MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Deep Learning Based Auto-Segmentation Method for Radiation Therapy of Head and Neck Cancer

A Amjad1*, Z Chen2 , M Awan1 , M Shukla1 , C Yang2 , Q Zhou2 , X Li1 , (1) Medical College of Wisconsin, Milwaukee, WI, (2) Manteia Medical Technologies, Milwaukee, WI

Presentations

(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 5

Purpose: In current practice, segmentation for radiation therapy (RT) planning of head and neck cancer (HNC) is generally time consuming and labor intensive. This study investigates the potential of a Deep Learning (DL) based automatic segmentation model for RT of HNC.

Methods: A research tool of DL based 3D auto-segmentation model employing a U-NET architecture, combined with multiple neural net design approaches including Resnet, DenseNet and Se-Block, was evaluated using the CT data sets from randomly selected 36 HNC patients. The algorithms considered the imbalance of data availability among different organs in loss design and used traditional image processing methods such as active contours to optimize both the neural network and the final contours. For each testing CT set, common structures including brain, brain stem, esophagus, eyes, mandible, parotid, and trachea were delineated manually by experienced radiation oncologists. The clinical practicality of the contours generated by the auto-segmentation model was evaluated by quantitatively comparing with the manual contours based on AAPM TG-132 recommendations; Dice Similarity Coefficient (DSC) > 0.8 and Mean Distance to Agreement (MDA) < 3 mm.

Results: A detailed organ-based comparative analysis revealed that for the 9 structures, contours created by the DL-based auto-segmentation algorithm were comparable to manual contours and were clinically acceptable under the TG-132 recommendations. The average DSC of 0.99, 0.85, 0.84, 0.91, 0.91, 0.89, 0.83, 0.83 and 0.87 was found for brain, brain stem, esophagus, left and right eyes, mandible, left and right parotid, and trachea respectively. A qualitative slice-by-slice analysis acknowledged that the model’s accuracy could be compromised if image artifact was present or if the quality of the training data was imperfect.

Conclusion: The newly developed DL-based auto-segmentation models, if trained with good quality data, can accurately deliver clinically acceptable contours on artifact-free CT for RT planning of HNC.

Funding Support, Disclosures, and Conflict of Interest: The work is funded by MCW Fotsch foundation and partially by Manteia Medical Technologies. ZC, CY and QZ are employees of Manteia.

Keywords

Not Applicable / None Entered.

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email