基于解的空间约束的超声前列腺图像分割算法

Ultrasonic prostate segmentation algorithm based on solution space constraint

超声前列腺图像分割是一项极具挑战的任务,目前传统检测算子面对灰度对比不显著的部位难以去辨别,神经网络则忽视了超声图像信噪比低的影响且消耗大量算力。为解决以上问题提出一种高效的基于解的空间约束的超声前列腺图像分割算法,将分割问题转化为求边界点问题,首先对法向量算子改进,改善其检测能力;然后使用降噪自编码器根据形状约束克服噪声优化解的空间;最后引入迭代算子将解的范围限制在极小的区域实现精准分割。实验表明,模型交并比型(IoU)达94.4%,DSC值约97.05%,精度高于当前热门的神经网络算法,且更轻量化。

Ultrasound prostate image segmentation is a very challenging task.At present,traditional detection operators are difficult to identify the parts with insignificant gray contrast.Neural networks ignore the effect of low signal-to-noise ratio of ultrasonic images and consume a lot of computing power.In order to solve the above problems,this paper proposes an efficient segmentation algorithm of ultrasonic prostate images based on spatial constraints of solutions,which transforms the segmentation problem into a boundary point problem.Firstly,the normal vector operator is improved to improve its detection ability.Then the denoising autoencoder is used to overcome the space of the noise optimization solution according to the shape constraint.Finally,iteration operator is introduced to limit the range of solution to a very small area to achieve accurate segmentation.Experiments show that the IoU of this model is 94.4%and the DSC value is about 97.05%,which is higher than the current popular neural network algorithm and lighter.