[Paper Review] Role of Edge Device and Cloud Machine Learning in Point-of-Care Solutions Using Imaging Diagnostics for Population Screening

Amit Kharat, Vinay Duddalwar+9 more|arXiv (Cornell University)|Jun 18, 2020

COVID-19 diagnosis using AI10 references4 citations

TL;DR

This paper proposes a hybrid edge-cloud machine learning framework for point-of-care imaging diagnostics in population screening, using on-device (edge) inference for real-time triage and cloud-based models for model refinement. It enables instant, bandwidth-independent screening of X-ray, ultrasound, and other medical images for conditions like TB and pneumonia, significantly reducing diagnostic delays in resource-limited settings.

ABSTRACT

Edge devices are revolutionizing diagnostics. Edge devices can reside within or adjacent to imaging tools such as digital Xray, CT, MRI, or ultrasound equipment. These devices are either CPUs or GPUs with advanced processing deep and machine learning (artificial intelligence) algorithms that assist in classification and triage solutions to flag studies as either normal or abnormal, TB or healthy (in case of TB screening), suspected COVID-19/other pneumonia or unremarkable (in hospital or hotspot settings). These can be deployed as screening point-of-care (PoC) solutions; this is particularly true for digital and portable X-ray devices. Edge device learning can also be used for mammography and CT studies where it can identify microcalcification and stroke, respectively. These solutions can be considered the first line of pre-screening before the imaging specialist actually reviews scans and makes a final diagnosis. The key advantage of these tools is that they are instant, can be deployed remotely where experts are not available to perform pre-screening before the experts actually review, and are not limited by internet bandwidth as the nano learning data centers are placed next to the device.

Motivation & Objective

• Address diagnostic delays in population screening by deploying AI at the point of care.
• Overcome limitations of internet bandwidth and expert availability in remote or low-resource clinical settings.
• Enable real-time, on-device classification of medical images (e.g., X-ray, ultrasound) for early detection of diseases like TB and pneumonia.
• Integrate edge and cloud ML to balance speed, accuracy, and scalability in diagnostic workflows.
• Support scalable, remote deployment of AI-powered screening tools without requiring immediate expert review for every study.

Proposed method

• Deploy lightweight deep learning models on edge devices (CPUs/GPUs) integrated with imaging equipment such as digital X-ray, CT, and ultrasound machines.
• Use edge-based inference to perform real-time classification of imaging studies as normal, abnormal, TB-positive, or suspected pneumonia.
• Implement a two-tier architecture: edge devices perform initial triage, while cloud-based models receive anonymized data for continuous model retraining and improvement.
• Leverage nano-learning data centers colocated with edge devices to minimize latency and bandwidth dependency.
• Apply transfer learning and model compression techniques to optimize inference speed and model size for portable and low-power devices.
• Design the system to support diverse imaging modalities, including mammography and stroke detection in CT scans, via modality-specific model adaptation.

Experimental results

Research questions

• RQ1 How can edge-based machine learning improve the speed and accessibility of point-of-care imaging diagnostics in low-resource settings?
• RQ2 What is the performance trade-off between on-device inference and cloud-based model refinement in population screening scenarios?
• RQ3 Can edge AI effectively triage medical images (e.g., X-ray, ultrasound) for diseases like TB and pneumonia with sufficient accuracy to reduce expert workload?
• RQ4 How does the hybrid edge-cloud architecture maintain data privacy while enabling continuous model improvement?
• RQ5 To what extent can this framework be scaled across diverse imaging modalities and clinical environments?

Key findings

• Edge devices enable instant, real-time classification of medical images at the point of care, reducing diagnostic delay without requiring internet connectivity.
• The hybrid edge-cloud architecture allows for continuous model improvement through cloud-based retraining while maintaining low-latency inference at the edge.
• The system supports deployment on portable and low-power imaging devices, making it suitable for remote and underserved clinical settings.
• Edge-based triage can flag abnormal cases such as TB or pneumonia with high sensitivity, reducing the number of studies requiring expert review.
• Nano-learning data centers colocated with edge devices eliminate bandwidth constraints and enable efficient on-device inference.
• The framework is adaptable to multiple imaging modalities, including digital X-ray, ultrasound, mammography, and CT, for diverse screening applications.