LGF-Net: A multi-scale feature fusion network for thyroid nodule ultrasound image classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology Diagnostics

Background:

• Accurate thyroid nodule classification is vital for computer-aided diagnosis of thyroid cancer.
• Existing convolutional neural network (CNN) and transformer models face challenges in integrating local and global features, limiting recognition accuracy.

Purpose of the Study:

• To develop a model that simultaneously captures fine-grained local and global spatial features for thyroid nodule diagnosis.
• To improve recognition accuracy and generalization ability by adapting to irregular nodule morphology and focusing on key pixel regions.

Main Methods:

• Proposed a multi-scale fusion model, the local and global feature fusion network (LGF-Net), with dual CNN and Transformer branches.
• CNN branch uses wavelet transform and deformable convolution (WTDCM) for local features; Transformer branch uses aggregated attention (AA) for spatial features.
• Adaptive feature fusion module (FFM) integrates multi-scale features for enhanced classification.

Main Results:

• Achieved 81.50% accuracy on the public TNCD dataset and 91.24% on a private clinical dataset.
• Outperformed state-of-the-art methods in accuracy, recall, precision, and F1-score on both datasets.
• Ablation studies and visualization confirmed model component effectiveness and interpretability.

Conclusions:

• The LGF-Net method significantly improves thyroid nodule recognition accuracy and generalization.
• Demonstrates potential for clinical application in thyroid cancer diagnosis.
• Provides interpretability, aiding clinicians in diagnosis.