Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
[Preprint]. 2023 Oct 16:arXiv:2310.10867v1.

Evolving Horizons in Radiotherapy Auto-Contouring: Distilling Insights, Embracing Data-Centric Frameworks, and Moving Beyond Geometric Quantification

Kareem A Wahid  1   2 , Carlos E Cardenas  3 , Barbara Marquez  4   5 , Tucker J Netherton  5 , Benjamin H Kann  6 , Laurence E Court  5 , Renjie He  2 , Mohamed A Naser  2 , Amy C Moreno  2 , Clifton D Fuller  2 , David Fuentes  1
Affiliations

Evolving Horizons in Radiotherapy Auto-Contouring: Distilling Insights, Embracing Data-Centric Frameworks, and Moving Beyond Geometric Quantification

Kareem A Wahid et al. ArXiv. .

Update in

No abstract available

PubMed Disclaimer

Conflict of interest statement

Conflicts of Interest: CDF has received travel, speaker honoraria and/or registration fee waivers unrelated to this project from: The American Association for Physicists in Medicine; the University of Alabama-Birmingham; The American Society for Clinical Oncology; The Royal Australian and New Zealand College of Radiologists; The American Society for Radiation Oncology; The Radiological Society of North America; and The European Society for Radiation Oncology. The other authors have no interests to disclose.

Figures

Figure 1.
Figure 1.
A deep learning model trained with a few highly consistent, i.e., high-quality, contours (green) was more closely aligned to the ground truth test data than a model trained with many inconsistent contours (red) for various head and neck cancer radiotherapy structures. The 95% Hausdorff distance (HD95) (a) and mean distance to agreement (mDTA) (b) were used as geometric performance quantification metrics. Lower values for both metrics indicate better performance. Reprinted from Henderson et al..
Figure 2.
Figure 2.
Consensus from a limited number of non-expert contours can approximate expert benchmarks. Specific plot is shown for the left parotid gland in a head and neck cancer case using the volumetric Dice similarity coefficient (DSC) as a performance quantification metric. The Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm was used to generate consensus contours. To explore consensus quality dynamics based on the number of non-expert inputs, bootstrap resampling selected random non-expert subsets with replacement to form consensus contours, which were then compared to expert consensus. Each dot represents the median from 100 bootstrap iterations with a 95% confidence interval (shaded area). The black dotted line indicates the median expert DSC interobserver variability (IOV). The gray dotted line indicates DSC performance for the maximum number of non-experts used in the consensus. For this example, three to four non-experts can approximate expert IOV benchmarks. As the number of non-experts in the consensus contour increases, performance generally improves before plateauing. Adapted from Lin et al..
Figure 3.
Figure 3.
Relatively small training sample sizes are needed to reach high geometric performance for deep learning auto-contouring models. The percentage of the volumetric Dice similarity coefficient (DSC) using different training sample sizes relative to the maximum DSC for individual contour structures is shown in different colors. Most organ-at-risk structures required ~40 patient samples to achieve 95% of the maximum possible performance; notably, lenses and optic nerves required 200 samples to achieve 95% of the maximum possible performance. Reprinted from Fang et al..
Figure 4.
Figure 4.
HEad and neCK TumOR (HECKTOR) contouring performance saturation. Contouring performance measured by volumetric Dice similarity coefficient. Green and blue dots correspond to the top 10% and median tumor contouring performance measured across all participating teams, respectively. The gray dotted line corresponds to a clinician expert interobserver variability benchmark. Data derived from corresponding HECKTOR conference proceedings.
See this image and copyright information in PMC

References

    1. Cardenas C. E., Yang J., Anderson B. M., Court L. E. & Brock K. B. Advances in Auto-Segmentation. Semin. Radiat. Oncol. 29, 185–197 (2019). - PubMed
    1. Santoro M., Strolin S., Paolani G., Della Gala G., Bartoloni A., Giacometti C., Ammendolia I., Morganti A. G. & Strigari L. Recent Applications of Artificial Intelligence in Radiotherapy: Where We Are and Beyond. NATO Adv. Sci. Inst. Ser. E Appl. Sci. 12, 3223 (2022).
    1. Naqa I. E. in Artificial Intelligence in Radiation Oncology and Biomedical Physics 1–23 (CRC Press, 2023).
    1. Hamid O. H. From Model-Centric to Data-Centric AI: A Paradigm Shift or Rather a Complementary Approach? in 2022 8th International Conference on Information Technology Trends (ITT) 196–199 (2022).
    1. Mackay K., Bernstein D., Glocker B., Kamnitsas K. & Taylor A. A Review of the Metrics Used to Assess Auto-Contouring Systems in Radiotherapy. Clin. Oncol. 35, 354–369 (2023). - PubMed

Publication types

LinkOut - more resources