Online

ESTRO 2020

Session Item

Poster highlights 16 RTT: Novel strategies for treatment planning and delivery
8200
Poster Highlights
RTT
17:09 - 17:17
Clinical evaluation of a deep network organ segmentation algorithm for radiation treatment planning
Sebastian Norbert Marschner, Germany
PH-0485

Abstract

Clinical evaluation of a deep network organ segmentation algorithm for radiation treatment planning
Authors: Claus Belka.(Department of Radiation Oncology, University Hospital LMU Munich, Munich, Germany), Guillaume Chabin.(Digital Technology & Innovation- Whole Body & Oncology USA, Siemens Healthineers, Princeton NJ, USA), Stefanie Corradini.(Department of Radiation Oncology, University Hospital LMU Munich, Munich, Germany), Manasi Datar.(Digital Technology & Innovation- Artificial Intelligence Germany, Siemens Healthineers, Erlangen, Germany), Aurélie Gaasch.(Department of Radiation Oncology, University Hospital LMU Munich, Munich, Germany), Bernhard Geiger.(Digital Technology & Innovation- Whole Body & Oncology USA, Siemens Healthineers, Princeton NJ, USA), Sasa Grbic.(Digital Technology & Innovation- Whole Body & Oncology USA, Siemens Healthineers, Princeton NJ, USA), Tobias Heimann.(Digital Technology & Innovation- Artificial Intelligence Germany, Siemens Healthineers, Erlangen, Germany), Sebastian Marschner.(Department of Radiation Oncology, University Hospital LMU Munich, Munich, Germany), Christian Moehler.(Advanced Therapies- Cancer Therapy, Siemens Healthineers, Erlangen, Germany), Julian Rosenman.(Department of Radiation Oncology, University of North Carolina at Chapel Hill, Chapel Hill NC, USA), Christian Thieke.(Department of Radiation Oncology, University Hospital LMU Munich, Munich, Germany), Fernando Vega.(Advanced Therapies- Cancer Therapy, Siemens Healthineers, Erlangen, Germany), Zhoubing Xu.(Digital Technology & Innovation- Whole Body & Oncology USA, Siemens Healthineers, Princeton NJ, USA)
Show Affiliations
Purpose or Objective

Auto contouring algorithms in radiotherapy aim at reducing the time needed for organ segmentation and to reduce inter-observer variations. In this work, we quantitatively evaluated a deep neural network algorithm which automatically contours lungs, heart, bladder and rectum as organs-at-risk on computed tomography (CT) for radiation treatment planning. 

         
Material and Methods

Thoracic CT images of 237 patients (for left/right lungs and the heart) and pelvic CT images of 102 patients (for bladder and rectum) were manually contoured by an expert. On the same CT images, a deep image-to-image network (DI2IN) algorithm, which was trained on a dataset acquired from other hospitals, generated the contours automatically without any user interaction. The manual and automatic contours were quantitatively compared using median Dice Similarity Coefficient (mDSC) and mean surface distance (MSD). The contours were also compared visually, and the areas with the biggest discrepancies between manual and automatic delineation were identified.

Results

In general, we observed high correlation between automatic and manual contours. The best results were obtained for the left/right lungs with a mDSC of 0.98±0.03/0.98±0.02 (standard deviation) and MSD of 1,7±6,9mm/1,3±2,9mm, and the bladder with a mDSC of 0.94±0.08 and MSD of 1,55±1,77mm. Slightly lower correlations were observed for the heart (mDSC of 0.91±0.02, MSD of 2.0±0.8mm) and the rectum (mDSC 0.86±0.09, MSD 2.16±1,3mm). Comparison of boundary values and visual inspection showed that for the heart automatic and manual contours mostly differed at the cranial boundary (with average deviation of 13.5mm), and for the rectum at both the cranial and the caudal boundary, with average deviations of 8.5 mm and 6.75mm, respectively. One possible reason for the deviations could be the fact that the training of the algorithm was based on contours of other hospitals which might have used different contouring guidelines.

Conclusion

The DI2IN algorithm automatically generated contours for organs at risk in the thorax and pelvis region which were very similar to the manual contours, making the contouring step in radiation treatment planning simpler and faster. Further improvements are expected when the algorithm will be trained or fine-tuned with a dataset generated in the same hospital to mitigate or eliminate deviations stemming from different contouring guidelines.