EUVIMEDEUVIMED entdecken
ValidierungsstudieOpen Access · CC BY 4.0

Cross‐anatomical evaluation of a deep‐learning auto‐contouring system: qualitative, geometric, and dosimetric validation

Sara Endo, Masaki Nakamura, Takeshi Fujisawa, Hidehiro Hojo, Hidenobu Tachibana

Journal of Applied Clinical Medical Physics · 2026

DOI: 10.1002/acm2.70662 ↗

Kurz gefasst

Validierungsstudie in der Zeitschrift Journal of Applied Clinical Medical Physics, 2026. Verfasst von Sara Endo und 4 weiteren Autor:innen. Beteiligte Einrichtung: Radiation Safety and Quality Assurance Division, National Cancer Center Hospital East, Chiba, Japan. Hauptthemen laut MEDLINE: Radiotherapy Planning, Computer-Assisted, Deep Learning, Organs at Risk, Neoplasms. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 12 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: JSPS KAKENHI.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

Accurate and consistent delineation of target and normal structures is essential for safe and effective radiotherapy. Compared with manual or atlas‐based methods, deep learning‐based auto‐contouring has demonstrated improved efficiency and reduced interobserver variability. However, its performance can vary across anatomical regions and clinical contexts, and comprehensive multi‐site validation remains limited. Existing studies have tended to assess only geometric similarity or a single region, providing insufficient insight into real‐world usability and dosimetric influence. A systematic evaluation that integrates physician‐based quality assessment, quantitative geometric metrics, and plan‐level dosimetry across diverse anatomical sites is needed to support broader clinical implementation of deep learning‐based auto‐contouring for differentiation between target and normal structures.

Purpose

To perform the first practice‐oriented, cross‐anatomical evaluation of the SYNAPSE Radiotherapy system (v1.5) and to assess its qualitative usability, geometric agreement (Dice similarity coefficient [DSC], mean distance to agreement [MDA], and maximum Hausdorff distance [HD]), and plan‐level dosimetric impact across six anatomical sites.

Methods

We retrospectively analyzed data for 120 patients across six sites, namely, the head, head and neck, thorax, abdomen, male pelvis, and female pelvis. Thirty organs at risk were auto‐segmented using the SYNAPSE Radiotherapy platform, and three board‐certified radiation oncologists graded the clinical usability of the platform on a 4‐point scale. Geometric agreement was quantified using the DSC, MDA, and maximum HD. Plan‐level dosimetric comparisons were performed for representative targets. Institutional review board approval was obtained (number 2017‐440).

Results

Overall, 76.1% of auto‐contours were qualitatively judged to be clinically usable (score ≥3). Geometric thresholds were achieved for most structures (DSC > 0.8 in 54.8%, MDA < 2.5 mm in 71.5%, and maximum HD < 12.5 mm in 50%), with only 8/29 being major outliers. Auto‐contours tended to be larger than manual contours in 72% of cases. Dosimetrically, 88.1% of organs showed dose‐volume differences within ±5%, and dose constraint achievement rates were comparable between auto‐contours and manual contours (93.9% vs. 95.%). Median doses to planning target volumes (e.g., prostate and iliac nodes) were also consistent, indicating minimal clinical impact.

Conclusions

The SYNAPSE Radiotherapy system showed high overall clinical usability with organ‐dependent variability. Our findings expand the platform‐specific evidence for the SYNAPSE Radiotherapy system by linking geometry to dosimetric impact and supporting clinical deployment with an organ‐specific quality assurance policy, while highlighting contexts in which targeted expert review remains necessary for safe clinical translation.

Schlagwörter

MeSH-Schlagwörter (MEDLINE)

Radiotherapy Planning, Computer-AssistedDeep LearningOrgans at RiskNeoplasmsImage Processing, Computer-AssistedHumansRadiotherapy DosageRetrospective StudiesRadiotherapy, Intensity-ModulatedMaleFemaleRadiometry

Stichwörter der Autor:innen

deep learning‐based auto‐contouringdosimetric impactgeometric accuracyorgans at risk (OARs)radiotherapy planning

Autor:innen und Institutionen

  • Radiation Safety and Quality Assurance Division, National Cancer Center Hospital East, Chiba, Japan.
  • Department of Radiation Oncology, National Cancer Center Hospital East, Chiba, Japan.

Förderung

  • JSPS KAKENHI

Zitieren

APA

Endo, S., Nakamura, M., Fujisawa, T., Hojo, H., & Tachibana, H. (2026). Cross‐anatomical evaluation of a deep‐learning auto‐contouring system: qualitative, geometric, and dosimetric validation. Journal of Applied Clinical Medical Physics. https://doi.org/10.1002/acm2.70662

Vancouver

Endo S, Nakamura M, Fujisawa T, Hojo H, Tachibana H. Cross‐anatomical evaluation of a deep‐learning auto‐contouring system: qualitative, geometric, and dosimetric validation. Journal of Applied Clinical Medical Physics. 2026. doi:10.1002/acm2.70662

BibTeX
@article{Endo2026,
  author = {Endo, Sara and Nakamura, Masaki and Fujisawa, Takeshi and Hojo, Hidehiro and Tachibana, Hidenobu},
  title = {Cross‐anatomical evaluation of a deep‐learning auto‐contouring system: qualitative, geometric, and dosimetric validation},
  journal = {Journal of Applied Clinical Medical Physics},
  year = {2026},
  doi = {10.1002/acm2.70662}
}
Kostenlos starten →