面向乳腺超声分类的低尺度形态特征重校准方法

Low-Scale Morphological Feature Recalibration Method for Breast Ultrasound Classification

针对乳腺超声图像具有类内差异大、类间差异小以及结节形状复杂多变等问题,提出一种形状特征重校准的乳腺超声图像算法,实现乳腺超声的自动化诊断.首先,构建端到端的网络模型,采用渐进训练方式,充分学习图像中更具辨别力的区域,获取更细粒度的特征信息;其次,提出分区打乱机制,降低网络中打乱图像时破坏结节区域所产生的噪声;然后,将模型底层提取的特征与通过掩膜图像获得的形状特征进行重校准,提出低尺度重校准损失函数;最后,构建一个包含1550张乳腺超声图像数据集LSRD(low-scale recalibration database),验证所提方法的有效性.实验结果表明:本文模型在LSRD上准确率94.3%、敏感性91.2%、特异性93.6%、ROC(receiver operator characteristic curve)与坐标围成的面积(area under curve,AUC)为0.941,均优于对比模型;在BUSI(breast ultrasound image)数据集上,相较于对比模型,其分类精度提升3.3%.

Breast ultrasound images have large intra-class differences,small inter-class differences,and complex and variable nodule shapes.In order to address these issues,a breast ultrasound image algorithm with morphological feature recalibration was designed to realize automatic diagnosis of breast ultrasound.First,an end-to-end network model was built,which adopted progressive training to fully learn the more discriminative regions in the image and obtain more fine-grained feature information.Secondly,a partition shuffle mechanism was proposed to reduce the noise caused by the disruption of the nodule region when the image was shuffled.Then,the features extracted from the bottom layer of the model were recalibrated with the morphological features obtained through the mask image,and a low-scale recalibration loss function was proposed.Finally,in order to verify the effectiveness of the proposed method,a low-scale recalibration database(LSRD)containing 1550 breast ultrasound images was constructed.The experimental results show that the accuracy of the proposed model on LSRD is 94.3%;the sensitivity is 91.2%;the specificity is 93.6%,and the area(AUC)under the receiver operator characteristic curve(ROC)is 0.941,all of which are superior to other comparison models.On the breast ultrasound image(BUSI)dataset,compared with the other models,the classification accuracy of the proposed model is improved by 3.3%.