基于U-Net的COVID-19病灶医学影像ZMINet分割模型

ZMINET: MEDICAL IMAGE SEGMENTATION MODELOF COVID-19 LESIONS BASED ON U-NET

针对COVID-19病灶医学影像边缘模糊及小病灶丢失的问题,基于多尺度多层级特征集成与多分支交互式学习策略对U-Net进行改进,提出一种基于U-Net的COVID-19病灶医学影像ZMINet分割模型。原图被缩放至多个尺度输入编码器;利用SIU-AIM模块多分支交互式地逐层学习融合层内与层间的多尺度关键特征,并将特征传入由SIM组成的解码器之对应层;SIM利用交互式学习策略以获取更丰富的多尺度信息表征,并自下而上集成多层特征;使用UAL作为损失函数指导模型输出更清晰的预测结果。在公开数据集上的对比实验表明,ZMINet分割模型在Dice、精确率、特异性和平均绝对误差等指标分别达到了79.2%、81.8%、96.8%和6.3%,与其他算法相比其性能得到了明显的提升。

Aimed at the problem of ambiguous edges and missing of small lesions in medical image segmentation of COVID-19 lesions,a medical image segmentation model named ZMINet of COVID-19 lesions is proposed,which improves U-Net based on multi-scale multi-level features integration and multi-branch interactive learning strategies.The original image was resized to multiple scales and inputted to the encoder.The SIU-AIM module applied multi-branch interactive learning strategy to study and fuse multi-scale key features from each layer and interlayers,and the resulting features were transmitted to the corresponding layer of the decoder composed by SIMs.The SIM module used interactive learning strategy to obtain richer multiple-scale feature expressions,and integrated multi-layer features from bottom to top.The UAL was used as the loss function to guide the model to output clearer prediction results.Comparative experiments on public datasets show that ZMINet has achieved 79.2%,81.8%,96.8%and 6.3%in dice,precision,specificity and mean absolute errors respectively,and the performance of ZMINet has significantly outperformed other algorithms.