COVID-SegNet: encoder-decoder-based architecture for COVID-19 lesion segmentation in chest X-ray.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• The COVID-19 pandemic has overwhelmed healthcare systems, necessitating advanced diagnostic tools.
• Current computer-aided diagnosis models for COVID-19 often lack precise localization of infected areas in chest X-rays.
• Accurate segmentation of lung lesions is crucial for effective medical diagnosis and treatment planning.

Purpose of the Study:

• To propose a novel UNet-based encoder-decoder architecture for precise COVID-19 lesion segmentation in chest X-rays.
• To enhance the segmentation model's performance using an attention mechanism and an atrous spatial pyramid pooling module.
• To evaluate the proposed model's effectiveness against state-of-the-art methods for COVID-19 detection.

Main Methods:

• Development of a UNet-based encoder-decoder network incorporating an attention mechanism.
• Integration of a convolution-based atrous spatial pyramid pooling module to refine feature extraction.
• Performance evaluation using Dice Similarity Coefficient and Jaccard Index metrics on chest X-ray datasets.

Main Results:

• The proposed model achieved a Dice Similarity Coefficient of 0.8325 and a Jaccard Index of 0.7132.
• The model demonstrated superior performance compared to the standard UNet model in COVID-19 lesion segmentation.
• An ablation study confirmed the significant contributions of the attention mechanism and specific dilation rates in the atrous spatial pyramid pooling module.

Conclusions:

• The developed UNet-based model effectively segments COVID-19 lesions in chest X-rays, offering improved diagnostic precision.
• The integration of attention mechanisms and atrous spatial pyramid pooling enhances segmentation accuracy.
• This approach provides a valuable tool for medical experts in diagnosing and managing COVID-19.