Li Xuanyi, Shi Weijun, Zhang Qianwen, Lin Xinhui, Sun Hongzan
Department of Radiology, Shengjing Hospital of China Medical University, Sanhao Street No. 36, Heping District, Shenyang, 110004, China.
The Second Clinical College of China Medical University, Shenyang, 110004, China.
BMC Cancer. 2025 Jun 5;25(1):1006. doi: 10.1186/s12885-025-14392-6.
To create an automated PET/CT segmentation method and radiomics model to forecast Mismatch repair (MMR) and TP53 gene expression in endometrial cancer patients, and to examine the effect of gene expression variability on image texture features.
We generated two datasets in this retrospective and exploratory study. The first, with 123 histopathologically confirmed patient cases, was used to develop an endometrial cancer segmentation model. The second dataset, including 249 patients for MMR and 179 for TP53 mutation prediction, was derived from PET/CT exams and immunohistochemical analysis. A PET-based Attention-U Net network was used for segmentation, followed by region-growing with co-registered PET and CT images. Feature models were constructed using PET, CT, and combined data, with model selection based on performance comparison.
Our segmentation model achieved 99.99% training accuracy and a dice coefficient of 97.35%, with validation accuracy at 99.93% and a dice coefficient of 84.81%. The combined PET + CT model demonstrated superior predictive power for both genes, with AUCs of 0.8146 and 0.8102 for MMR, and 0.8833 and 0.8150 for TP53 in training and test sets, respectively. MMR-related protein heterogeneity and TP53 expression differences were predominantly seen in PET images.
An efficient deep learning algorithm for endometrial cancer segmentation has been established, highlighting the enhanced predictive power of integrated PET and CT radiomics for MMR and TP53 expression. The study underscores the distinct influences of MMR and TP53 gene expression on tumor characteristics.
创建一种自动化的PET/CT分割方法和放射组学模型，以预测子宫内膜癌患者的错配修复（MMR）和TP53基因表达，并研究基因表达变异性对图像纹理特征的影响。
在这项回顾性探索性研究中，我们生成了两个数据集。第一个数据集包含123例经组织病理学确诊的患者病例，用于开发子宫内膜癌分割模型。第二个数据集包括249例用于MMR预测和179例用于TP53突变预测的患者，数据来源于PET/CT检查和免疫组织化学分析。基于PET的注意力U型网络用于分割，随后结合配准后的PET和CT图像进行区域生长。使用PET、CT和组合数据构建特征模型，并根据性能比较进行模型选择。
我们的分割模型训练准确率达到99.99%，骰子系数为97.35%，验证准确率为99.93%，骰子系数为84.81%。PET+CT组合模型对两个基因均显示出卓越的预测能力，在训练集和测试集中，MMR的AUC分别为0.8146和0.8102，TP53的AUC分别为0.8833和0.8150。MMR相关蛋白异质性和TP53表达差异主要见于PET图像。
已建立一种高效的子宫内膜癌分割深度学习算法，突出了PET和CT联合放射组学对MMR和TP53表达的增强预测能力。该研究强调了MMR和TP53基因表达对肿瘤特征的不同影响。
