OCCMNet: Occlusion-Aware Class Characteristic Mining Network for multi-class artifacts detection in endoscopy.

Area of Science:

• Medical imaging analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Multi-class artifact detection in endoscopic imaging is vital for accurate diagnosis.
• Challenges include data imbalance, inter-class similarity, and occluded artifacts.
• Existing methods struggle with the complexity of simultaneous multi-class artifact identification.

Purpose of the Study:

• To develop an advanced deep learning network for simultaneous detection of eight classes of endoscopic artifacts.
• To address challenges of data imbalance, artifact similarity, and occlusion in endoscopic images.
• To improve the accuracy and reliability of automated endoscopic artifact detection.

Main Methods:

• Proposed the Occlusion-Aware Class Characteristic Mining Network (OCCMNet).
• Incorporated a Dual-Branch Class Rebalancing Module (DCRM) for data distribution balancing.
• Utilized a Class Discrimination Enhancement Module (CDEM) to improve inter-class distinction.
• Implemented a Global Occlusion-Aware Module (GOAM) to handle occluded artifacts by inferring obscured regions.

Main Results:

• OCCMNet demonstrated superior performance on the public EndoCV2020 dataset.
• Achieved a 3.5-6.5% improvement in mAP50 compared to state-of-the-art methods.
• Effectively handled challenges of data imbalance, similarity, and occlusion in multi-class artifact detection.

Conclusions:

• OCCMNet significantly enhances multi-class endoscopic artifact detection accuracy.
• The proposed network shows great potential in reducing clinical interference and diagnostic errors.
• This approach advances the application of computer vision in endoscopic diagnostics.