Optimizing Deep Learning Architectures for Mobile Medical Imaging Devices

EasyChair Preprint no. 13784

Abstract

The integration of deep learning (DL) into mobile medical imaging devices represents a transformative leap in healthcare delivery, enabling real-time diagnosis and analysis at the point of care. This research investigates the optimization of deep learning architectures specifically tailored for mobile medical imaging devices, addressing the unique challenges posed by their limited computational resources, energy constraints, and the need for high accuracy in medical diagnosis.

We begin by reviewing existing deep learning models used in medical imaging, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their hybrid variations. These models, while effective on high-performance computing systems, often exhibit reduced efficacy when deployed on mobile devices due to hardware limitations. To overcome these challenges, we propose a multi-faceted optimization strategy encompassing model compression, efficient architecture design, and hardware-software co-optimization.

Model compression techniques, including pruning, quantization, and knowledge distillation, are explored to reduce the model size and computational load without compromising diagnostic accuracy. We introduce novel pruning methods that leverage the inherent sparsity in medical images, significantly reducing the number of parameters while maintaining critical diagnostic features. Quantization techniques are refined to ensure minimal loss in model precision, and knowledge distillation is employed to transfer knowledge from complex models to lightweight architectures suitable for mobile deployment.

Efficient architecture design focuses on developing lightweight neural network architectures, such as MobileNets and SqueezeNets, specifically optimized for medical imaging tasks. We propose enhancements to these architectures,

Keyphrases: deep learning, Edge Computing, efficient architecture, Federated Learning, hardware-software co-optimization, mHealth., mobile medical imaging, Model Compression

@Booklet{EasyChair:13784,
  author = {Kayode Sheriffdeen},
  title = {Optimizing Deep Learning Architectures for Mobile Medical Imaging Devices},
  howpublished = {EasyChair Preprint no. 13784},

  year = {EasyChair, 2024}}