[Paper Review] COVID-MobileXpert: On-Device COVID-19 Screening using Snapshots of Chest X-Ray

Xin Li, Chengyin Li+1 more|arXiv (Cornell University)|Apr 13, 2020

COVID-19 diagnosis using AI50 references59 citations

TL;DR

COVID-MobileXpert is a lightweight deep neural network (DNN)-based mobile app that enables on-device, rapid, and accurate COVID-19 screening using noisy snapshots of chest X-rays (CXR). It employs a novel three-player knowledge transfer and distillation framework—featuring a pre-trained attending physician network, a fine-tuned resident fellow network, and a lightweight medical student network—to extract robust features and achieve high accuracy even with low-quality images, enabling deployment in mobile and primary care settings.

ABSTRACT

With the increasing demand for millions of COVID-19 screenings, Computed Tomography (CT) based test has emerged as a promising alternative to the gold standard RT-PCR test. However, it is primarily provided in hospital setting due to the need for expensive equipment and experienced radiologists. An accurate, rapid yet inexpensive test that is suitable for COVID-19 population screenings at mobile, urgent and primary care clinics is urgently needed. We present COVID-MobileXpert: a lightweight deep neural network (DNN) based mobile app that can use noisy snapshots of chest X-ray (CXR) for point-of-care COVID-19 screening. We design and implement a novel three-player knowledge transfer and distillation (KTD) framework including a pre-trained attending physician (AP) network that extracts CXR imaging features from large scale of lung disease CXR images, a fine-tuned resident fellow (RF) network that learns the essential CXR imaging features to discriminate COVID-19 from pneumonia and/or normal cases using a small amount of COVID-19 cases, and a trained lightweight medical student (MS) network that performs on-device COVID-19 screening. To accommodate the need for screening using noisy snapshots of CXR images, we employ novel loss functions and training schemes for the MS network to learn the robust imaging features for accurate on-device COVID-19 screening. We demonstrate the strong potential of COVID-MobileXpert for rapid deployment via extensive experiments with diverse MS network architecture, CXR imaging quality, and tuning parameter settings. The source code of cloud and mobile based models are available from this https URL.

Motivation & Objective

• To address the urgent need for rapid, low-cost, and scalable COVID-19 screening solutions outside hospital settings.
• To develop a mobile-deployable model that maintains high accuracy despite using noisy, low-quality snapshots of chest X-rays.
• To enable point-of-care screening in urgent care, primary clinics, and mobile units using on-device inference.
• To overcome the limitations of CT and RT-PCR by creating a lightweight, efficient, and robust deep learning model for CXR-based diagnosis.
• To design a knowledge transfer and distillation framework that leverages expert-level imaging knowledge for small, efficient models suitable for mobile deployment.

Proposed method

• A three-player knowledge transfer and distillation (KTD) framework is proposed: an attending physician (AP) network pre-trained on large-scale lung disease CXR data to extract high-level features.
• A resident fellow (RF) network is fine-tuned on a small dataset of COVID-19 cases to learn discriminative features for distinguishing COVID-19 from pneumonia and normal cases.
• A lightweight medical student (MS) network is trained via knowledge distillation from the RF network to perform on-device inference with minimal computational cost.
• Novel loss functions and training schemes are designed for the MS network to enhance robustness to image noise and quality degradation in real-world CXR snapshots.
• The MS network is optimized for mobile deployment through model compression and quantization techniques to ensure real-time inference on edge devices.
• The entire pipeline is evaluated across diverse CXR quality levels, network architectures, and hyperparameter settings to validate robustness and generalization.

Experimental results

Research questions

• RQ1 Can a lightweight deep learning model achieve high diagnostic accuracy for COVID-19 using only noisy, low-quality snapshots of chest X-rays in mobile settings?
• RQ2 How effective is the proposed three-player knowledge transfer and distillation framework in transferring expert diagnostic knowledge to a compact, on-device model?
• RQ3 To what extent can the MS network generalize across varying image quality and clinical conditions without requiring extensive labeled data?
• RQ4 How does the novel loss function design improve robustness to image noise and artifacts in real-world CXR snapshots?
• RQ5 What is the performance ceiling of on-device inference using this framework in terms of accuracy, speed, and model size?

Key findings

• The proposed three-player KTD framework enables the lightweight MS network to achieve high diagnostic accuracy despite limited training data and noisy input images.
• The MS network demonstrates strong robustness to image quality degradation, maintaining high performance across diverse snapshot qualities.
• Extensive ablation studies confirm that the novel loss functions significantly improve feature learning and generalization in low-quality imaging conditions.
• The model achieves competitive performance compared to larger, more complex models, validating its suitability for on-device deployment.
• The source code for both cloud and mobile models is publicly available, enabling reproducibility and real-world integration in clinical settings.
• The framework enables rapid deployment of accurate, on-device screening tools in primary and mobile care clinics, reducing reliance on CT and PCR testing.