文章：

确保可重复性与部署模型：Image2Radiomics框架下图像处理对胰腺神经内分泌肿瘤模型性能影响的评估

Ensuring Reproducibility and Deploying Models with the Image2Radiomics Framework: An Evaluation of Image Processing on PanNET Model Performance

原文发布日期：1 August 2025

DOI: 10.3390/cancers17152552

类型: Article

开放获取: 是

英文摘要：

Background/Objectives: To evaluate the importance of image processing in a previously validated model for detecting pancreatic neuroendocrine tumors (PanNETs) and to introduce Image2Radiomics, a new framework that ensures reproducibility of the image processing pipeline and facilitates the deployment of radiomics models. Methods: A previously validated model for identifying PanNETs from CT images served as the reference. Radiomics features were re-extracted using Image2Radiomics and compared to those from the original model using performance metrics. The impact of nine alterations to the image processing pipeline was evaluated. Prediction discrepancies were quantified using the mean ± SD of absolute differences in PanNET probability and the percentage of classification disagreement. Results: The reference model was successfully replicated with Image2Radiomics, achieving a Cohen’s kappa coefficient of 1. Alterations to the image processing pipeline led to reductions in model performance, with AUC dropping from 0.87 to 0.71 when image windowing was removed. Prediction disagreements were observed in up to 45% of patients. Even minor changes, such as switching the library used for spatial resampling, resulted in up to 21% disagreement. Conclusions: Reproducing image processing pipelines remains challenging and limits the clinical deployment of radiomics models. While this study is limited to one model and imaging modality, the findings underscore a common risk in radiomics reproducibility. The Image2Radiomics framework addresses this issue by allowing researchers to define and share complete processing pipelines in a standardized way, improving reproducibility and facilitating model deployment in clinical and multicenter settings.

摘要翻译： 

背景/目的：评估图像处理在先前已验证的胰腺神经内分泌肿瘤（PanNETs）检测模型中的重要性，并介绍Image2Radiomics这一新框架，该框架可确保图像处理流程的可重复性，并促进放射组学模型的部署。方法：以先前验证的CT图像识别PanNETs模型为参照标准。使用Image2Radiomics重新提取放射组学特征，并通过性能指标与原模型特征进行比较。评估了图像处理流程中九项调整的影响。使用PanNET概率绝对差值的均值±标准差及分类不一致百分比量化预测差异。结果：使用Image2Radiomics成功复现了参照模型，Cohen's kappa系数达到1。图像处理流程的调整导致模型性能下降，当去除图像窗宽窗位调整时，AUC从0.87降至0.71。在高达45%的患者中出现预测不一致。即使是细微调整（如更换空间重采样库）也可导致高达21%的不一致。结论：图像处理流程的复现仍具挑战性，限制了放射组学模型的临床部署。尽管本研究仅针对单一模型和成像模态，但结果揭示了放射组学可重复性面临的普遍风险。Image2Radiomics框架通过允许研究者以标准化方式定义和共享完整处理流程，有效解决了这一问题，提升了可重复性，并促进了临床及多中心场景中的模型部署。

原文链接：

Ensuring Reproducibility and Deploying Models with the Image2Radiomics Framework: An Evaluation of Image Processing on PanNET Model Performance