From CNN to SNN: Leveraging Neural Processing Units for Skin Lesion Classification

Skin lesions represent an increasingly prevalent dermatological concern that calls for proactive screening among the general population. Despite growing efforts to address this issue, effective active screening requires the adoption of up-to-date technologies by healthcare professionals. Recently, the integration of low-cost imaging devices has enabled more accessible health monitoring, allowing skin lesions to be evaluated using affordable tools such as smartphones or non-medical grade cameras. The emergence of Neural Processing Units (NPUs), capable of executing AI-based models efficiently on-device, marks a significant step forward in the development of low-cost, portable medical devices. These devices can perform inference directly, leveraging the growing number of labelled dermatological datasets to support accurate and timely diagnosis in real-world settings. Current datasets are typically used in conjunction with Convolutional Neural Networks (CNNs), which have proven highly effective for image classification tasks such as skin lesion analysis. However, CNNs present limitations in terms of memory footprint and computational efficiency, particularly in edge or resource-constrained environments. To address these challenges, the conversion of CNNs into Spiking Neural Networks (SNNs) offers a promising alternative. This approach enables the deployment of models on NPUs, allowing for fast, energy-efficient, and secure on-device inference-capabilities that are especially important for real-time medical applications requiring low latency and power-aware processing. As a result, we propose an end-to-end pipeline for converting a CNN to an SNN, enabling the deployment of a skin lesion classification model on an NPU. This pipeline was successfully implemented using the Akida MetaTF framework.