CAM-Wnet: An effective solution for accurate pulmonary embolism segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pulmonary embolism (PE) is a significant cause of morbidity, necessitating early detection and treatment.
• Manual segmentation of PE from CT scans is time-consuming and prone to errors.
• Accurate PE segmentation is crucial for targeted treatment and improving patient survival rates.

Purpose of the Study:

• To develop an automatic and efficient deep neural network (CAM-Wnet) for PE segmentation from CT images.
• To address challenges in PE segmentation, including varied emboli sizes, shapes, and low tissue contrast.
• To incorporate coordinate attention (CA) mechanisms and pyramid pooling modules (PPMs) for enhanced segmentation.

Main Methods:

• A stacked U-Net architecture (CAM-Wnet) was proposed, comprising a coarse U-Net and a subdivision U-Net.
• The coarse U-Net utilized a pretrained VGG-19 encoder and CA residual blocks (CARBs) in its decoder.
• The subdivision U-Net refined segmentation using CARBs, with PPMs integrated between encoder-decoder layers for global context, and an improved focal loss function for training.

Main Results:

• The CAM-Wnet achieved high segmentation accuracy on the China-Japan Friendship Hospital dataset, with Precision (0.9703), Recall (0.963), IoU (0.9353), and F1-score (0.9665).
• Experimental results demonstrated the method's effectiveness in automatically and accurately segmenting PE in lung CT images.
• The model's performance on a liver tumor dataset confirmed its generalization ability.

Conclusions:

• The proposed CAM-Wnet effectively leverages multiscale pooling and attention mechanisms to capture global and semantic information.
• The method significantly improves PE segmentation in lung CT images, offering a valuable tool for clinical diagnosis.
• CAM-Wnet shows potential for assisting clinicians in the diagnosis and treatment planning of PE.