Hardware Acceleration for Machine Learning in Medical Imaging: Case Study on Colorectal Polyp Segmentation

The rapid evolution of machine learning (ML) and deep learning (DL) has spurred the development of specialized hardware accelerators to meet the growing computational demands. In the field of cancer analysis, particularly for colorectal polyp segmentation, the need for robust and fast ML models is critical for accurate and timely diagnosis and treatment. Dedicated Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs) are prominent among the specialized hardware. This paper provides a comprehensive review of the current state-of-the-art, comparing the training performance of ML models on Central Processing Units (CPUs), dedicated GPUs, and TPUs, focusing on their application in colorectal polyp cancer segmentation. The proposed method utilizes five hardware accelerators: NVIDIA A100, L4, and T4 GPUs, Google TPU v2, and Intel Core i9-13900KF CPU. The results show that the A100 GPU (0.17 hours) is the most efficient, followed by the L4 (0.43 hours), T4 (1.04 hours), CPU (3.45 hours), and TPU v2 (10.94 hours).