Oops, you're using an old version of your browser so some of the features on this page may not be displaying properly.

MINIMAL Requirements: Google Chrome 24+Mozilla Firefox 20+Internet Explorer 11Opera 15–18Apple Safari 7SeaMonkey 2.15-2.23

ePoster Display

876P - Automatic segmentation of organs at risk in head and neck using 3D-UNets

Date

16 Sep 2021

Session

ePoster Display

Topics

Tumour Site

Head and Neck Cancers

Presenters

Swapneel Datta

Citation

Annals of Oncology (2021) 32 (suppl_5): S786-S817. 10.1016/annonc/annonc704

Authors

S. Datta1, A. Traverso2, S. Mehrkanoon1, A. Briassouli1

Author affiliations

  • 1 Data Science And Knowledge Engineering, Maastricht University, 6211LK - Maastricht/NL
  • 2 Oncology And Developmental Biology, Maastricht University Medical Center (MUMC), 6202 AZ - Maastricht/NL

Resources

Login to get immediate access to this content.

If you do not have an ESMO account, please create one for free.

Abstract 876P

Background

Manual segmentations of OARs is time consuming even for an expert radiologer. On average it takes an expert about 8 hours to manually segment all OARs. However, autosegmenting mechanisms using deep learning have proven to be able to segment organs in a fraction of the time with almost as good accuracy as an oncologist. Autosegmenting algorithms are much less prone to inter observer variability in that the interpretation of some organ boundaries differ per oncologist. Manual segmentation is also prone to missegmentations in some cases. Autosegmentation is the first step towards planning for radiation therapy and the clinical workflow can be improved using the incorporation of deep learning models during the clinical workflow.

Methods

Using the standard loss function (Sorensen-Dice coefficient) for training, larger organs are weighted more than smaller organs, leading to the larger organs being segmented more correctly. We propose a method that will lead to better performance of the model on small organs with poor surrounding contrast. In particular, 3D-Unets were used with the surface distance and SLF.

Results

It was observed that the Unet with the SLF preprocessing and hard region weighted loss performed the best, reaching around 77% validation dice scores. The Unet trained with Dice performed the second best, reaching around 70% validation dice scores. The Unet trained with generalized surface distance loss and SLF performed the worst (around 58%), but it was able to classify smaller organs like the lens better than the other models, reaching about 100% in some samples, which is not observed with the model trained on the Dice score.

Conclusions

The SLF model trained on the hard region weighted loss performed the best overall. Some benefits to using the surface distance metric was observed, as small organs like the lens were segmented better using the model trained on it despite being the worst performer overall. In the future, a combination of these loss functions with importance weights may be explored for a better overall model. This model may in practice be used in combination with consultation with an expert radiologer to correct the segmentations in clinical workflows, thereby reducing the time taken for the radiotherapy planning and ultimately a quicker treatment for the paitents.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.

This site uses cookies. Some of these cookies are essential, while others help us improve your experience by providing insights into how the site is being used.

For more detailed information on the cookies we use, please check our Privacy Policy.

Customise settings
  • Necessary cookies enable core functionality. The website cannot function properly without these cookies, and you can only disable them by changing your browser preferences.