A multi-channel deep convolutional neural network for multi-classifying thyroid diseases.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid diseases and cancer incidence are rising globally.
• Current deep learning models primarily focus on binary classification of thyroid cancer.
• Limited research exists on multi-class thyroid disease classification and co-diagnosis.

Purpose of the Study:

• To develop a novel multi-channel convolutional neural network (CNN) for multi-class thyroid disease classification.
• To address the challenge of diagnosing co-existing thyroid disease types.
• To enhance diagnostic accuracy by integrating computed tomography characteristics and multi-scale feature maps.

Main Methods:

• Proposed a novel multi-channel CNN architecture for thyroid disease diagnosis.
• Utilized computed tomography characteristics for comprehensive thyroid gland assessment.
• Investigated feature map concatenation at different scales to boost CNN performance.

Main Results:

• The multi-channel CNN outperformed the standard single-channel CNN in accuracy, precision, recall, and F1-score.
• Achieved high diagnostic accuracy (0.909±0.048) and specificity (0.994±0.001).
• Demonstrated consistent performance across gender groups, with accuracies of 0.908 (female) and 0.901 (male).

Conclusions:

• The proposed multi-channel CNN model exhibits excellent generalization capabilities.
• The model shows significant potential for clinical deployment as a decision support tool.
• This approach advances the multi-class classification and co-diagnosis of thyroid diseases.