目的 基于深度学习，利用盆腔MRI图像,构建子宫轮廓、子宫内膜及子宫肌瘤病灶的自动分割模型，并评价其分割性能；其次，构建并测试一种全新的子宫肌瘤FIGO分型决策树专家系统。方法 本研究回顾性纳入199例子宫肌瘤患者，在盆腔MRI的轴位T2加权成像（T2-weighted imaging, T2WI）和矢状位T2WI上，分别使用深睿医疗科研平台逐一勾画出子宫轮廓、子宫内膜及子宫肌瘤，并选择对应的子宫肌瘤FIGO分型，得到手工标注标签作为金标准，将199名患者按8：2的比例随机分配至训练集和测试集。基于nnU-Net深度学习算法，通过自适应的方法，根据训练集数据的大小分布以及GPU容量，自动化配置数据预处理方式（图像重采样和像素值正则化）和模型结构。通过对标注数据进行学习建立子宫轮廓、子宫内膜及子宫肌瘤的自动分割模型。采用5折交叉验证的方式训练模型，把数据分成5份，每个模型使用其中的4份进行模型训练，得到5个模型。最终，使用5个模型的集成预测结果作为最终的输出，采用相似度系数（Dice similarity coefficient，DSC）、豪斯多夫距离的95分位数（95%Hausdorff distance，HD 95）、平均表面距离（Average surface distance，ASD）来评估模型的分割性能。然后使用手工标注的分割图像作为分型模型的输入对象，从图像中自动提取出肌瘤轮廓及肌瘤和周边组织的关系信息，根据指南规范和放射科医师的诊断经验，构建子宫肌瘤FIGO分型决策树专家系统，并计算分型模型的准确率、查全率、查准率、F1值。结果 子宫轮廓、子宫内膜和子宫肌瘤在轴位T2WI上的自动分割模型的平均DSC分别为0.920、0.734、0.807；平均HD95分别为7.324、11.057、8.558mm；平均ASD分别为1.091、2.201、1.725mm；在矢状位T2WI上的DSC分别为0.913、0.808、0.781；HD95分别为7.426、6.654、9.772mm；平均ASD分别为1.236、0.719、2.034mm。子宫肌瘤FIGO分型决策树专家系统的总体准确率为0.869，宏查全率为0.859，宏查准率为0.765，宏F1值为0.792。结论 本研究构建了基于nnU-Net网络结构的子宫轮廓、子宫肌瘤及子宫内膜自动分割模型，达到了较好的分割性能，有一定的应用前景；此外，本研究还构建了子宫肌瘤FIGO分型决策树专家系统，可以准确地对子宫肌瘤进行分型，诊断效能良好，未来可用于指导临床决策。

Objective: To construct an automatic segmentation model for uterine contour, endometrium and fibroid lesions based on deep learning using pelvic MRI images and evaluate its segmentation performance; secondly, to construct and test a new FIGO classification system decision tree expert system for uterine fibroids. Methods: In this study, 199 patients with uterine fibroids were retrospectively included, and the uterine contour, endometrium and fibroids were outlined one by one using the Deepwise Medical Research Platform on axial T2-weighted imaging (T2-weighted imaging,T2WI) and sagittal T2WI of pelvic MRI, respectively, and the corresponding FIGO type of fibroids was selected to obtain manual labeling labels as the gold standard, randomly assign 199 patients to the training set and test set in the ratio of 8:2. Based on the nnU-Net deep learning algorithm, the data pre-processing method (image resampling and pixel value regularization) and model structure are automatically configured according to the size distribution of the training set data and the GPU capacity through an adaptive approach. Automatic segmentation models for uterine contours, endometrium and fibroids are established by learning from labeled data. A 5-fold crossvalidation approach is used to train the models by dividing the data into 5 parts and using 4 of them for model training for each model to obtain 5 models. Finally, the integrated prediction results of the 5 models were used as the final output. The similarity coefficient (DSC), the 95th percentile of the Hausdorff distance (95% Hausdorff distance, HD95), the mean surface distance (Average surface distance, ASD) are used to evaluate the segmentation performance of the model. Then, using the manually annotated segmented images as the input object of the fractionation model, the information of the fibroid contour and the relationship between the fibroid and the surrounding tissues were automatically extracted from the images, and the expert system of FIGO classification decision tree for uterine fibroids was constructed according to the guideline specifications and the diagnostic experience of radiologists, and the recall rate, the precision rate, the detection accuracy, and the F1 value of the fractionation model were calculated. Results: The mean DSC of the automatic segmentation models for uterine contour, endometrium and fibroid on axial T2WI were 0.920, 0.734, 0.807; mean HD95 were 7.324, 11.057, 8.558mm; mean ASD were 1.091, 2.201, 1.725mm; on sagittal T2WI were 0.913, 0.808, 0.781; HD95 were 7.426, 6.654, 9.772 mm; mean ASD were 1.236, 0.719, 2.034 mm, respectively. The overall accuracy of the FIGO classification system decision tree expert system for uterine leiomyoma was 0.869, the macroscopic recall rate was 0.859, the macroscopic precision rate was 0.765, and the macro F1 value was 0.792. Conclusion: This study constructed an automatic segmentation model of uterine contour, endometrium, and uterine fibroids based on the nnU-Net network structure, which achieved a good segmentation performance and has some application prospects; in addition, this study also constructed a FIGO classification decision tree expert system for uterine fibroids, which can accurately classify uterine fibroids with good diagnostic efficacy and can be used to guide clinical decisions in the future.