Integration of High-Performance Computing and Artificial Intelligence for Accelerated Clinical Diagnosis: A Comparative Study Using Cloud and On-Premise Infrastructure

Abstract

This study assesses the amalgamation of High-Performance Computing (HPC) architectures with deep learning models for expedited brain MRI analysis in clinical diagnostic environments. We conducted a systematic comparison of three computing configurations available to Latin American universities: cloud-based CPU instances (Linode G8 Dedicated), cloud-based GPU instances (Linode RTX4000 Ada), and an on-premise workstation (Dell Precision 7960 Tower with dual RTX 5000 Ada GPUs). We utilized a dataset of 200 annotated brain magnetic resonance imaging (MRI) scans for binary classification of neurological abnormalities (presence/absence of lesions ≥5mm, including white matter hyperintensities, tumors, and vascular malformations) as determined by consensus of two board-certified neuroradiologists to train ResNet-50 and Vision Transformer models, assessing training efficiency, inference delay, energy consumption, and cost-effectiveness. The results indicate that the Dell Precision workstation attained an 11.1× acceleration in training duration (12.8 versus 142.7 minutes) relative to CPU-only cloud instances, while inference latency was minimized to 8.7 ms per image.