DAACN: Dense Atrous Asymmetric Convolutional Network for HIFU Treatment Target Region Extraction

In the context of HIFU treatment monitoring for ultrasonic image segmentation, there is a trade-off between the segmentation accuracy and the complexity of semantic segmentation methods. High-accuracy methods tend to be computationally complex, while low-complexity methods are prone to missegmentation issues related to the shape and boundary of the treatment target region. To address this challenge., this paper proposes a novel approach called Dense Atrous Asymmetric Convolutional Network (DAACN). DAACN is a lightweight network with an encoder-decoder architecture that incorporates non-symmetric convolutions and dense atrous convolutions. It achieves efficient extraction of image features while maintaining a low computational complexity. Additionally, this study proposes the use of multi-cross-entropy loss to supervise the network and preserve more details and contextual information during the training process, thereby further improving the accuracy of the segmentation. The experimental results demonstrate that the DAACN trained with the multi-cross-entropy loss function achieved a parameter count of 3.22M. Compared to other networks, it exhibits excellent segmentation accuracy, with MIoU and DSC values of 95.07% and 97.47% respectively.