基于全尺度跳跃连接的TransUNet医学图像分割网络

TransUNet medical image segmentation network based on full-scale jump connectivity

医学图像分割是医学图像智能分析的关键技术之一,其中深度学习更是为医学图像分割作出了不少的贡献。为进一步提高U型网络对医学图像分割的精确度,提出基于全尺度跳跃连接和权重深度监督的改进Transformer医学图像分割网络(full-scale jump connectivity and weight depth supervision for improved Transformer medical image segmentation networks, FS-TransUNet3+),该模型通过全尺度跳跃连接的结构,使模型聚合了多层抽象语义特征和空间信息,并且降低了模型的网络参数,提高计算效率。同时采用权重深度监督(weight depth supervision, WDS)的方式,提高了特征学习的表征能力和图像识别精度,细化了目标区域的边界,以及改进了特征聚合机制,将混合编码器与解码器的语义信息进行拼接,强化模型在图像的边缘注意力。在多个医学图像复杂场景分割,如细菌图像数据集、肝脏肿瘤分割挑战数据集、多模态脑肿瘤分割挑战数据集上的参数量、FLOPS、样本相似性指标(DSC)、交并比(IoU)、像素精度(PA)、95HD指标有所改进,有效验证了各个部分的性能,综合分割效果优于其他网络。

semantic features and spatial information.And reduces the network parameters of the model,Improve the computational efficiency.At the same time,the weight depth supervision(WDS)is adopted to improve the representation ability of feature learning and the recognition accuracy of the image,the boundary of the target area is refined,and the feature aggregation mechanism is improved,and the semantic information of the hybrid encoder and decoder is stitched together,so as to strengthen the edge attention of the model in the image.In multiple medical image complex scene segmentation,such as bacterial image data set,liver tumor segmentation challenge data set,Params,FLOPS,DSC,IoU,PA,95HD indicators on the multimodal brain tumor segmentation challenge data set were improved,effectively verified the performance of each part,and the comprehensive segmentation effect is better than other networks.