A HIGH-EFFICIENCY VLSI CNN ACCELERATOR FOR REAL-TIME BIOANALYTICAL IMAGING ON EDGE DEVICES

Abstract

The current trend of the rapid growth of artificial intelligence in healthcare has put pressure on developing real-time, low-power processing of biomedical images on mobile devices, but the majority of convolutional neural network (CNN) models cannot be easily reconfigured to operate in edge environments. The new VLSI CNN accelerator proposed will provide high throughput and energy-efficient inference of the bioanalytical imaging application. Off-chip memory bandwidth and processing element underutilization limitations are overcome using the proposed architecture that is based on a scalable systolic array compute engine and on-chip memory tiling, lightweight data compression, and adaptive power management. Analysis of clinical relevance using blood smear (BCCD) and histopathology (CAMELYON16) datasets and models LeNet-5 and MobileNetV1 has demonstrated that the accelerator can achieve up to 94.8 GOPS throughput, 1.12 TOPS/W energy efficiency, and <7 ms inference latency per image region, compared to the industry-leading accelerators, including Eyeriss and TPU-Edge. The platform continues to show classification accuracy of less than 1 percent of floating-point baselines in diagnostic workflows. Such outcomes establish the feasibility of the proposed accelerator in next-generation point-of-care and laboratory automation, opening the door to scalable and feasible edge AI applications in real-time bioanalytical imaging.