Optimizing AI for White Blood Cell Analysis: A Multi-Objective Neural Architecture Search Approach

Abstract

White Blood Cells (WBC) contain vital elements for identifying several health complications such as infections, immunity problems and even cancer. Most of the existing traditional Artificial Intelligence (AI) architectures for WBC detection and classification problems rely mainly on model accuracy, while other important aspects such as model size, time for inference, and energy consumption are critical in restricted use cases like mobile devices and edge computing platforms. To this end, this paper proposes a new framework called Multi-Objective Neural Architecture Search (MO-NAS) that aims to search for architectures that optimize for multiple objectives at once – accuracy, model size, inference time, and energy consumption. The performance of the proposed MO-NAS framework was tested using 8,500 AI images acquired from Hiwa Hospital with an overall classification accuracy of 96.4% within five WBC subtypes. Thereby also, the model achieved a faster inference time of 145ms per image, a smaller model size of 32.4 MB from 45MB, and a 25% energy advantage. It provides a path to scientific, high-quality AI models in WBC diagnostics and other medical imaging applications, towards improving individualized medicine and real-time clinical usage.