Charbel Charlotte, Kwok Henry C, Miranda Joao, Zheng Junting, El Homsi Maria, El Amine Mohammad Ali, Chhabra Shalini, Danilova Sofia, Gangai Natalie, Petkovska Iva, Capanu Marinela, Vanguri Rami S, Chakraborty Jayasree, Horvat Natally
Department of Radiology, Oncologic Imaging Division, NYU Langone Health, New York, NY, USA.
Department of Surgery, Faculty of Medicine and Health Science, The University of Auckland | Te Waipapa Taumata Rau, Auckland, Aotearoa, New Zealand; Department of Radiology, Counties Manukau District, Health New Zealand, Aotearoa, New Zealand.
Eur J Radiol. 2025 Apr;185:112019. doi: 10.1016/j.ejrad.2025.112019. Epub 2025 Feb 26.
To assess the reliability of rectal MRI radiomic features across reader expertise level, image segmentation technique, and timing of rectal MRI.
This retrospective single-institutional study included consecutive patients with rectal adenocarcinoma who underwent total neoadjuvant therapy from January 2018 to June 2018. Baseline and restaging rectal MRI T2-weighted images were segmented independently by six radiologists (two fellows, two non-rectal radiologists, and two rectal radiologists). Four segmentation strategies were used and varied by image segmentation technique and timing of rectal MRI: (a) baseline volume of interest (VOI), (b) baseline region of interest (ROI), (c) restaging VOI, and (d) restaging ROI. Inter-reader agreement on each extracted radiomic feature was evaluated using the intra-class correlation coefficient (ICC).
Among 24 patients (16 men; median age, 56 years [interquartile range: 49-62]), 1,595 radiomic features were extracted. Baseline VOI segmentation achieved the highest inter-reader agreement rate, with 68 % (1,079/1,595) of radiomic features having an ICC > 0.7. Restaging ROI segmentation achieved the worst inter-reader agreement rate, with only 26 % (415/1,595) of radiomic features having an ICC > 0.7. First-order statistics and Gray Level Co-occurrence Matrix (GLCM) feature subgroups showed high inter-reader agreement rates, and the application of 'Square Root' and 'LOG Sigma' filters resulted in improved inter-reader agreement rates relative to original images. The expertise level of radiologists performing the segmentations did not affect the distribution of inter-agreement rates according to image segmentation technique or timing of rectal MRI.
Radiomic features were more reliable when extracted from baseline (vs. restaging) rectal MRIs and using 3D volume of interest (vs. 2D region of interest) segmentation, independent of the expertise level of the radiologists performing the segmentation.
Radiomic studies on rectal MRI employ various segmentation strategies and few assess their impact on reproducibility. Establishing the optimal segmentation method enhances radiomics model generalizability, potentially bridging the gap in clinical translation and improving clinical management of patients.
评估直肠MRI影像组学特征在不同阅片者专业水平、图像分割技术以及直肠MRI检查时间方面的可靠性。
这项回顾性单机构研究纳入了2018年1月至2018年6月期间接受全新辅助治疗的连续性直肠腺癌患者。6名放射科医生（2名住院医师、2名非直肠放射科医生和2名直肠放射科医生）分别独立分割基线和重新分期的直肠MRI T2加权图像。采用了四种分割策略，根据图像分割技术和直肠MRI检查时间的不同而有所变化：(a)基线感兴趣体积(VOI)，(b)基线感兴趣区域(ROI)，(c)重新分期VOI，以及(d)重新分期ROI。使用组内相关系数(ICC)评估阅片者之间对每个提取的影像组学特征的一致性。
在24例患者（16例男性；中位年龄56岁[四分位间距：49 - 62岁]）中，共提取了1595个影像组学特征。基线VOI分割实现了最高的阅片者间一致性率，68%（1079/1595）的影像组学特征ICC>0.7。重新分期ROI分割的阅片者间一致性率最差，只有26%（415/1595）的影像组学特征ICC>0.7。一阶统计量和灰度共生矩阵(GLCM)特征亚组显示出较高的阅片者间一致性率，相对于原始图像，应用“平方根”和“LOG西格玛”滤波器可提高阅片者间一致性率。进行分割的放射科医生的专业水平并未影响根据图像分割技术或直肠MRI检查时间得出的一致性率分布。
从基线（相对于重新分期）直肠MRI中提取影像组学特征并使用三维感兴趣体积（相对于二维感兴趣区域）分割时，影像组学特征更可靠，且与进行分割的放射科医生的专业水平无关。
关于直肠MRI的影像组学研究采用了各种分割策略，很少评估它们对可重复性的影响。建立最佳分割方法可提高影像组学模型的通用性，有可能弥合临床转化方面的差距并改善患者的临床管理。
