Abstract: Traditional segmentation algorithms have problems such as low accuracy, low precision and time- consuming. Therefore, an adaptive segmentation algorithm for 3D breast ultrasound images using improved deep learning is proposed. First, the images are pre-processed, and the deep multiple example learning method is used to detect the lesion image blocks and remove the normal image blocks. Second, the breast ultrasound image data set is augmented and processed for neural network training. Then, a residual convolutional neural network model is constructed, a residual learning unit is designed, a feature mapping is formed by combining the augmented dataset. A softmax function is used to train the network and perform feature block judgment. Finally, combined with threshold settings, the achieves adaptive segmentation of 3D breast ultrasound images is realized. The results show that the proposed algorithm can complete image segmentation more carefully, and the average running time of the algorithm is 52. 3 s, the image segmentation accuracy is 95. 5% , the F1 score value is high, and the overall performance is good, which provides a reference for the application of convolutional neural network segmentation.

Key words: 3D breast ultrasound, lesion detection, data set amplification, residual convolutional neural network, deep learning