基于乳腺X线摄影的神经网络深度学习模型对乳腺病变的诊断价值

Value of a neural network deep learning model based on mammography in the diagnosis of breast lesions

目的 构建一种基于乳腺X线摄影的乳腺良恶性病变深度学习模型并评估其诊断效能。方法 回顾性收集行乳腺X线摄影检查且经超声引导下穿刺活检或术后获得病理结果的乳腺肿瘤女性病人210例,平均年龄(53.3±11.2)岁。根据病理结果将病人分为良性组82例,恶性组128例。按照2∶1比例随机将病人分为训练集140例(良性53例,恶性87例)和验证集70例(良性29例,恶性41例)。以残差网络(ResNet)深度神经网络框架为基础,基于训练集数据构建鉴别乳腺肿块良恶性的诊断模型,并使用验证集数据评估该诊断模型的诊断效能。由高、中、低年资3名放射诊断医师分别对验证集肿块良恶性进行主观评价。良恶性组间以及训练集和验证集间病人基本临床资料比较采用t检验、Mann-Whitney U检验。采用加权Kappa方法分析不同年资放射医师和深度学习模型间的一致性。采用多阅片者-多病例(MRMC)的受试者操作特征(ROC)曲线下面积(AUC)评估深度学习模型和不同年资放射医师的诊断效能,并采用DeLong检验比较各分析对象之间的诊断效能。结果 训练集和验证集中,恶性组的病人年龄和病灶直径均大于良性组(均P<0.05)。深度学习模型与低、中、高年资医师对良恶性肿瘤分类的一致性中等或较好(κ值分别为0.546、0.656、0.788)。在训练集或验证集中,深度学习模型与高年资医师对乳腺良恶性病变诊断的AUC值差异无统计学意义(均P>0.05),而高于中、低年资放射医师(均P<0.05)。结论 基于神经网络的深度学习模型在乳腺X线摄影良恶性病变鉴别诊断中的诊断效能较高,相当于高年资放射医师水平,可为临床医生决策提供参考依据。

Objective To construct a deep learning model based on mammography for differentiating benign from malignant breast lesions and to assess its diagnostic performance.Methods A retrospective analysis was conducted on 210 female patients with breast tumors who underwent mammography and received pathology results via ultrasound-guided biopsy or post-surgical examination,with an average age of 53.3±11.2 years.Patients were categorized into a benign group(82 cases)and a malignant group(128 cases)based on pathology results.They were randomly divided at a 2∶1 ratio into a training set(140 cases,with 53 benigns and 87 malignants)and a validation set(70 cases,with 29 benigns and 41 malignants).A diagnostic model was constructed based on the ResNet deep neural network framework to distinguish between benign and malignant breast lesions using the training set,and its diagnostic performance was evaluated with the validation set.Three radiologists with high,medium,and low experience levels independently performed subjective assessments of benign and malignant lesions in the validation set.Clinical data comparisons between benign and malignant groups and between training and validation sets were analyzed using t-tests and Mann-Whitney U tests.The weighted Kappa method was applied to evaluate consistency between the deep learning model and radiologists at different experience levels.The area under the receiver operating characteristic(ROC)curve(AUC)was calculated using the multi-reader,multi-case(MRMC)method to assess the diagnostic performance of the deep learning model and radiologists.DeLong’s test was used to compare diagnostic performance across groups. Results In both the training and validation sets, patients in the malignant group were older and had larger lesion diameters compared to the benign group (both P<0.05). The consistency between the deep learning model and radiologists with low, medium, and high experience levels in classifying benign and malignant tumors was moderate to good (κ values of 0.546, 0.656, and 0.788, respectively). In both the training and validation sets, the AUC of the deep learning model for diagnosing benign and malignant breast lesions was comparable to that of the highly experienced radiologist (both P>0.05), and was higher than that of radiologists with medium and low experience levels (both P<0.05). Conclusion The neural network-based deep learning model shows high efficiency in the differential diagnosis of benign and malignant lesions on mammography, comparable to that of highly experienced radiologists, and may provide valuable reference support for clinical decision-making.