Source-linked AI summary
COVID-19 Infection Localization and Severity Grading from Chest X-ray Images
Anas M. Tahir, Muhammad E. H. Chowdhury, Amith Khandakar, Tawsifur Rahman, Yazan Qiblawey, Uzair Khurshid, Serkan Kiranyaz, Nabil Ibtehaz, M Shohel Rahman, Somaya Al-Madeed, Khaled Hameed, Tahir Hamid, Sakib Mahmud, Maymouna Ezeddin
TL;DR
Existing COVID-19 CXR studies were constrained by small datasets and limited infection localization or severity grading. This paper builds a large benchmark with human-machine-generated lung masks and develops a unified segmentation-based approach, achieving strong lung segmentation, infection localization, and detection results. Its conclusion is that the system supports detecting, localizing, and quantifying COVID-19 infection from X-ray images, while the authors note limited performance in localization and severity grading among prior studies.
Problem
Prior COVID-19 deep-learning studies used limited CXR repositories and showed limited infection localization and severity grading.
Method
The study constructs a large benchmark with ground-truth lung masks and develops a unified approach for lung segmentation, infection localization, and infection quantification from CXRs.
Results
96.11% lung-segmentation IoU and 97.99% DSC were achieved, while infection localization reached 83.05% IoU and 88.21% DSC.
Takeaways & Limitations
The study reports a system that detects, localizes, and quantifies COVID-19 infection from X-ray images.
Takeaways & Limitations
Prior studies had limited performance in infection localization and severity grading of COVID-19 pneumonia.
Abstract
from arXiv · showhide
Coronavirus disease 2019 (COVID-19) has been the main agenda of the whole world, since it came into sight in December 2019 as it has significantly affected the world economy and healthcare system. Given the effects of COVID-19 on pulmonary tissues, chest radiographic imaging has become a necessity for screening and monitoring the disease. Numerous studies have proposed Deep Learning approaches for the automatic diagnosis of COVID-19. Although these methods achieved astonishing performance in detection, they have used limited chest X-ray (CXR) repositories for evaluation, usually with a few hundred COVID-19 CXR images only. Thus, such data scarcity prevents reliable evaluation with the potential of overfitting. In addition, most studies showed no or limited capability in infection localization and severity grading of COVID-19 pneumonia. In this study, we address this urgent need by proposing a systematic and unified approach for lung segmentation and COVID-19 localization with infection quantification from CXR images. To accomplish this, we have constructed the largest benchmark dataset with 33,920 CXR images, including 11,956 COVID-19 samples, where the annotation of ground-truth lung segmentation masks is performed on CXRs by a novel human-machine collaborative approach. An extensive set of experiments was performed using the state-of-the-art segmentation networks, U-Net, U-Net++, and Feature Pyramid Networks (FPN). The developed network, after an extensive iterative process, reached a superior performance for lung region segmentation with Intersection over Union (IoU) of 96.11% and Dice Similarity Coefficient (DSC) of 97.99%. Furthermore, COVID-19 infections of various shapes and types were reliably localized with 83.05% IoU and 88.21% DSC. Finally, the proposed approach has achieved an outstanding COVID-19 detection performance with both sensitivity and specificity values above 99%.
1 Introduction
COVID-19 CXR research is motivated by diagnostic ambiguity and limitations of existing imaging and datasets. The study therefore develops a unified, benchmarked system for lung segmentation, infection localization, quantification, and detection.
- Motivation: CT offers higher sensitivity than RT-PCR in some studies but is slower, costlier, less applicable, and limited for early COVID-19 cases.X-ray imaging is described as cheaper, faster, more available, and involving less radiation than CT.
- Motivation: COVID-19 symptoms and radiographic features overlap with other viral and inflammatory lung diseases, complicating diagnosis and potentially delaying treatment.Chest radiographs can support screening, but distinguishing COVID-19 from other pneumonia remains difficult.
- Research gap: Prior deep-learning studies reported high COVID-19 detection performance but commonly relied on datasets containing only a few hundred COVID-19 CXR images.The paper links this data scarcity to inadequate evaluation and difficulty generalizing results in practice.
- Research gap: Previous studies largely focused on detection or classification, with limited assessment of infection localization and severity grading.The paper identifies this limitation as restricting usability and robustness for clinical use.
- Contributions: COVID-QU contains 11,956 COVID-19, 11,263 Non-COVID, and 10,701 Normal CXR images, with ground-truth lung masks created through human-machine collaboration.The dataset and masks are intended as a public benchmark for more reliable evaluation of detection, localization, and quantification.
- Contributions: The proposed system combines lung segmentation with COVID-19 infection localization and quantification from CXR images.Experiments compare segmentation networks to identify suitable models for the accomplished tasks.
2 The Benchmark COVID-QU Dataset
The COVID-QU benchmark compiles diverse chest X-ray sources into a large, three-class dataset and uses collaborative human-machine annotation to generate lung masks at scale.
- Only posterior-to-anterior or anterior-to-posterior chest X-rays were considered because these views are widely used by radiologists.
- The dataset combines publicly available repositories with varying formats, resolutions, quality, and signal-to-noise levels after removing duplicates and severely low-quality or over-exposed images.
- COVID-19 images were collected from BIMCV-COVID19+, a German medical school, SIRM, Github, Kaggle, Twitter, and another COVID-19 repository.
- The human-machine approach addresses the exhaustive, subjective production of pixel-wise lung masks by iteratively training, evaluating, correcting, and selecting segmentation outputs.
3 Methods
The proposed system uses parallel encoder-decoder CNNs to segment lungs and COVID-19 infection regions, then uses the resulting masks for localization, quantification, and detection. It evaluates U-Net, U-Net++, and FPN architectures with multiple encoder backbones on CXR datasets.
- Pipeline: Two ED-CNNs process each CXR in parallel to generate binary lung and COVID-19 infection masks.The masks are superimposed on the CXR to localize and quantify infected lung regions.
- Network models: U-Net, U-Net++, and FPN models were tested with ResNet18, ResNet50, DenseNet121, DenseNet161, and InceptionV4 encoders.The networks use encoder-decoder structures for contextual representation and spatial localization, with architecture-specific decoder connections.
- Segmentation: Pixel-wise SoftMax outputs assign each pixel to background or lung for lung segmentation and background or lesion for infection segmentation.Predicted masks are thresholded at 0.5, while lung masks additionally undergo hole filling and removal of small regions; infection masks are constrained to lung areas.
- Detection and quantification: The infection mask supports COVID-19 classification by identifying whether predicted infection pixels occur within lung regions.At least one predicted infection pixel classifies an image as COVID-19 positive; infection extent is also computed overall and per lung.
- Data and evaluation: Lung segmentation uses COVID-QU, while infection segmentation and COVID-19 detection use 2,913 samples with infection masks from QaTa-Cov19.Images are resized to 256×256 pixels, with 80% training, 20% testing, and five-fold cross-validation using part of the training data for validation.
- Evaluation: Segmentation is evaluated with Accuracy, IoU, and DSC, while COVID-19 detection uses Accuracy, Precision, Sensitivity, F1-score, and Specificity.Segmentation metrics assess pixel-level foreground and background classification; detection metrics are computed per CXR sample.
4 Experimental Results
Experiments compare segmentation and detection configurations across unseen test sets. The results show strong lung and infection segmentation, high COVID-19 detection performance, and sub-100-ms system inference under the reported setup.
- Lung Segmentation: 96.11% IoU and 97.99% DSC were achieved by the FPN model with DenseNet121 for lung segmentation.The networks also segmented lung regions across COVID-19 and non-COVID-19 pneumonia cases with different severity levels.
- Infection Segmentation: 87.08% DSC for parallel segmentation exceeded 86.84% DSC for the cascaded configuration.The parallel scheme was therefore used for the remaining infection-segmentation experiments.
- Infection Segmentation: 83.05% IoU and 88.21% DSC were obtained by U-Net++ with a DenseNet121 encoder for infection segmentation.FPN with InceptionV4 reached 83.08% IoU and 88.13% DSC, while U-Net with ResNet18 led U-Net models with 82.92% IoU.
- Qualitative Evaluation: The top-three networks reliably segmented infections across small, medium, and large regions and mild, moderate, severe, and critical cases.Figure 5 also presents localization and severity grading for one 42-year-old female patient across the first three hospital days.
- COVID-19 Detection: 99.66% sensitivity was achieved by U-Net with DenseNet121 and FPN with ResNet18 for COVID-19 detection.All models exceeded 97% sensitivity and specificity; U-Net++ with ResNet18 achieved 100% specificity.
5 Conclusion
The study addresses limited CXR datasets and weak infection-localization and severity-grading capabilities by combining lung segmentation, infection localization, and quantification. Using the COVID-QU benchmark and collaborative annotation, the system achieved strong segmentation and detection results while retaining future scope for reducing model complexity.
- Dataset and contribution: 33,920 CXR images, including 11,956 COVID-19, 11,263 Non-COVID pneumonia, and 10,701 Normal images, comprise the COVID-QU benchmark dataset.Ground-truth lung masks were constructed using a collaborative human-machine approach intended to reduce annotation labor and subjectivity.
- Results: 96.11% IoU and 97.99% DSC were achieved for lung segmentation on COVID-QU.The result was reported after extensive experiments on the benchmark dataset.
- Results: 83.05% IoU and 88.21% DSC were achieved for localizing COVID-19 infections of various severity.The system localized infections with different shapes and severity levels.
- Results: Sensitivity and specificity values above 99% were achieved for COVID-19 detection.The study presents this as an unprecedented detection performance.
- Conclusion: The proposed system combines lung segmentation with COVID-19 infection detection, localization, and quantification from X-ray images.The authors state that this integrated use of lung and infection segmentation can assist severity diagnosis and progression follow-up.
- Future work: Future work targets robust quantization and model compression to reduce model complexity and accelerate inference.The authors plan to explore heterogeneous network models for this purpose.
Funding
The work received support from the Qatar University COVID19 Emergency Response Grant. Responsibility for the work and its claims remains with the authors.
- Funding: The Qatar University COVID19 Emergency Response Grant provided support for the work.
- Funding: The grant is identified as QUERG-CENG-2020-1.
- Responsibility: The authors state that responsibility for the work and claims made herein is solely theirs.