the presence of contrast. Radformation generates more structures than MIMVista. The clinical implementation of the AI contouring tools will have profound impacts in Radiation Oncology practices. This will improve treatment planning quality by standardizing contour delineation process whereby the clinicians can dedicate more time on reviewing and fine-tuning, if necessary, the contours. AI can generate more structures than are typically delineated by the clinicians routinely and can be of assistance for dose tuning and toxicity assessment.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email