A pre-trained convolutional neural network based method for thyroid nodule diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Thyroid nodules present heterogeneous appearances and vague boundaries in ultrasound images, complicating accurate diagnosis of malignancy.
• Distinguishing between benign and malignant thyroid nodules is challenging for clinicians, necessitating advanced diagnostic tools.

Purpose of the Study:

• To develop and validate a hybrid deep learning method for accurate thyroid nodule diagnosis.
• To enhance diagnostic performance by fusing features from multiple pre-trained convolutional neural networks (CNNs).

Main Methods:

• A hybrid diagnostic method was proposed, fusing two pre-trained CNNs with distinct architectures.
• Feature maps from convolutional filters, pooling, and normalization layers of the CNNs were combined.
• A softmax classifier utilized the fused feature maps for thyroid nodule classification.

Main Results:

• The proposed hybrid CNN method achieved a diagnostic accuracy of 83.02%±0.72% on a dataset of 15,000 ultrasound images.
• Fusion of the two CNN models resulted in significant performance improvement compared to individual networks.
• The method demonstrated accurate and effective diagnosis of thyroid nodules.

Conclusions:

• The hybrid CNN approach offers a promising tool for improving the accuracy of thyroid nodule diagnosis.
• The fusion strategy effectively leverages features from multiple networks, enhancing diagnostic capabilities.
• This method shows significant potential for clinical application in thyroid cancer screening and diagnosis.