S 3 TU-Net: Structured convolution and superpixel transformer for lung nodule segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of lung adenocarcinoma nodules in CT images is crucial for clinical staging and diagnosis.
• Irregular nodule shapes and ambiguous boundaries present challenges for current segmentation methods.

Purpose of the Study:

• To introduce S3TU-Net, a hybrid CNN-Transformer architecture for enhanced lung nodule segmentation.
• To improve feature extraction, fusion, and global context modeling for more accurate nodule detection.

Main Methods:

• Developed S3TU-Net, integrating structured convolution blocks (DWF-Conv/D2BR-Conv) for multi-scale feature extraction and overfitting mitigation.
• Incorporated S2-MLP Link for enhanced multi-level feature fusion via spatial-shift skip connections.
• Utilized a residual-based superpixel vision transformer (RM-SViT) for efficient long-range dependency capture.

Main Results:

• S3TU-Net achieved a Dice score of 89.04%, precision of 90.73%, and IoU of 90.70% on the LIDC-IDRI dataset.
• The model outperformed recent methods by 4.52% in Dice score.
• Validation on the EPDB dataset demonstrated generalizability with a Dice score of 86.40%.

Conclusions:

• S3TU-Net effectively bridges local feature sensitivity and global context awareness for lung nodule segmentation.
• The hybrid architecture offers a robust tool for clinical decision support in lung cancer diagnosis.
• This approach advances the accuracy and reliability of automated nodule segmentation in CT imaging.