Click here to

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

Performance Evaluation of AI-Based Automatic Segmentation Modules for Head and Neck Cancer Patients

Y Liao*, R Injerd, G Tolekidis, N Joshi, J Turian, Rush University Medical Center, Chicago, IL


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

ePoster Forums

Purpose: Artificial Intelligence (AI)-based auto segmentation is an emerging technology in Radiation Oncology. Radiotherapy for head-and-neck cancer requires accurate contouring of numerous organs at risk (OARs), which can be a time-consuming task. To improve the efficiency of this task, we evaluated six commercially available AI contouring modules to assess and rank their performance before a clinical implementation decision is to be finalized.

Methods: We retrospectively selected 12 head-and-neck cases and input the CT images through the six chosen modules: Carina AI, Limbus, Manteia AccuContour, MIM ProtégéAI, Mirada DLCExpert, and Radformation AutoContour. Agreement between the AI-generated contours and the clinical contours drawn by the physician was evaluated in terms of mean distance to agreement (MDA) and Dice similarity coefficient (DSC). In addition, contouring time, organ list, system requirements, and trainability were also compared. A qualitative assessment of the AI contours by 2 dosimetrists and 2 physicists using a 4-point scoring scale was also performed.

Results: Overall, AI generated contours were accurate for high contrast and relatively large organs, such as: mandible, brain, parotids, eyes, and submandibular glands with MDA ~1mm, and DSC ~0.8-0.9 from all modules. However, for low contrast regions, more complex and/or smaller structures, such as: brachial plexus and chiasm, DSC decreases to <0.5. The score varies among modules for long structures like spinal cord and esophagus due to the ill-defined superior/inferior border. The AI segmentation time ranged from 1 to 7 minutes. AI modules generated 21-50 structures per patient, out of which 21-28 overlapped with the physician’s 32-structure list. Our qualitative review of the AI segmentations revealed that Limbus module was the top performer.

Conclusion: Commercially available AI segmentation modules can provide an efficient alternative for generating reliable OARs for head and neck treatment planning purposes. Clinical implementation will need a more comprehensive evaluation.




IM/TH- image Segmentation: CT

Contact Email