【摘要】 目的 基于多模态磁共振成像与深度学习的融合技术，构建DenseNet121及InceptionV3两种卷积神经网络模型，实现新生儿急性胆红素脑病的早期精准诊断。方法 选取覆盖苍白球区域的连续T1加权图像、T2加权图像、表观扩散系数图作为卷积神经网络输入，在T1WI、T2WI、ADC三种单模态数据集上分别训练DenseNet121(SoftMax)、InceptionV3(SoftMax)、DenseNet121+SVM和I nceptionV3+SVM，比较不同分类器的性能。基于DenseNet121和InceptionV3提取深度特征，采用支持向量机分类器，对不同模态特征融合方法进行比较。运用五倍交叉验证方法来评价模型的泛化能力，采用曲线下面积(AUC)、灵敏度、特异度、准确度、精确度、以及F1评分(F1-score)评估模型的分类性能。结果 在单模态特征中，InceptionV3+SVM和DenseNet121+SVM在T1WI上的AUC值最高，T2WI次之，ADC最低。InceptionV3+SVM和DenseNet121+SVM在双模态特征融合中的AUC值高于任何一个单模态特征的AUC值。在三模态特征融合中，T1WI+T2WI+ADC特征融合的AUC值高于单模态及双模态特征的AUC值，其中DenseNet121+SVM表现最佳，AUC值为0.8730，准确度为88.17%。结论 多模态MRI结合深度学习为新生儿急性胆红素脑病的早期诊断和及时治疗提供了客观的影像学依据。更多还原

【Abstract】 Objective Based on the fusion technology of multimodal magnetic resonance imaging and deep learning, two convolutional neural network models, DenseNet121 and InceptionV3, are constructed to achieve early and accurate diagnosis of neonatal acute bilirubin encephalopathy. Methods Continuous T1-weighted imaging(T1 WI), T2-weighted imaging(T2 WI), and apparent diffusion coefficient(ADC) maps covering the globus pallidus region were selected as inputs for the convolutional neural network(CNN). Four models—DenseNet121(with SoftMax), InceptionV3(with SoftMax), DenseNet121+SVM, and InceptionV3+SVM—were trained on the three single-modal datasets(T1 WI, T2 WI, and ADC) to compare the performance of different classifiers. Based on the deep features extracted by Dense Net121 and InceptionV3, a support vector machine(SVM) classifier was employed to evaluate different feature fusion strategies. A five-fold cross-validation approach was used to assess the models’ generalization ability, while classification performance was evaluated using metrics including the area under the curve(AUC), sensitivity, specificity, accuracy, precision, and F1-score. Results Among single-modal features, InceptionV3+SVM and DenseNet121+SVM exhibited the highest AUC values on T1 WI, followed by T2 WI, with ADC performing the lowest. For dual-modal feature fusion, both Inception V3+SVM and DenseNet121+SVM achieved higher AUC values than any single-modal feature. In multimodal feature fusion(T1 WI+T2 WI+ADC), the AUC value surpassed those of both single-modal and dual-modal features, with DenseNet121+SVM demonstrating the best performance—achieving an AUC of 0.8730 and an accuracy of 88.17%. Conclusion Multimodal MRI combined with deep learning provides objective imaging evidence for the early diagnosis and timely treatment of neonatal acute bilirubin encephalopathy.更多还原