腺体和息肉的自动分割是人工智能辅助结直肠腺癌诊断的基础，但医学图像中的分割目标大小、形状多变，基于单一的卷积神经网络的自动分割方法已陷入瓶颈。基于此，提出了一种卷积神经网络和自注意力相结合的双分支网络（LG UNet），用以提升分割的精度。首先，基于U-Net设计了Local UNet分支，利用卷积神经网络的优势，学习分割目标的局部信息。然后在Global Transformer分支中，利用Transformer全局依赖关系的学习能力来优化分割细节。最后在编码过程中通过交叉融合模块将Local分支和Global分支的特征图进行融合，将两者优势互补。在腺体分割挑战数据集Glas的两个测试子集Test A和Test B上，以Dice系数和交并比（IOU）系数为主要评价指标，LG UNet的测试结果分别为93.62%、88.44%和88.17%、80.49%。在息肉分割数据集Kvasir-SEG上，LG UNet的Dice系数和IOU系数分别为85.63%和77.82%。实验结果表明，结合Transformer和卷积神经网络优势的LG UNet在腺体和息肉分割上取得了更好的性能。

Two-dimensional (2D) Te nanosheets were successfully fabricated through the liquid-phase exfoliation (LPE) method. The nonlinear optical properties of 2D Te nanosheets were studied by the open-aperture Z-scan technique. Furthermore, the continuous wave mode-locked Nd:YVO₄ laser was successfully realized by using 2D Te as a saturable absorber (SA) for the first time, to the best of our knowledge. Ultrashort pulses as short as 5.8 ps were obtained at 1064.3 nm with an output power of 851 mW. This primary investigation indicates that the 2D Te SA is a promising photonic device in the fields of ultrafast solid-state lasers.

在高速光电子器件的微波封装过程中,需要综合考虑封装寄生参数和芯片寄生参数对器件高频性能的影响。利用封装寄生参数对芯片寄生参数的补偿作用,成功实现了10 Gb/s电吸收调制激光器(EML)的高频封装。通过封装前后芯片和器件的小信号频率响应测试结果对比,器件的反射参数和传输参数有所改善,3 dB带宽达到10 GHz;并进行了10 Gb/s速率的光纤传输实验,经过40 km光纤传输后通道代价不到1 dBm(误码率为10-12),满足10 Gb/s长距离光纤传输系统的要求。