Inf-Net: Automatic COVID-19 Lung Infection Segmentation from CT Images (2004.14133v4)

Abstract: Coronavirus Disease 2019 (COVID-19) spread globally in early 2020, causing the world to face an existential health crisis. Automated detection of lung infections from computed tomography (CT) images offers a great potential to augment the traditional healthcare strategy for tackling COVID-19. However, segmenting infected regions from CT slices faces several challenges, including high variation in infection characteristics, and low intensity contrast between infections and normal tissues. Further, collecting a large amount of data is impractical within a short time period, inhibiting the training of a deep model. To address these challenges, a novel COVID-19 Lung Infection Segmentation Deep Network (Inf-Net) is proposed to automatically identify infected regions from chest CT slices. In our Inf-Net, a parallel partial decoder is used to aggregate the high-level features and generate a global map. Then, the implicit reverse attention and explicit edge-attention are utilized to model the boundaries and enhance the representations. Moreover, to alleviate the shortage of labeled data, we present a semi-supervised segmentation framework based on a randomly selected propagation strategy, which only requires a few labeled images and leverages primarily unlabeled data. Our semi-supervised framework can improve the learning ability and achieve a higher performance. Extensive experiments on our COVID-SemiSeg and real CT volumes demonstrate that the proposed Inf-Net outperforms most cutting-edge segmentation models and advances the state-of-the-art performance.

• The paper introduces Inf-Net, a novel deep learning architecture that uses parallel partial decoders and attention modules for precise COVID-19 lung segmentation.
• The model incorporates a semi-supervised framework to overcome scarce labeled data, achieving a Dice score of 0.739 and an MAE of 0.064.
• Inf-Net extends to multi-class infection labeling, enhancing diagnostic precision by differentiating infection types and improving treatment assessments.

Inf-Net: Automatic COVID-19 Lung Infection Segmentation from CT Images

The paper "Inf-Net: Automatic COVID-19 Lung Infection Segmentation from CT Images" addresses the pressing issue of accurately and efficiently segmenting lung infections in CT images due to COVID-19 using Deep Learning techniques. This challenge arises primarily from the high variability in infection presentations and the limited availability of annotated training data. To meet these challenges, the authors propose a novel architecture, Inf-Net, which leverages advanced features like a parallel partial decoder and attention mechanisms to enhance segmentation accuracy, combined with a semi-supervised learning strategy to mitigate the label scarcity problem.

Key Contributions and Methodology

1. Inf-Net Architecture: The Inf-Net model is designed to automatically identify infected regions from chest CT slices. It employs several innovative components:

- Parallel Partial Decoder (PPD): This module aggregates high-level features to generate a coarse global map that roughly localizes potential infection areas. - Reverse Attention (RA) Modules: These modules refine the initial coarse predictions by iteratively focusing on difficult regions, thereby enhancing the boundary detection and implicit edge-learning. - Edge Attention (EA) Module: This module explicitly models the boundary information to improve the final segmentation details.

1. Semi-Supervised Learning Framework: To combat the limited availability of labeled data, the authors introduce a semi-supervised framework that leverages unlabeled CT images. They utilize a pseudo-label generation strategy to progressively enhance the model's training dataset, ensuring more robust learning and better generalization.
2. Multi-Class Labeling Extension: Recognizing the clinical importance of distinguishing between different types of lung infections (e.g., ground-glass opacities (GGO) and consolidations), the paper extends Inf-Net for multi-class infection labeling. It integrates the initial infection segmentation into a multi-class segmentation network like FCN8s or U-Net to differentiate various infection types effectively.

Performance Evaluation

Quantitative evaluations demonstrate that Inf-Net, augmented with semi-supervised learning (Semi-Inf-Net), achieves significant improvements over existing models, such as U-Net, U-Net++, and various attention-based UNets. Specifically, Semi-Inf-Net achieves a Dice score of 0.739 and Mean Absolute Error (MAE) of 0.064 on the COVID-SemiSeg dataset, outperforming conventional counterparts by noteworthy margins.

Qualitative results exhibit the model’s robustness, yielding accurate segmentation even in complex infection presentations, further validated by tests on real CT volumes. The superiority of Inf-Net is prominently visible in its ability to delineate infection boundaries accurately, which competitive models fail to achieve consistently.

Practical and Theoretical Implications

The practical implications of Inf-Net are considerable:

• Enhanced Diagnostic Capability: By providing detailed and accurate segmentations of lung infections, Inf-Net enhances radiologists’ ability to assess disease severity and progression.
• Improved Efficiency: The automated segmentation reduces the manual effort required, which is especially valuable during pandemics where rapid and scalable diagnostic processes are critical.
• Quantitative Assessment: Inf-Net’s extension for multi-class infection labeling provides a nuanced understanding of the infection types, aiding in personalized treatment strategies and better patient outcomes.

Theoretically, Inf-Net introduces methodological advancements such as:

• Effective Feature Aggregation: The parallel partial decoder efficiently combines high-level features to generate a global context, highlighting the importance of context-aware predictions in medical image analysis.
• Attentive Refinement Mechanisms: The integration of RA and EA modules underlines the role of attention mechanisms in refining segmentation outputs, specifically in medical images with low-intensity contrasts.

Future Directions

Ongoing work may focus on integrating COVID-19 detection and infection segmentation into a unified diagnostic framework. This holistic approach would not only classify COVID-19 cases but also provide immediate infection region segmentation, paving the way for comprehensive diagnostic tools.

Additionally, leveraging Generative Adversarial Networks (GANs) or Conditional Variational Autoencoders (CVAE) for data augmentation could address the challenge of dataset scarcity even further. This would enhance model performance by synthesizing realistic training samples, thus enriching the training dataset without the need for extensive manual annotations.

Conclusion

The Inf-Net framework, with its innovative use of parallel partial decoders, reverse attention modules, and semi-supervised learning, presents a significant step forward in the automatic and accurate segmentation of COVID-19 lung infections from CT images. Its practical and theoretical contributions hold substantial promise for the rapid screening, quantitative assessment, and treatment evaluation in clinical settings, particularly amid global health crises like the COVID-19 pandemic. The proposed methodologies also offer valuable insights for broader applications in medical image segmentation and computer vision.

By openly providing the code and datasets, the authors facilitate further research and development in this critical domain, fostering advancements that may substantially improve global health responses to infectious diseases.