山东医药

目的 构建融合妇科查体信息与影像组学特征的子宫颈癌放疗靶区自动勾画模型，为子宫颈癌精准放疗提供技术支撑。方法 将2022年1月—2023年12月收治的288例子宫颈癌放疗患者，采用随机数表法按7∶2的比例划分为训练集(224例)和验证集(64例);将2024年1月—2024年12月收治的32例子宫颈癌放疗患者作为独立测试集。采集患者放疗前1周内妇科查体记录，经自然语言处理转化为结构化特征向量；同步采集CT及MRI(T_2WI/DWI)图像，由资深医师勾画放疗靶区作为专家参考标准。使用PyRadiomics软件提取126项影像组学特征，经LASSO回归筛选68项关键特征。以定位CT图像作为唯一输入，采用经典3D U-Net架构构建单一CT模型。以配准后T_2WI/DWI双序列及CT图像为3通道输入，沿用3D U-Net架构，仅调整输入通道适配多序列融合数据，构建单一MRI模型。以CT图像、MRI融合图像、7维妇科查体特征向量为输入，在改进U-Net架构基础上增加跨模态注意力特征融合模块，构建融合模型。使用训练集进行模型训练，使用验证集进行实时监控。使用独立测试集对模型性能进行评价，包括勾画准确性[Dice相似系数(DSC)、95%豪斯多夫距离(HD₉₅)、体积误差率]、勾画效率、勾画一致性(医师间Kappa系数)三个维度。结果 融合模型的DSC高于其余两种模型，HD₉₅和体积误差率均低于其余两种模型(P均<0.05)。10名放疗科医师纯手动勾画、融合模型辅助勾画平均耗时分别为(58.6±8.3)min、(9.2±1.5)min,两者相比，P<0.05。10名放疗科医师纯手动勾画、融合模型辅助勾画的Kappa系数分别为0.62(95%CI:0.55～0.69)、0.81(95%CI:0.75～0.87),两者相比，P<0.05。结论 成功构建融合妇科查体信息与影像组学特征的子宫颈癌放疗靶区自动勾画模型，可显著提升子宫颈癌放疗靶区勾画准确性与一致性，缩短耗时。

Objective To construct an automatic segmentation model for radiotherapy target volumes of cervical cancer by integrating gynecological examination information and radiomics features, so as to provide reliable technical support for precise radiotherapy of cervical cancer. Methods A total of 288 patients with cervical cancer who received radiotherapy were enrolled. They were divided into the training set(224 cases) and validation set(64 cases) at a ratio of 7∶2 using a random number table method. Another 32 patients with cervical cancer who received radiotherapy were selected as the independent test set. Gynecological examination records within 1 week before radiotherapy were collected and converted into structured feature vectors through natural language processing(NLP). Meanwhile, computed tomography(CT) and magnetic resonance imaging [MRI, including T₂-weighted imaging(T_2WI) and diffusion-weighted imaging(DWI)] images were collected, and the radiotherapy target volumes were delineated by senior radiation oncologists as the expert reference standard. A total of 126 radiomics features were extracted using PyRadiomics software, and 68 key features were screened out by least absolute shrinkage and selection operator(LASSO) regression. A single CT-based model was constructed using the classic 3D U-Net architecture with positioning CT images as the only input. A single MRI-based model was built with the same 3D U-Net architecture, using registered T_2WI/DWI dual sequences and CT images as 3-channel input, and only adjusting the input channels to adapt to multi-sequence fusion data. A fusion model was established on the basis of the improved U-Net architecture with the addition of a cross-modal attention feature fusion module, taking CT images, fused MRI images, and 7-dimensional gynecological examination feature vectors as inputs. The training set was used for model training, the validation set for real-time monitoring, and the independent test set for evaluating model performance from three dimensions: segmentation accuracy [including Dice similarity coefficient(DSC), 95% Hausdorff distance(HD₉₅), and volume error rate], delineation efficiency, and inter-observer consistency(inter-physician Kappa coefficient). Results The fusion model showed a significantly higher DSC, and significantly lower HD₉₅ and volume error rate than the other two single-modality models(all P<0.05). The average time consumed by 10 radiation oncologists for manual delineation and fusion model-assisted delineation was(58.6 ± 8.3) minutes and(9.2 ± 1.5) minutes, respectively, with a statistically significant difference(P<0.05). The Kappa coefficients of manual delineation and fusion model-assisted delineation by 10 radiation oncologists were 0.62(95% CI: 0.55-0.69) and 0.81(95% CI: 0.75-0.87), respectively, and the difference was statistically significant(P<0.05). Conclusions The automatic segmentation model for cervical cancer radiotherapy target volumes integrating gynecological examination information and radiomics features is successfully constructed. This model can significantly improve the accuracy and consistency of cervical cancer radiotherapy target volume delineation, and shorten the delineation time.