|
| Reference | Mathematical models and methods | Applicable scene | Data set | Result | Limitations and basis |
|
| [30] | Image enhancement by wavelet transform and improved AlexNet network structure for ultrasound image segmentation | Wavelet transform, which has an excellent effect in image enhancement; AlexNet has low network complexity and fewer parameters, and performs well in the case of insufficient samples | 325 breast ultrasound images provided by the ultrasound department of a tertiary hospital | The indicators of the segmentation results are better than the existing methods | The model only uses 325 original images, and the sample data set is too small, resulting in poor robustness and adaptability; the extracted abstract features are limited, which limits the segmentation ability of the full convolutional network |
|
| [31] | Improving convolutional neural networks using residual learning, sobel enhancement + GLCM | It can effectively solve the problem of network degradation and gradient fragmentation | Intravascular ultrasound images of 63 patients with carotid atherosclerosis marked by professional physicians in the fourth military medical university | 87.1% accuracy | Using the residual network and increasing the depth of the network, the neural network model becomes more complex |
|
| [32] | Improved CNN architecture | Video image classification | Video image of an echocardiogram | A classification accuracy of 92.10% can be achieved | The generalization ability is weak, two 2D CNNs are used, and the interframe motion information of the time dimension is not considered |
|
| [33] | Deep DenseNet and transfer learning | It can alleviate the problem of gradient disappearance and strengthen feature propagation; efficiently reuse features; and reduce the number of parameters | 7230 ultrasonic images of 6 abdominal organs | Classification accuracy rate 86.40% | Network training takes a long time, and DenseNet uses dense connections, which increases the amount of network parameters and calculations |
|
| [34] | Improved U-Net’s modified video segmentation algorithm | Suitable for image segmentation, especially in medical image segmentation | 7230 ultrasonic images of 6 abdominal organs | Segmentation accuracy rate 81.72% | The guidance algorithm for identifying and locating the liver is not reasonable and effective, and the guidance method that only increases the liver area by moving the probe |
|
| [35] | Automatic detection of intima and media-adventitia boundaries in coronary IVUS images based on deep convolutional networks DFCN-1 and DFCN-2, combined with stacked funnel networks and generative adversarial learning for automatic detection of key tissue boundaries | Stacked funnel network can achieve automatic labeling, and generative adversarial learning can alleviate the problem that deep convolutional networks require a large number of labeled medical ultrasound images | 435 20 MHz IVUS images of the international standard IVUS image public database SetB | The test result is 94.00% | Insufficient anti-interference ability and cross-domain segmentation generalization ability; low-lying, high-lying and other pixel-level regional errors are difficult to avoid |
|
