Optimization of deep learning models for inference in low resource environments.

Area of Science:

• Medical Artificial Intelligence (AI)
• Deep Learning (DL) in Healthcare
• Computational Imaging

Background:

• Deep learning (DL) models hold significant potential for revolutionizing healthcare applications.
• Clinical translation of AI is often limited by substantial hardware requirements and computational costs.
• Need for efficient AI solutions applicable in diverse healthcare environments, including low-resource settings.

Purpose of the Study:

• To assess the effectiveness of optimization techniques for DL models in various healthcare AI workloads.
• To evaluate the impact of these optimizations on model performance across different hardware configurations.
• To determine if optimization improves AI model efficiency without sacrificing accuracy for clinical use.

Main Methods:

• Evaluated optimization techniques on DL models for segmentation and classification tasks.
• AI workloads included brain extraction (MRI), colorectal cancer delineation (histopathology), and diabetic foot ulcer classification (RGB imaging).
• Quantitatively assessed model runtime (speedup, latency, memory usage) and model utility on unseen data across hardware setups.

Main Results:

• Optimization techniques significantly improved model runtime, including reduced latency and memory usage.
• These improvements were achieved without compromising the model's utility or accuracy on unseen data.
• Demonstrated substantial speedups in inference times across diverse AI healthcare applications.

Conclusions:

• Optimization techniques are effective in enhancing the performance of DL models for healthcare applications.
• These methods can facilitate the clinical translation of AI, particularly in low-resource environments.
• Optimized AI models are more practical for real-world healthcare applications, increasing accessibility in underserved regions.