UBDR-FB:一种针对脂肪型乳腺超声大数据识别方法

UBDR-FB: A Method of Ultrasound Big Data Recognition for Fat Breast

目的:新疆地区女性多为脂肪型乳腺,B超回声影像具有很强的特殊性,当前国内外的深度学习模型识别率偏低。方法:为提升新疆地区乳腺超声影像的识别精确度,提升辅助诊断效果,本文提出了一种基于U-Net网络的超声乳腺肿瘤的分割方法UBDR-FB(ultrasound big data recognition for fatty breast)。UBDR-FB包含编码和解码两部分,在编码的过程中,UBDR-FB利用下采样来捕获不同分辨率的病理图像的特征信息;在解码的过程中,对特征信息进行上采样融合至初始像素。同时,UBDR-FB使用了跳跃连接方法以提高超声影像的识别效率。结果:本文收集了8430张新疆脂肪型乳腺超声影像,经过标准化处理后,利用UBDR-FB深度学习模型建立了具有新疆地域特征的超声模型。结果:UBDR-FB方法识别脂肪型乳腺规则病灶准确率均值为80%,最高可达92%。

Objective: Most women in Xinjiang are fat mammary glands. B ultrasound echo images have strong particularity, and the recognition rate of in depth learning models at home and abroad is low. Method: In order to improve the accuracy of breast ultrasound image recognition and the effectiveness of assisted diagnosis in Xinjiang region, this paper presents a U-Net based ultrasound breast tumor segmentation method UBDR-FB(Ultrasound Big Data Recognition for Fatty Breast). UBDR-FB contains two parts, encoding and decoding. In the encoding process, UBDR-FB uses downsampling to capture the feature information of pathological images with different resolutions. During the decoding process, the feature information is up-sampled and fused to the initial pixel. At the same time, UBDR-FB uses a skip connection method to improve the efficiency of ultrasound image recognition. Result: This paper collects 8430 ultrasound images of fat type breast in Xinjiang. After standardization, UBDR-FB deep learning model is used to build an ultrasound model with regional characteristics of Xinjiang. Conclusion: The experimental results ultimately show that the average accuracy of UBDR-FB method in identifying fat-type breast regular lesions is 80%, up to 92%.