上海科技大学知识管理系统

Implicit neural representation (INR)-based medical imaging techniques are becoming popular due to their efficiency in representation and ability to produce high-quality reconstructions. Nevertheless, INR-based methods are computational demanding, posing challenges for deployment in practical applications. To address the computation challenge while maintain system flexibility for different applications, we propose a RISC-Vbased domain specific acceleration system for INR-based medical image reconstruction. Our system features a dedicated computing engine for reconstruction and a RISC-V processor for system configuration and runtime management. To validate the proposed system, we implemented it using VerilogHDL and constructed a prototype system on field programmable gate array (FPGA). The experimental results affirm the advantages of the proposed system over existing designs in terms of flexibility and efficiency.