RAD-UNet: Research on an improved lung nodule semantic segmentation algorithm based on deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models struggle with segmenting small target pixels in CT images, leading to under- or oversegmentation of lung lesions.
• Existing convolutional neural network models face challenges in accurately extracting feature information for lung nodule segmentation in CT scans.

Purpose of the Study:

• To propose an improved convolutional neural network segmentation model, RAD-UNet, for enhanced lung nodule segmentation in CT images.
• To address the limitations of under- and oversegmentation in current deep learning models for CT image analysis.

Main Methods:

• The RAD-UNet model features a U-Net encoder-decoder architecture with a ResNet residual network replacing the original encoder.
• Incorporates an atrous spatial pyramid pooling module after the encoder and an improved decoder with a cross-fusion feature module utilizing channel and spatial attention.
• The model was evaluated on the LIDC dataset and a clinical CT dataset.

Main Results:

• RAD-UNet demonstrated superior lung lesion segmentation performance compared to SegNet and U-Net on both datasets.
• Achieved a mean Intersection over Union (mIoU) of 87.76% and 88.13%, and an F1-score of 93.56% and 93.72% respectively.
• The model effectively solved under- and oversegmentation problems, significantly improving segmentation accuracy.

Conclusions:

• The improved RAD-UNet method provides more accurate pixel-level segmentation for lung tumors in CT images.
• RAD-UNet outperforms existing models like SegNet and U-Net in identifying lung nodules.
• The study successfully enhanced image segmentation performance for CT scans, particularly for lung nodule detection.