Research on breast ultrasound image segmentation method based on deep learning
Xiaofeng Song, Yu Zhao, Ke Yi, Yuejin Zhang · 2024
In the medical image segmentation task, many studies choose to increase the network depth to achieve higher detection accuracy, but the network parameters of the deep convolutional neural network are too large, and the number of medical images is very complex, leading to a long segmentation time. Therefore, to solve the above problems, an improved U-SegNet network was proposed to segment breast tumors. The experimental results show that the improved U-SegNet reduces the experimental parameters while ensuring the accuracy. Due to the different shapes of breast tumors and the unclear boundary of ultrasonic image shapes of malignant tumors, the segmentation accuracy of malignant tumors is less than that of benign tumors in the task of medical image segmentation. A new segmented network, an optimized network based on RefineNet, is proposed, which uses two different scale inputs sent to ResNet101 and ResNet34 for downsampling to obtain feature maps with different resolutions. Finally, through experimental comparison, the experimental results show that the segmentation accuracy of this method is improved.