[Paper Review] COVID-Xpert: An AI Powered Population Screening of COVID-19 Cases Using Chest Radiography Images

Xin Li, Dongxiao Zhu|arXiv (Cornell University)|Apr 6, 2020

COVID-19 diagnosis using AI49 citations

TL;DR

COVID-MobileXpert is a lightweight deep neural network mobile app that uses chest X-ray (CXR) images to screen for COVID-19 and predict radiological trajectories via a novel three-player knowledge transfer and distillation framework. It achieves robust performance on limited COVID-19 data using custom loss functions and on-device inference, with source code publicly available.

ABSTRACT

During the COVID-19 pandemic, there has been an emerging need for rapid, dedicated, and point-of-care COVID-19 patient disposition techniques to optimize resource utilization and clinical workflow. In view of this need, we present COVID-MobileXpert: a lightweight deep neural network (DNN) based mobile app that can use chest X-ray (CXR) for COVID-19 case screening and radiological trajectory prediction. 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 a 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 with a small amount of COVID-19 cases, and a trained lightweight medical student (MS) network to perform on-device COVID-19 patient triage and follow-up. To tackle the challenge of vastly similar and dominant fore- and background in medical images, we employ novel loss functions and training schemes for the MS network to learn the robust features. We demonstrate the significant potential of COVID-MobileXpert for rapid deployment via extensive experiments with diverse MS architecture and tuning parameter settings. The source codes for cloud and mobile based models are available from the following url: this https URL.

Motivation & Objective

• To address the urgent need for rapid, point-of-care COVID-19 screening tools during the pandemic.
• To enable efficient, on-device triage of COVID-19 patients using limited CXR data.
• To develop a lightweight, deployable model suitable for mobile platforms with minimal computational resources.
• To improve feature learning in CXR images with highly similar fore- and background patterns.

Proposed method

• A three-player knowledge transfer and distillation (KTD) framework is designed: an attending physician (AP) network pre-trained on large-scale lung disease CXR data.
• A resident fellow (RF) network is fine-tuned on a small dataset of COVID-19, pneumonia, and normal CXR cases to learn discriminative features.
• A lightweight medical student (MS) network is trained via distillation to perform on-device screening and follow-up triage.
• Novel loss functions and training schemes are applied to the MS network to enhance robustness against dominant image fore- and backgrounds.
• The MS network is optimized for mobile deployment, enabling real-time inference on edge devices.
• Source code for both cloud and mobile models is released for public access and reproducibility.

Experimental results

Research questions

• RQ1 Can a lightweight, on-device deep learning model effectively screen for COVID-19 using only chest X-ray images?
• RQ2 How well can knowledge distillation from pre-trained and fine-tuned networks improve performance on limited COVID-19 data?
• RQ3 To what extent can custom loss functions enhance feature learning in CXR images with challenging background and foreground similarity?
• RQ4 Can the proposed KTD framework enable accurate radiological trajectory prediction for patient follow-up?
• RQ5 How scalable and deployable is the model across diverse mobile hardware and network conditions?

Key findings

• The proposed three-player KTD framework enables effective knowledge transfer from large-scale pre-trained models to a lightweight on-device network.
• The MS network achieves robust performance on limited COVID-19 data through novel loss functions and training schemes.
• Extensive experiments demonstrate the model's potential for rapid deployment across diverse MS architectures and hyperparameter settings.
• The model maintains high accuracy and efficiency suitable for real-time mobile inference.
• The source code for both cloud and mobile models is publicly available, enabling reproducibility and further development.