Hardware-Compatible U-Net for Low-Dose PET Reconstruction

This work presents a hardware-compatible U-Net designed for the reconstruction of low-dose PET images, integrating depthwise separable convolution and quantization techniques to enhance efficiency and compatibility with Field Programmable Gate Array implementations. By leveraging DWSC, a 7.42x reduction in parameter count is achieved, significantly decreasing model size and improving power efficiency, while maintaining reconstruction quality. Quantization further reduces the model size by 4x and substantially improves inference speed, making the model well-suited for resource-constrained environments.