A lung nodule segmentation model based on the transformer with multiple thresholds and coordinate attention.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate lung nodule segmentation is critical for early lung cancer detection.
• Current deep learning models often struggle with extracting edge/semantic information and long-range dependencies.

Purpose of the Study:

• To develop an advanced lung nodule segmentation model, MCAT-Net, to overcome limitations in existing deep learning approaches.
• To enhance the extraction of edge, semantic, and spatial information for improved segmentation accuracy.

Main Methods:

• Proposed MCAT-Net incorporating a multi-threshold feature separation module for multi-level feature extraction.
• Integrated coordinate attention mechanism to improve spatial information utilization and sensitivity to nodule positions.
• Employed a transformer to capture long-range dependencies and enhance global information integration.

Main Results:

• MCAT-Net achieved a Dice Similarity Coefficient (DSC) of 88.29% on the LIDC-IDRI dataset and 78.51% on the LNDb dataset.
• Sensitivities reached 86.33% (LIDC-IDRI) and 75.05% (LNDb), demonstrating robust performance.
• The model effectively integrated multi-level features and long-range dependencies for superior segmentation.

Conclusions:

• MCAT-Net shows significant potential for accurate lung nodule segmentation, aiding in the early diagnosis of lung cancer.
• The proposed architectural components effectively address limitations in feature extraction and dependency modeling in deep learning segmentation.