Подавление шума на ультразвуковом изображении с использованием генеративных состязательных сетей
Аннотация
Ultrasound imaging is important for diagnosing noninvasive lesions when developed diagnosis may not be possible.
The research represent a convolutional neural network for image classification to detect fragments in ultrasound images. As a first application, fragments of different types and sizes were first placed in phantom tissue and then in pig femoral tissue.
The architecture of the algorithm was stepwise optimized to minimize verification losses and maximize F1. The final version of the algorithm, trained on sets of tissue phantom images, had an F1 score of 0.95 and an area under the ROC curve of 0.95.It returned more than 90% accuracy for each of the eight types of fragments. When trained only with pig image sets, the optimized format of the algorithm achieved even higher values: F1 and area under the ROC curve 0.99. Overall, the developed algorithm provides high classification accuracy for both phantom tissues and animal tissues.
The research represent a convolutional neural network for image classification to detect fragments in ultrasound images. As a first application, fragments of different types and sizes were first placed in phantom tissue and then in pig femoral tissue.
The architecture of the algorithm was stepwise optimized to minimize verification losses and maximize F1. The final version of the algorithm, trained on sets of tissue phantom images, had an F1 score of 0.95 and an area under the ROC curve of 0.95.It returned more than 90% accuracy for each of the eight types of fragments. When trained only with pig image sets, the optimized format of the algorithm achieved even higher values: F1 and area under the ROC curve 0.99. Overall, the developed algorithm provides high classification accuracy for both phantom tissues and animal tissues.