Click here to

Session: Imaging: AI in Imaging [Return to Session]

Autosegmentation of Videofluoroscopic Swallow Studies Using U-Net for Improved Analysis of Radiation Induced Dysphagia in Head and Neck Cancer Patients

N Shaheen1*, R Pena-Chavez2, C Ulmschneider3, R Burdick2,4, J Yee4,5, A Kurosu6, N Rogus-Pulia2,4,5, B Bednarz1, (1)Department of Medical Physics, University of Wisconsin - Madison(2) Department of Communication Sciences and Disorders, University of Wisconsin - Madison(3) Department of Surgery, University of Wisconsin - Madison(4) Department of Medicine, University of Wisconsin - Madison(5) Geriatric Research Education and Clinical Center, William S. Middleton Memorial Veterans Hospital Madison, Wi, USA


SU-IePD-TRACK 1-6 (Sunday, 7/25/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]

Purpose: Radiation induced dysphagia is a common complication from head and neck cancer (HNC) treatment. Dysphagic patients will experience malnutrition and a large increased risk for aspiration pneumonia, leading to future hospitalizations. Current dysphagia diagnosis and treatment requires extensive manual analysis of a videofluoroscopic swallow study (VFS). Current methods are time consuming and solely qualitative. This work develops an auto segmentation algorithm for assisted analysis of VFS and further implementation of automated quantitative measures.

Methods: 40 VFS were collected from a varying severity dysphagic cohort. Shielded edges of the images and non-swallowing related motion were cropped. Images were standardized and resized to 256 x 256 pixels. Hyoid, bolus, and C2 – C4 vertebrae contours were done on unprocessed images and underwent the same cropping and resizing. A 2D modified U – Net structure was used. The network was trained using 80% of available image slices (1386,4490, 2881) for the hyoid, bolus, and C2-C4 separately. A non-decaying learning rate and Adam optimizer in TensorFlow was used.

Results: Dice coefficients on the remaining validation images were 0.87, 0.92, and 0.95 for the hyoid, bolus, and C2-C4, respectively.

Conclusion: Resulting dice coefficients reflected accurate contours. The C2-C4 contours were most similar given the lack of motion and contrast in that region. With the Hyoid and C2-C4, edge pixels were most uniformly missed around the contour. The bulk of bolus contour was accurate, but small portions (residue) were missed in some cases. Identification of residue is debated with manual analysis leading to required further work for accurate bolus contouring.

Funding Support, Disclosures, and Conflict of Interest: UW - Madison Fall Research Competition



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email