Energy-Efficient VLSI Architecture for Real-Time Retinal Disease Detection using Deep Learning

Efficient processing of high-resolution fundus images is indispensable for the real-time detection of retinal diseases. This study introduces a novel very large-scale integration (VLSI) architecture that utilizes convolutional neural network models to resolve concerns regarding latency, power consumption, and memory efficiency. While detecting retinal disorders, the proposed design achieves a 97.1% classification accuracy with a latency of 40 ms and 0.9W of power. Utilizing Verilog HDL and 28nm CMOS technology, the product is equipped with a fault-tolerant control unit, efficient processing elements, and hierarchical memory management. The architecture's capacity to manage 33 images per second has been demonstrated through testing on benchmark datasets, all while maintaining energy efficiency and throughput. The proposed method considerably outperforms GPU-based methods, particularly in embedded and portable devices, and paves the way for fast, real-time medical diagnostics.