Neural Network Architectures for Automated Thyroid Nodule Diagnosis: A Comparative Study with Attention-Enhanced Multimodal Integration

Automated diagnosis of thyroid nodules using neural networks has gained substantial traction due to rapid advancements in deep learning and medical imaging. This study presents a comprehensive comparison of state-of-the-art neural network architectures ResNet, DenseNet, U-Net, Inception, and MobileNet for classifying thyroid nodules as benign or malignant. We introduce a novel Attention-Enhanced Multimodal Integration (AEMI) framework that dynamically weights the importance of ultrasound image features and patient clinical data. Using a dataset of 1,000 annotated thyroid ultrasound images with corresponding clinical metadata, we evaluate model performance across multiple metrics including accuracy, sensitivity, specificity, and F1-score. Our proposed AEMI framework with DenseNet-121 backbone achieves superior results (95.3% accuracy, 94.8% sensitivity, 95.7% specificity), outperforming conventional approaches. Additionally, we propose an interpretability module that generates visual explanations of diagnostic decisions, addressing the clinical need for transparent AI systems. Ablation studies confirm the significance of both the attention mechanism and multimodal integration. This work provides valuable insights for selecting suitable architectures for clinical thyroid nodule diagnosis applications and demonstrates how attention mechanisms can enhance diagnostic accuracy while maintaining model interpretability.