Department of Radiology, Obstetrics and Gynecology Hospital, Fudan University, P R China.
Department of Interventional Radiology, DaLian Women and Children's Medical Group, P R China.
Acad Radiol. 2024 Dec;31(12):4985-4995. doi: 10.1016/j.acra.2024.05.037. Epub 2024 Jun 6.
To establish and validate a predictive multi-machine learning model for the long-term efficacy of uterine artery embolization (UAE) in the treatment of adenomyosis based on habitat subregions.
Patients who underwent UAE for adenomyosis at institution A between November 2015 and June 2018 were included in the training cohort and those at institution B between June 2017 and June 2019 were included in the test cohort. The regions of interest (ROI) were manually segmented on the T2-weighted images (T2WI). The ROIs were subsequently partitioned into habitat subregions using k-means clustering. Radiomic features were extracted from each subregion on T1WI, T2WI, apparent diffusion coefficient, and contrast-enhanced images. The least absolute shrinkage and selection operator (LASSO) was used to select the subregion radiomics features. With the improvement in patients' symptoms at 36 months post-UAE, the habitat subregion features were trained using six machine-learning classifiers. The most suitable classifier was chosen based on model performance to establish the habitat radiomics model (HRM). The efficacy of the model was validated using both the training and test cohorts. Finally, a whole-region radiomics model (WRM) and clinical model (CM) were established. The Delong test was used to compare the predictive performance of the habitat subregion model and the two other models.
The study included 258 patients, 191 in the training cohort and 67 in the test cohort. The ROIs were divided into four habitat subregions. Radiomics features were extracted from different sequence images of the subregions. After LASSO regression, 24 habitat subregion features were included in the model. Based on the receiver operating characteristic curve analysis, the area under the curve (AUC) of the HRM was 0.921 (95% CI, 0.857-0.985, training) and 0.890 (95% CI, 0.736-1.000, test). The AUCs for the WRM were 0.805 (95% CI, 0.737-0.872, training) and 0.693 (95% CI, 0.497-0.889, test). Compared to the HRM, the difference in predictive performance was statistically significant (p = 0.008, training; p = 0.007, test). The AUCs for the CM were 0.788 (95% CI, 0.711-0.866, training) and 0.735 (95% CI, 0.566-0.903, test). Compared to the HRM, there was a statistically significant difference in the training cohort (p = 0.014) but not in the test cohort (p = 0.186).
The HRM can predict the long-term efficacy of UAE in the treatment of adenomyosis. The predictive performance was superior to that of both the WRM and CM, serving as an effective tool to assist interventional physicians in clinical decision-making.
基于生境亚区，建立并验证一种预测子宫动脉栓塞术（UAE）治疗子宫腺肌病长期疗效的多机器学习模型。
将 2015 年 11 月至 2018 年 6 月在机构 A 接受 UAE 治疗的腺肌病患者纳入训练队列，将 2017 年 6 月至 2019 年 6 月在机构 B 接受 UAE 治疗的患者纳入测试队列。在 T2 加权图像（T2WI）上手动对感兴趣区域（ROI）进行分段。随后使用 K 均值聚类法将 ROI 划分为生境亚区。从 T1WI、T2WI、表观扩散系数和对比增强图像的每个亚区提取放射组学特征。使用最小绝对收缩和选择算子（LASSO）选择亚区放射组学特征。在 UAE 治疗后 36 个月，根据患者症状改善情况，使用六种机器学习分类器对生境亚区特征进行训练。根据模型性能选择最合适的分类器，建立生境放射组学模型（HRM）。使用训练和测试队列验证模型的有效性。最后，建立全区域放射组学模型（WRM）和临床模型（CM）。采用 Delong 检验比较生境亚区模型与另外两种模型的预测性能。
共纳入 258 例患者，其中训练队列 191 例，测试队列 67 例。ROI 被分为四个生境亚区。从不同序列图像的亚区提取放射组学特征。经过 LASSO 回归，24 个生境亚区特征被纳入模型。基于受试者工作特征曲线分析，HRM 在训练队列中的曲线下面积（AUC）为 0.921（95%置信区间，0.857-0.985），在测试队列中的 AUC 为 0.890（95%置信区间，0.736-1.000）。WRM 在训练队列中的 AUC 为 0.805（95%置信区间，0.737-0.872），在测试队列中的 AUC 为 0.693（95%置信区间，0.497-0.889）。与 HRM 相比，预测性能的差异具有统计学意义（p=0.008，训练；p=0.007，测试）。CM 在训练队列中的 AUC 为 0.788（95%置信区间，0.711-0.866），在测试队列中的 AUC 为 0.735（95%置信区间，0.566-0.903）。与 HRM 相比，在训练队列中存在统计学差异（p=0.014），但在测试队列中无统计学差异（p=0.186）。
HRM 可预测 UAE 治疗子宫腺肌病的长期疗效。预测性能优于 WRM 和 CM，可作为辅助介入医师临床决策的有效工具。
