Abstract:

Accurate breast nodule segmentation in ultrasound images is very challenging due to the low resolution and noise of ultrasound imaging， as well as the complexity and variability of the shape and texture of nodules， therefore， an end-to-end automatic segmentation method of breast nodules in ultrasound images based on multi-scale and cross-spatial feature fusion was proposed. Firstly， a Multi-scale Feature Extraction and Fusion （MFEF） module was designed to enable the network to have multi-scale feature extraction ability by fusing four convolutional paths with different receptive fields. Then， for the multi-scale observation and information filtering of high-level semantic information， a Scale-aware Feature Aggregation （SFA） module was used at the bottleneck layer to enhance the deep feature extraction ability in the encoding stage. Besides， a Cross-spatial Residual Fusion （CRF） module was designed and applied to the skip connection between the encoder and decoder to fuse information among different encoding layers in a cross-spatial way and implement information complementarity between different encoding layers， further extract information features of encoding layer and narrow the difference between peer layers of encoder and decoder to better compensate for the information loss in the decoding stage. Experimental results on a public ultrasound breast nodule dataset show that the proposed method achieves DICE coefficient of 0.888， which is 0.033 to 0.094 higher than those of the mainstream deep learning segmentation models UNet， AttUNet， ResUNet++ and SKUNet， and is 0.001 to 0.068 higher than those of the improved models such as CF²-Net， Estan， FS-UNet and SMU-Net in the same dataset. The subjective visualization of the segmentation result of the proposed method is closest to the gold standards provided by experts， verifying that the proposed mehtod can segment the breast nodule area more accurately.

Key words: multi-scale feature fusion, cross-spatial feature fusion, Scale-aware Feature Aggregation (SFA), Convolutional Neural Network (CNN), breast nodule segmentation in ultrasound images

摘要：

针对超声成像分辨率低且存在噪声、结节形态与纹理复杂多变导致超声乳腺结节精确分割较为困难的问题，提出一种融合多尺度特征提取和跨空间特征融合的超声乳腺结节端到端自动分割方法。首先，设计一种多尺度特征提取与融合（MFEF）模块，通过融合4条具有不同感受野的卷积路径使网络具有多尺度的特征提取能力。其次，为对高级语义信息进行多尺度观察和信息筛选，在瓶颈层采用尺度感知与特征聚合（SFA）模块，以增强编码阶段的深层特征提取能力。此外，设计跨空间残差融合（CRF）模块，并将它应用在编、解码器间的跳跃连接上。该模块一方面对不同编码层进行跨空间信息融合，实现不同编码层间的信息互补；另一方面进一步提取编码层信息特征，缓解编解码对等层之间的语义差异，从而更好地补偿解码阶段的信息损失。在公开的超声乳腺结节数据集上的实验结果显示，所提方法的DICE系数可达0.888，同主流的深度学习分割模型UNet、AttUNet、ResUNet++、SKUNet相比，提高了0.033~0.094，对比相同数据集中的改进模型如CF²-Net、ESTAN、FS-UNet、SMU-Net，提高了0.001~0.068。所提方法分割结果图的主观视觉效果与专家给出的金标准最接近，能更加准确地分割出乳腺结节区域。

关键词: 多尺度特征融合, 跨空间特征融合, 尺度感知与特征聚合, 卷积神经网络, 超声乳腺结节分割