边界增强像素级多模态声像图融合方法评估良、恶性乳腺肿瘤

目的 观察以边界增强多模态乳腺声像图像素级特征融合方法评估良、恶性乳腺肿瘤性质的价值。方法 基于乳腺肿瘤B型声像图提取边界增强图像,于超声弹性复合声像图中提取纯弹性信息图像。对多模态乳腺肿瘤声像图进行像素级特征融合,形成边界特征增强的融合图像,再以卷积神经网络(CNN)进行分类;评估融合方法分类良、恶性乳腺肿瘤的性能,并与单模态方法、特征级融合方法、无边界增强像素级图像融合方法及其他CNN模型进行对比。结果 边界增强像素级特征融合方法有助于CNN提取乳腺肿瘤特征,分类良、恶性乳腺性能最佳,其分类准确率为85.71%,特异度为85.49%,敏感度为86.16%,模型稳定。结论 边界特征增强像素级多模态声像图融合方法可用于判断良、恶性乳腺肿瘤。

Boundary enhancement pixel-level image fusion of multi-modal breast ultrasound images for evaluating benign and malignant breast tumors

Objective To observe the value of boundary enhancement pixel-level feature fusion of multi-modal breast ultrasound images for evaluating benign and malignant breast tumors. Methods Boundary-enhanced images were extracted from B-mode ultrasonograms of breast tumors, and pure-elastic information images were extracted from elastic composite ultrasonograms. Then multi-modal breast tumor ultrasonograms were fused with pixel-level features to form fused images with boundary feature enhancement, which were then classified using convolutional neural network (CNN). The performance of fusion method in classifying breast tumors was analyzed and compared with that of single-modal method, feature-level fusion method, pixel-level image fusion method without boundary enhancement, also with other CNN models. Results Boundary-enhanced pixel-level feature fusion was helpful for extracting features of breast tumors with the best classification performance using CNN, with classification accuracy of 85.71%, specificity of 85.49%, sensitivity of 86.16%,and the model was stable. Conclusion Pixel-level multi-modal ultrasound image fusion with boundary feature enhancement could be used to judge benign and malignant breast tumors.