基于特征尺度的平面波医学影像重建

相较于传统的线扫成像,平面波成像由于其超快的成像速度被广泛应用,但其成像质量较差,影响医生对肿瘤以及血管疾病的精确诊断,现有技术虽然可以提高成像质量,但会降低成像帧频,无法满足临床医学上超快成像的需求。针对上述问题,提出了一种基于生成对抗网络(generative adversarial network, GAN)的图像重建方法:MF-GAN(generative adversarial network with multiscales and feature extraction)。采用基于U-Net的生成器,在编码器中结合多尺度感受野提取不同层次的信息,在解码器中提出了叠加采样机制(fusion-sampling mechanism, FSM),并结合交叉自注意力(criss-cross self-attention, CCSA)分别提取局部和全局特征。在PICMUS 2016数据集上进行训练,利用组合损失规范该模型的收敛方向,相较主流基于深度学习和波束合成的方法,在点目标、囊肿目标和体内图像中的重建效果均有明显提升。综上所述,MF-GAN能够解决平面波图像病灶部位不清晰的问题,重建出高质量的平面波图像。

Detecting and Tracking Moving Targets on Omnidirectional Vision

A method was presented to implement the detecting and tracking of moving targets through omnidirec-tional vision. The method combined optical flow with particle filter arithmetic, in which optical flow field was used to detect and locate moving targets and particle filter was used to track the detected moving objects. According to the circular image character of omnidirectional vision, the calculation equation of optical flow field and the tracking arithmetic of particle filter were improved based on the polar coordinates at the omnidirectional center. The edge of a randomly moving object could be detected by optical flow field and was surrounded by a reference region in the particle filter. A dynamic motion model was established to predict particle state. Histograms were used as the fea-tures in the reference region and candidate regions. The mutual information (MI) and Gaussian function were com-bined to calculate particle weights. Finally, the state of tracked object was computed by the total particle states with weights. Experiment results show that the proposed method could detect and track moving objects with better real-time performance and accuracy.