Adaptive Hardware/Software Co-Design for Medical Image Processing: From Systematic Review to Adaptive Framework Development

In modern computing, hardware/software co-design represents a pivotal paradigm for advancing image processing systems. This paper presents a multi-faceted investigation that evolves from a systematic review to comprehensive framework development. First, we conduct a systematic review of 22 studies on hardware/software co-design methods for image processing, following PRISMA guidelines. Our analysis identifies key techniques including FPGA-based acceleration, high-level synthesis, and dynamic reconfiguration that address computational bottlenecks. Second, we validate these principles through experimental simulation using the MedMNIST dataset across CPU, FPGA, ARM, and GPU platforms. Results demonstrate FPGA’s optimal balance between latency (10.8 ms) and energy efficiency (51.1% reduction versus CPU), ARM’s superiority in energy-constrained scenarios (0.047 J per inference), and GPU’s maximum throughput (31,335 FPS) for high-volume applications. Third, building on these insights, we propose and develop the Adaptive Hybrid Co-Design Framework (AHCDF), evolving through five iterative versions that progressively address performance-reliability trade-offs. The Enhanced AHCDF version achieves 117,405 FPS throughput with 34.0% clinical constraint satisfaction through machine learning-driven hardware partitioning. Our comparative analysis reveals evolutionary trade- offs: while initial versions prioritize reliability (100% success rate), later versions achieve throughput improvements of up to 14.3× over baseline hardware. The framework demonstrates GPU dominance (50-76% allocation) in throughput-optimized configurations, while maintaining ARM’s energy efficiency for wearable applications. These findings provide quantitative evidence supporting co-design strategies and offer actionable deployment guidelines: ARM for portable diagnostics, FPGA for clinical workstations, and GPU for centralized screening. The paper concludes by highlighting directions for adaptive, secure, and scalable image processing systems that balance performance, energy efficiency, and clinical applicability.