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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• The COVID-19 pandemic created an urgent need for rapid diagnostic tools.
• Automated detection of lung infections from CT scans can support traditional healthcare strategies.
• Challenges in COVID-19 CT segmentation include varied infection appearance and low contrast, hindering deep learning model training due to data scarcity.

Purpose of the Study:

• To propose a novel deep learning network, Inf-Net, for automated segmentation of COVID-19 infected regions in CT images.
• To develop a semi-supervised learning framework to overcome limitations of insufficient labeled data for training deep models.

Main Methods:

• Inf-Net utilizes a parallel partial decoder for global feature aggregation and a novel attention mechanism (implicit reverse attention and explicit edge-attention) to enhance boundary representation.
• A semi-supervised segmentation framework employing a random propagation strategy was developed, requiring minimal labeled data and leveraging abundant unlabeled data.

Main Results:

• The proposed Inf-Net achieved state-of-the-art performance in segmenting COVID-19 lung infections on COVID-SemiSeg and real CT datasets.
• The semi-supervised framework demonstrated improved learning ability and higher segmentation performance compared to traditional methods.
• Inf-Net outperformed existing cutting-edge segmentation models.

Conclusions:

• Inf-Net provides an effective solution for automated COVID-19 lung infection segmentation from CT scans.
• The semi-supervised approach successfully addresses the challenge of limited labeled data in medical image analysis.
• This work advances the state-of-the-art in medical image segmentation for infectious disease detection.