DRR4Covid: Learning Automated COVID-19 Infection Segmentation from Digitally Reconstructed Radiographs (2008.11478v1)

Abstract: Automated infection measurement and COVID-19 diagnosis based on Chest X-ray (CXR) imaging is important for faster examination. We propose a novel approach, called DRR4Covid, to learn automated COVID-19 diagnosis and infection segmentation on CXRs from digitally reconstructed radiographs (DRRs). DRR4Covid comprises of an infection-aware DRR generator, a classification and/or segmentation network, and a domain adaptation module. The infection-aware DRR generator is able to produce DRRs with adjustable strength of radiological signs of COVID-19 infection, and generate pixel-level infection annotations that match the DRRs precisely. The domain adaptation module is introduced to reduce the domain discrepancy between DRRs and CXRs by training networks on unlabeled real CXRs and labeled DRRs together.We provide a simple but effective implementation of DRR4Covid by using a domain adaptation module based on Maximum Mean Discrepancy (MMD), and a FCN-based network with a classification header and a segmentation header. Extensive experiment results have confirmed the efficacy of our method; specifically, quantifying the performance by accuracy, AUC and F1-score, our network without using any annotations from CXRs has achieved a classification score of (0.954, 0.989, 0.953) and a segmentation score of (0.957, 0.981, 0.956) on a test set with 794 normal cases and 794 positive cases. Besides, we estimate the sensitive of X-ray images in detecting COVID-19 infection by adjusting the strength of radiological signs of COVID-19 infection in synthetic DRRs. The estimated detection limit of the proportion of infected voxels in the lungs is 19.43%, and the estimated lower bound of the contribution rate of infected voxels is 20.0% for significant radiological signs of COVID-19 infection. Our codes will be made publicly available at https://github.com/PengyiZhang/DRR4Covid.

• The paper introduces an infection-aware DRR generator that simulates X-ray images from 3D CT volumes for varied COVID-19 training data.
• It integrates a domain adaptation module using MMD to bridge the gap between synthetic DRRs and real chest X-rays, achieving superior segmentation metrics.
• The modular FCN framework enables both infection localization and quantification, paving the way for broader pulmonary disease diagnostics.

DRR4Covid: A Modular Approach to COVID-19 Infection Segmentation

Introduction

"DRR4Covid: Learning Automated COVID-19 Infection Segmentation from Digitally Reconstructed Radiographs" introduces a technique for COVID-19 infection segmentation on chest X-rays (CXRs) using digitally reconstructed radiographs (DRRs). The paper proposes a modular framework that enhances the utility of CXRs by bridging the gap between the photo-realistic synthetic DRRs and real-world CXRs using domain adaptation. The methodological innovations are primarily focused on the infection-aware DRR generator and domain adaptation techniques that facilitate accurate segmentation of COVID-19 infections without the requirement of annotated CXRs.

Methodology

Infection-Aware DRR Generation

The paper's central innovation lies in its infection-aware DRR generator, which simulates X-ray images from 3D CT volumes containing infection annotations. This generator allows the modulation of the radiological signs of COVID-19 infections, thus providing a varied dataset for effectively training segmentation algorithms. Unlike traditional methods that require precise, pixel-level annotations, the proposed generator automatically aligns pixel-level infection annotations precisely with the generated DRRs.

Domain Adaptation Strategy

Despite the photorealistic nature of synthetic DRRs, domain discrepancies with actual CXRs can impede model performance. To address this, DRR4Covid incorporates a domain adaptation module based on Maximum Mean Discrepancy (MMD). By training networks on both labeled synthetic DRRs and unlabeled real CXRs, the module minimizes domain shifts, enabling effective inference on real data.

Network Architecture

The implementation adopts a fully convolutional network (FCN) architecture, augmented with a classification and segmentation header. This dual-header design facilitates not only infection localization but also quantification, providing comprehensive infection measurement capabilities. The platform's modularity allows the integration of various off-the-shelf deep learning and domain adaptation techniques for expanded application scope.

Experimental Results

The methodology is empirically validated on a dataset comprising 794 COVID-19 positive and 794 normal cases, achieving formidable classification metrics: an accuracy of 0.954, an AUC of 0.989, and a F1-score of 0.953. The segmentation performance is similarly robust, with an accuracy of 0.957, an AUC of 0.981, and a F1-score of 0.956, surpassing traditional segmentation approaches. The research highlights the infection-aware DRR generator's ability to estimate the sensitivity of X-ray imaging, suggesting a detection limit in COVID-19-infected lung volume at 19.43% ± 16.29%.

Practical Implications and Limitations

The modular design of DRR4Covid indicates its adaptability to other infection types and lung lesions beyond COVID-19, labeling this broader scope as DRR4Lesion. While the paper demonstrates the potential in automated segmentation, it highlights the need for further research to acquire pixel-level annotations in CXRs for cross-validation.

Conclusion

DRR4Covid presents a novel methodology that revolutionizes COVID-19 infection segmentation through DRR-based training and sophisticated domain adaptation. It provides a pathway towards more precise and efficient infection diagnostics using readily available imaging technology, with implications for broader applications in medical imaging and AI-based diagnostic frameworks. Future projects could focus on further enhancing segmentation accuracy and broadening the algorithm's application to other pulmonary diseases.