Ultrasonic image segmentation

1. introduction

The parameters of human organs can be obtained by ultrasonic image segmentation , It is of great significance to evaluate the function of human organs . But ultrasonic image has speckle noise 、 Regional blur 、 Weak boundary 、 It is difficult to locate the region of interest （ROI） Other questions , As a result, the current automatic segmentation technology can not guarantee the segmentation accuracy , But only relying on manual segmentation of the target area has a huge workload , And subjective factors are strong .

2. Method

At present, ultrasonic image segmentation generally needs image preprocessing 、 ROI localization and image segmentation ：

① Image pre-processing stage is image denoising and image enhancement .

② The location of the region of interest is to select the target location to be segmented .

③ Finally, the target region is segmented .

Medical image segmentation methods are mainly divided into two categories ： One is segmentation algorithm based on training set learning , Including supervised learning and semi supervised learning , It needs to be trained by manually labeling images . Such methods include segmentation algorithms based on active shape model and recently popular semantic segmentation algorithms based on deep learning , The former depends on the positioning of the initial contour , The latter requires manual production of a large number of training sets , A lot of work . The other is the segmentation algorithm based on clustering , That is, unsupervised segmentation algorithm , There is no need to label the training set manually , There has been a Markov random field segmentation algorithm 、 Active contour model optimized by genetic algorithm 、 The combination of standardized cutting methods K Mean algorithm 、 Based on adaptive kernel bandwidth mean shift Image segmentation algorithm 、 Mean shift - Region expansion fitting (MS- RSF) Model, etc .

3. Ultrasonic image segmentation method based on deep learning

There are several problems in ultrasonic image segmentation based on deep learning ：

• The extracted features are redundant ;
• Lack of dimension data , It makes the network prone to over fitting .

Current image segmentation networks often combine attention mechanism .