PDAtt-Unet: Pyramid Dual-Decoder Attention Unet for Covid-19 infection segmentation from CT-scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical imaging, particularly CT scans, is crucial for diagnosing and quantifying COVID-19 infection.
• Accurate segmentation of COVID-19 infection from CT scans is essential for effective disease management.
• Existing deep learning architectures like Att-Unet have limitations in precise infection segmentation.

Purpose of the Study:

• To develop advanced deep learning architectures for improved COVID-19 infection segmentation in CT scans.
• To enhance the Att-Unet architecture by introducing PAtt-Unet and DAtt-Unet for better spatial awareness and guided segmentation.
• To propose a hybrid loss function to address challenges with blurry boundaries in infection segmentation.

Main Methods:

• Proposed PAtt-Unet architecture to leverage input pyramids for preserving spatial information across encoder layers.
• Developed DAtt-Unet architecture to guide COVID-19 infection segmentation within lung lobes.
• Introduced a combined PDAtt-Unet architecture and a hybrid loss function for enhanced segmentation accuracy.
• Evaluated proposed models on four datasets using intra- and cross-dataset scenarios, comparing against baseline and state-of-the-art methods.

Main Results:

• PAtt-Unet and DAtt-Unet demonstrated improved performance over the standard Att-Unet for COVID-19 infection segmentation.
• The combined PDAtt-Unet architecture further enhanced segmentation accuracy.
• The proposed PDEAtt-Unet, utilizing the hybrid loss function, outperformed all other tested methods, including Unet, Unet++, Att-Unet, InfNet, SCOATNet, and nCoVSegNet.
• PDEAtt-Unet successfully addressed segmentation challenges across diverse datasets and evaluation scenarios.

Conclusions:

• The novel PAtt-Unet, DAtt-Unet, and PDAtt-Unet architectures significantly advance the capabilities of deep learning for COVID-19 infection segmentation.
• The hybrid loss function is effective in overcoming blurry boundary segmentation issues.
• PDEAtt-Unet represents a state-of-the-art solution for accurate and robust COVID-19 segmentation in medical imaging.