Authors - M.Murugesen, Priyanka P Abstract - Deep learning–based medical image models have achieved expert level performance in GPU-based research environments [1–3]. However, relia ble deployment in real clinical systems remains challenging due to constraints related to power consumption, hardware stability, and long-term operation. While prior studies have focused on improving model architectures or hardware accelerators [4,5], relatively limited attention has been devoted to systematical ly managing the transition from GPU-based development to NPU-based de ployment environments. This study formulates the GPU-to-NPU transition as an independent deployment research problem. Rather than proposing a new model architecture, we focus on preserving functional equivalence when trans ferring a validated GPU-trained medical vision model to an NPU-based infer ence environment. The proposed framework consists of reference model fixa tion, intermediate representation (IR)-based conversion [13–15], operator com patibility management, inference pipeline alignment, and output-level function al equivalence validation. The framework is evaluated through deployment of a ResNet-50–based pa thology classification model on a commercial ATOM NPU platform. Experi mental results demonstrate a 99.1% agreement rate (991/1,000 samples) be tween GPU-based and NPU-based inference outputs, confirming consistent de cision behavior despite architectural differences. These findings indicate that deployment reliability depends more on execution environment control and preprocessing alignment than on model architecture modification. By redefining deployment as a structured research problem, this work pro vides a reproducible methodology for translating research-grade medical AI models into energy-efficient NPU inference systems under practical clinical constraints.
Open Zoom