A Case Study of DenseNet, ResNet and Vision Transformers for Thyroid Nodule Analysis in Medical Imaging
Year: 2025
Type: article
Abstract: In order to classify thyroid nodules using ultrasound imaging [1], this study assesses the effectiveness of three deep learning models: Vision Transformer (ViT), DenseNet, and ResNet. Seven thousand thyroid ultrasound pictures from Morocco's Hassan II Hospital (2005–2022) were utilized as the dataset. Accuracy, F1-score, sensitivity, and specificity were important parameters. DenseNet did somewhat better with 89.3% accuracy and F1-score than ResNet, which had 87.7% accuracy and an 87.8% F1-score. ViT outperformed both, achieving 91.5% accuracy and a 91.4% F1-score, demonstrating superior global context capture. ResNet excels in gradient flow optimization, DenseNet in feature propagation for smaller datasets, and ViT in versatility but requires larger datasets. The study highlights trade-offs between transformer-based and CNN-based architectures, emphasizing the importance of dataset characteristics and task requirements for optimal diagnostic outcomes in medical imaging.
Authors: Karima Bahmane, Hamid Aksasse, Brahim Alkhalil Chaouki
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4409600409