Integrated optical matrix operator for biomedical image segmentation using deep residual U-Net

We present an on-chip optical matrix operator designed for multichannel 1×1 optical convolution, which is deployed in a deep residual U-Net for biomedical image segmentation. The integrated optical matrix operator is fabricated on a silicon-on-insulator (SOI) chip that supports fully reconfigurable 5×5 matrix operation and is used in a deep residual U-net that requires 3-channel 1×1 convolution operation. The multichannel optical 1×1 convolution operation is experimentally validated by using the deep residual U-Net to achieve precise segmentation of pneumonia lesion region in lung CT images. The on-chip optical matrix operator features high scalability and large-scale parallel operation and may find great application in future optical neural networks.