Thyroid intelligent diagnosis based on THMSNet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Endocrinology

Background:

• Thyroid disease diagnosis relies on differentiating benign and malignant nodules.
• Traditional methods like ultrasound and TI-RADS have limitations including interobserver variability and time constraints.
• Deep learning models like CNNs and transformers show promise but struggle with multiscale features and global dependencies.

Purpose of the Study:

• To develop an advanced deep learning model for accurate thyroid nodule diagnosis.
• To overcome limitations of existing diagnostic methods and deep learning approaches.
• To improve clinical decision-making in thyroid disease management.

Main Methods:

• Proposed THMSNet, a hybrid architecture integrating pyramid structure for multiscale features and Mamba for long-range dependencies.
• Incorporated Serial Channel-Spatial Attention Module (SCSAM) for enhanced feature representation.
• Utilized Truth-Value Calibration (TVC) algorithm to align predictions with pathological standards.
• Evaluated on 7,288 thyroid ultrasound images (3,282 benign, 4,006 malignant).

Main Results:

• THMSNet achieved 91.15% accuracy, 93.28% recall, and 96.92% AUROC.
• Outperformed ResNet (86.03% accuracy) and DenseNet (95.50% AUROC).
• Ablation studies demonstrated significant performance gains from each module: pyramid architecture (+7.83% accuracy), Mamba (+2.99%), SCSAM (+6.94%), and TVC (+6.94%).

Conclusions:

• THMSNet offers a robust and clinically applicable solution for thyroid nodule diagnosis.
• The model combines advanced feature extraction, attention mechanisms, and probability calibration for high accuracy.
• THMSNet serves as a valuable tool for assisting radiologists in clinical practice.