Energy-Efficient 3D Convolution Using Interposed Memory Accelerator eXtension 2 for Medical Image Processing

Energy-efficient medical image processing is crucial in mobile or remote healthcare situations where traditional GPU-based solutions are not feasible. Recently, three-dimensional (3D) image processing has gained significant importance in fields like computer vision, machine learning, natural language processing, and medical diagnosis. 3D convolution neural networks (CNN) have outperformed state-of-the-art in many visual recognition tasks, such as medical imaging, video processing and analysis, and 3D Object recognition. 3D CNNs excel at feature extraction but impose a computational burden, mainly from convolution layers. Their cubic complexity growth hinders speed and overall performance, requiring complexity reduction in these layers, as they dominate 3D CNN calculations. To tackle these challenges, we introduce a novel energy-efficient computational solution of the Interposed Memory Accelerator eXtension 2 (IMAX2), a Coarse-Grained Linear Array (CGLA) developed in our laboratory, which outperforms the RTX3090 by achieving 7.37 times the efficiency in Tops/W for 3D CNNs. Unlike fixed Application-Specific Integrated Circuits, IMAX2 offers remarkable flexibility for computations, making it a versatile choice for complex tasks. Specifically, we optimized the computational bottleneck of 3D convolutions within the U-Net architecture, a specialized CNN model designed for segmentation tasks. Our initial findings demonstrate that IMAX2 empowers medical image analysis with 3D CNNs while achieving superior energy efficiency. This study opens up new possibilities for Computer-Aided Diagnosis and AI-driven medical imaging solutions in settings where conventional high-power systems are impractical.