一种改进的DeepLab V3+的医学图像分割方法

Improved medical image segmentation method based on DeepLab V3+

医学图像分割是疾病诊断任务中的关键技术,针对现常用的图像分割网络存在对小物体预测不准以及因局部信息丢失产生网格伪像等问题,设计了一种基于注意机制的DeepLab V3+图像分割方法。在编码(Encoder)部分添加YOLOv5网络中Focus+CBL+CSP的级联式特征提取模块,增强网络能力,同时在Encoder和解码(Decoder)网络加入瓶颈注意力机制,增加目标权重,优化特征提取,获取更多的细节、纹理信息。实验结果证明,该方法优于其他分割方法,改进后的模型在像素准确率比DeepLab V3+提高了0.82%,平均交并比(MIoU)上提高0.99%,改进后的模型提高了小目标组织特征提取的能力,实现了更高精度分割。

Medical image segmentation is a key technology in disease diagnosis.Aiming at the problems of the commonly used image segmentation network,such as poor prediction of small objects and grid artifacts caused by local information loss,this paper designed a DeepLab V3+image segmentation method based on attention mechanism.In the Encoder part,the cascaded feature extraction module of Focus+CBL+CSP in YOLOv5 network is added to enhance the network representation ability.At the same time,the bottleneck attention mechanism is added to the Encoder and Decoder network to increase the target weight,optimize feature extraction,and obtain more details and texture information.The experimental results showed that the proposed method outperformed other methods.The improved model achieved 0.82% higher pixel accuracy than DeepLab V3+and 0.99% higher pixel accuracy than MIoU.The improved model improves the ability of feature extraction of small target organization and achieves higher precision segmentation.