基于人工智能的自动分割技术在盆腔区域高性能锥形束计算机断层扫描成像系统中的评估

Evaluation of artificial intelligence-based autosegmentation for a high-performance cone-beam computed tomography imaging system in the pelvic region.

作者信息

Sluijter Judith H, van de Schoot Agustinus J A J, Yaakoubi Abdelmounaim El, de Jong Maartje, van der Knaap-van Dongen Martine S, Kunnen Britt, Sijtsema Nienke D, Penninkhof Joan J, de Vries Kim C, Petit Steven F, Dirkx Maarten L P

机构信息

Department of Radiotherapy, Erasmus MC Cancer Institute, University Medical Center Rotterdam, Rotterdam, The Netherlands.

出版信息

Phys Imaging Radiat Oncol. 2024 Dec 9;33:100687. doi: 10.1016/j.phro.2024.100687. eCollection 2025 Jan.

DOI:10.1016/j.phro.2024.100687

PMID:39802649

原文链接:https://pmc.ncbi.nlm.nih.gov/articles/PMC11721864/

Abstract

BACKGROUND AND PURPOSE

A novel ring-gantry cone-beam computed tomography (CBCT) imaging system shows improved image quality compared to its conventional version, but its effect on autosegmentation is unknown. This study evaluates the impact of this high-performance CBCT on autosegmentation performance, inter-observer variability, contour correction times and delineation confidence, compared to the conventional CBCT.

MATERIALS AND METHODS

Twenty prostate cancer patients were enrolled in this prospective clinical study. Per patient, one pair of high-performance CBCT and conventional CBCT scans was included. Three observers manually corrected contours generated by the artificial intelligence (AI) model for prostate, seminal vesicles, bladder, rectum and bowel. Differences between AI-based and manual corrected contours were quantified using Dice Similarity Coefficient (DSC) and 95th percentile of Hausdorff distance (HD95). Autosegmentation performance and interobserver variation were compared using a random effects model; correction times and confidence scores using a paired -test and Wilcoxon signed-rank test, respectively.

RESULTS

Autosegmentation performance showed small, but statistically insignificant differences. Interobserver variability, assessed by the intraclass correlation coefficient, was significantly different across most organs, but these were considered clinically irrelevant (maximum difference = 0.08). Mean contour correction times were similar for both CBCT systems (11:03 versus 11:12 min; p = 0.66). Delineation confidence scores were significantly higher with the high-performance CBCT scans for prostate, seminal vesicles and rectum (4.5 versus 3.5, 4.3 versus 3.5, 4.8 versus 4.3; all p < 0.001).

CONCLUSION

The high-performance CBCT did not (clinically) improve autosegmentation performance, inter-observer variability or contour correction time compared to conventional CBCT. However, it clearly enhanced user confidence in organ delineation for prostate, seminal vesicles and rectum.

摘要

背景与目的

一种新型环形机架锥形束计算机断层扫描（CBCT）成像系统相较于传统版本，图像质量有所提高，但其对自动分割的影响尚不清楚。本研究评估了这种高性能CBCT与传统CBCT相比，对自动分割性能、观察者间变异性、轮廓校正时间和勾画置信度的影响。

材料与方法

20例前列腺癌患者纳入了这项前瞻性临床研究。每位患者均包括一对高性能CBCT和传统CBCT扫描。三名观察者手动校正由人工智能（AI）模型生成的前列腺、精囊、膀胱、直肠和肠的轮廓。基于AI的轮廓与手动校正轮廓之间的差异使用骰子相似系数（DSC）和豪斯多夫距离第95百分位数（HD95）进行量化。自动分割性能和观察者间变异使用随机效应模型进行比较；校正时间和置信度分数分别使用配对t检验和Wilcoxon符号秩检验。

结果

自动分割性能显示出微小但无统计学意义的差异。通过组内相关系数评估的观察者间变异性在大多数器官中存在显著差异，但这些差异被认为在临床上不相关（最大差异=0.08）。两种CBCT系统的平均轮廓校正时间相似（11:03对11:12分钟；p=0.66）。对于前列腺、精囊和直肠，高性能CBCT扫描的勾画置信度分数显著更高（4.5对3.5、4.3对3.5、4.8对4.3；所有p<0.001）。