Source-linked AI summary
Review of Artificial Intelligence Techniques in Imaging Data Acquisition, Segmentation and Diagnosis for COVID-19
Feng Shi, Jun Wang, Jun Shi, Ziyan Wu, Qian Wang, Zhenyu Tang, Kelei He, Yinghuan Shi, Dinggang Shen
TL;DR
COVID-19 imaging requires effective tools because available imaging datasets are limited and RT-PCR-related diagnostic challenges persist. This review synthesizes AI methods across X-ray and CT acquisition, segmentation, diagnosis, and follow-up, reporting applications across the imaging pipeline and representative promising results. It also identifies small samples, possible overfitting, and insufficient data quality as boundaries requiring further dataset development.
Problem
COVID-19 X-ray and CT scans are not widely available, hindering AI research and development, while early disease may show negative radiological signs and RT-PCR can have high false-negative rates.
Method
The paper reviews AI-empowered X-ray and CT imaging across acquisition, segmentation, diagnosis, prognosis, follow-up, datasets, and open challenges.
Results
The review finds AI applied across the COVID-19 imaging pipeline, with reported examples including 92.9% detection accuracy for Bayesian VGG16 and 87.5% accuracy for CT-based severity assessment.
Takeaways & Limitations
AI-empowered X-ray and CT platforms support safer, accurate, and efficient COVID-19 imaging workflows and clinical assessment within the reviewed scope.
Takeaways & Limitations
Many current AI segmentation and diagnosis studies use small samples, which may lead to overfitting and limits clinical usefulness without improved data quality and quantity.
Abstract
from arXiv · showhide
(This paper was submitted as an invited paper to IEEE Reviews in Biomedical Engineering on April 6, 2020.) The pandemic of coronavirus disease 2019 (COVID-19) is spreading all over the world. Medical imaging such as X-ray and computed tomography (CT) plays an essential role in the global fight against COVID-19, whereas the recently emerging artificial intelligence (AI) technologies further strengthen the power of the imaging tools and help medical specialists. We hereby review the rapid responses in the community of medical imaging (empowered by AI) toward COVID-19. For example, AI-empowered image acquisition can significantly help automate the scanning procedure and also reshape the workflow with minimal contact to patients, providing the best protection to the imaging technicians. Also, AI can improve work efficiency by accurate delination of infections in X-ray and CT images, facilitating subsequent quantification. Moreover, the computer-aided platforms help radiologists make clinical decisions, i.e., for disease diagnosis, tracking, and prognosis. In this review paper, we thus cover the entire pipeline of medical imaging and analysis techniques involved with COVID-19, including image acquisition, segmentation, diagnosis, and follow-up. We particularly focus on the integration of AI with X-ray and CT, both of which are widely used in the frontline hospitals, in order to depict the latest progress of medical imaging and radiology fighting against COVID-19.
I. INTRODUCTION
The review examines AI-enhanced X-ray and CT imaging for COVID-19 across acquisition, segmentation, diagnosis, prognosis, and follow-up. It emphasizes safer, more efficient workflows amid diagnostic and occupational challenges.
- RT-PCR can be inadequate during severe outbreaks and has high false-negative rates, increasing the clinical value of accessible chest X-ray and thoracic CT.
- COVID-19 imaging-based diagnosis generally proceeds through pre-scan preparation, image acquisition, and disease diagnosis.CT acquisition includes a breath-hold scan, reconstruction, and transmission through PACS for subsequent reading.
- AI applications span dedicated imaging platforms, lung and infection segmentation, clinical assessment and diagnosis, prognosis, and related research.
- Contactless acquisition is important because conventional positioning requires technicians to work closely with patients, creating viral-exposure risks.
- AI visual sensors estimate scan range and other parameters from patient imagery, while 3D body reconstruction supports positioning and scanning decisions.Approaches include anatomical keypoints, parametric human meshes, and depth-based 3D patient-body inference.
C. Applications in COVID-19
COVID-19 imaging applications include contactless mobile-CT workflows and AI-assisted image analysis. The section also frames segmentation as a prerequisite for subsequent assessment while noting limited direct COVID-19 segmentation work.
- C. Applications in COVID-19: Mobile CT platforms and monitoring cameras enabled contactless imaging workflows with isolated scan and control rooms.The reviewed mobile platform combines AI-based pre-scan and diagnosis systems with separate entrances to reduce unnecessary interaction.
- C. Applications in COVID-19: AI-assisted patient monitoring and positioning can support remote technician operation during scanning.The workflow uses camera imagery, visual and audio prompts, motion analysis, and automated patient-positioning steps.
- Segmentation: Segmentation delineates lungs, lobes, bronchopulmonary segments, and infection regions for subsequent assessment and quantification.
- Segmentation: CT COVID-19 segmentation commonly uses U-Net, UNet++, and VB-Net, whereas X-ray segmentation is more challenging because ribs confound soft-tissue contrast.The passage states that no method had yet been developed specifically for segmenting COVID-19 X-ray images, while an Attention-U-Net was adopted for lung segmentation.
- Segmentation: Table I summarizes representative image-segmentation methods used in COVID-19 applications.
A. Segmentation of Lung Regions and Lesions
Segmentation methods target either lung regions or lung lesions. Lung-region segmentation is typically a prerequisite, while lesion localization remains challenging and may use attention mechanisms.
- COVID-19 segmentation methods are grouped into lung-region-oriented and lung-lesion-oriented approaches.
- Lung-region methods separate whole lungs or lobes from background regions and are considered a prerequisite for COVID-19 applications.
- Lesion or nodule detection is challenging, and attention mechanisms can provide an additional localization approach in screening.
B. Segmentation Methods
COVID-19 image segmentation delineates lungs and lesions for diagnosis, measurement, follow-up, and severity-related analysis. U-Net variants and human-guided or weakly supervised strategies address segmentation needs and limited annotations.
- Segmentation architectures: U-Net and its variants are widely used to segment lung regions and lesions in COVID-19 images.The reviewed variants include 3D U-Net, V-Net, VB-Net, and UNet++.
- Training challenges: Limited annotated data makes robust lesion segmentation difficult because manual delineation is labor-intensive and time-consuming.The literature addresses this limitation with radiologist interaction, diagnostic attention, pseudo-masks, and weakly supervised learning.
- Diagnostic applications: Segmentation outputs support COVID-19 diagnosis by supplying lung or lesion regions to downstream classification models.Examples include lung segmentation before classification and lesion highlighting for classification.
- Quantitative applications: Segmentation also enables quantitative analysis of infection progression, severity, and lesion distribution.Reported applications include infection volumes and ratios, longitudinal progression, ground-glass opacity assessment, and percentage of infection.
- Segmentation purpose: Segmentation delineates lungs, lobes, bronchopulmonary segments, and infected regions for subsequent assessment and quantification.These regions of interest support analysis of infection extent and distribution.
IV. AI-ASSISTED DIAGNOSIS
COVID-19 imaging diagnosis is challenging because chest CT contains hundreds of slices and the disease can resemble other pneumonias. Radiologists therefore need substantial time and experience to interpret these studies.
- Diagnostic burden: Chest CT diagnosis requires substantial specialist time because each study may contain hundreds of slices.X-ray and CT are widely used for fast acquisition, but CT volume increases the reading burden.
- Diagnostic ambiguity: COVID-19 can show manifestations similar to other pneumonias, making experienced radiological interpretation important.The passage frames diagnostic experience as necessary for distinguishing similar disease appearances.
A. X-ray based Screening of COVID-19
X-ray is commonly used as a first-line COVID-19 modality but is less sensitive than chest CT, particularly in early or mild disease. AI studies mainly classify COVID-19 against pneumonia or healthy subjects, while limited datasets constrain robustness and generalizability.
- Modality characteristics: X-ray is a typical first-line modality, but it is generally less sensitive than 3D chest CT for COVID-19.Chest radiographs were normal in early or mild disease in a reported study.
- Modality characteristics: Abnormal chest radiographs were reported in 69% of patients at admission and 80% later during hospitalization.These values describe detection at two different time points.
- AI screening studies: AI-based X-ray studies primarily classify COVID-19 against other pneumonia, healthy subjects, or non-COVID-19 cases.The reviewed studies use convolutional and residual-network models, including Bayesian CNNs, ResNet-based models, and COVID-Net.
- AI screening studies: COVID-Net achieved 83.5% testing accuracy on 5941 X-ray images spanning healthy, bacterial-pneumonia, viral-pneumonia, and COVID-19 cases.The dataset included 45 COVID-19 images, alongside substantially larger non-COVID-19 groups.
- Evidence limitations: Most reviewed studies contain only 70 COVID-19 images from two online datasets, limiting robustness evaluation and clinical generalizability.The passage also identifies unknown subject severity and early detection as unresolved concerns.
B. CT-based Screening and Severity Assessment of COVID-19
CT-based AI studies address COVID-19 classification, differentiation from other pneumonias, and severity assessment using segmentation, deep learning, and handcrafted-feature pipelines. Reported results are generally promising, while severity prediction remains important for treatment planning and ICU-related estimation.
- Pneumonia differentiation: CT models also distinguish COVID-19 from other pneumonias and healthy subjects despite similar radiological appearances.Reported approaches first segment candidate infection regions or use CNN-based classification.
- Representative results: 90% sensitivity, 96% specificity, and 0.96 AUC were reported by COVNet for identifying COVID-19 from community-acquired pneumonia and non-pneumonia.COVNet used shared-weight 2D-slice processing with max-pooling on a large chest CT dataset.
- Severity assessment: 93.3% true positive rate, 74.5% true negative rate, and 87.5% accuracy were reported for CT-based COVID-19 severity assessment.The random-forest model used infection volumes and ratios from anatomically segmented lung regions.
- Summary: CT-based COVID-19 diagnosis has produced generally promising results, while screening and severity prediction warrant further investigation.Severity prediction is linked in the passage to ICU-event estimation and treatment planning.
V. AI IN FOLLOW-UP STUDIES
AI-empowered COVID-19 follow-up aims to evaluate patients’ treatment responses and investigate potential problems, but remains challenging and sparsely studied. Existing work demonstrates data-driven tracking of infection changes to support subsequent clinical decisions.
- Follow-up evaluates patients’ responses and investigates potential problems after clinical treatment.
- AI-empowered follow-up is challenging because COVID-19 has a long incubation period and high infectivity.
- Most current studies address pre-diagnosis, while COVID-19 follow-up research remains very limited.
- A machine learning and visualization approach tracks changes in infection-region volume, density, and other clinical factors, then automatically generates clinical reports.
- Follow-up remains an open issue, with segmentation, diagnosis, quantification, and assessment methods proposed as possible guides for future development.
VI. PUBLIC IMAGING DATASETS FOR COVID-19
COVID-19 imaging research is hindered by the limited availability of X-ray and CT data. Several public collections provide early examples of COVID-19 images from clinical and literature-based sources.
- COVID-19 X-ray and CT scans are not widely available, greatly hindering AI research and development.
- Cohen et al.’s COVID-19 Image Data Collection contains 123 frontal-view X-rays.
- The COVID-CT dataset contains 288 CT slices from confirmed cases collected from more than 700 COVID-19 preprints.
- The Coronacases Initiative shares confirmed COVID-19 cases through its website.
VII. DISCUSSION AND FUTURE WORK
AI has been applied across COVID-19 imaging, while future work must address data quality, radiation and scanning workflows, labeling constraints, multicenter validation, and limited follow-up studies. The review identifies practical directions for extending these applications.
- AI has been applied across the entire COVID-19 imaging-based diagnosis pipeline, but substantial future work remains.
- AI-empowered acquisition can improve efficiency and protect medical staff, with future applications targeting scan quality and reduced patient radiation dosage.
- Small samples and negative early-stage radiological signs limit clinical usefulness, motivating larger datasets of clinically collected X-ray and CT images.
- Incomplete, inexact, or inaccurate labels and costly manual annotation motivate weakly supervised, self-supervised, and transfer-learning methods.
- The review proposes borrowing prognosis methods, combining in-hospital and out-of-hospital tracking, and pursuing multidisciplinary integration for COVID-19 follow-up.
VIII. CONCLUSION
The paper reviews how AI supports safe, accurate, and efficient COVID-19 imaging across the full application pipeline, focusing on X-ray and CT. It also emphasizes that imaging provides only partial patient information and should be combined with clinical and laboratory data.
- The review covers AI-empowered imaging platforms, clinical diagnosis, and pioneering research across the entire COVID-19 imaging pipeline.
- X-ray and CT are used to demonstrate the effectiveness of AI-empowered medical imaging for COVID-19.
- Imaging provides only partial information, so screening, detection, and diagnosis should also use clinical manifestations and laboratory examination results.
- AI is positioned to fuse imaging, clinical, and laboratory data for diagnosis, analysis, and follow-up.