Abstract

This study presents the development and evaluation of a novel, multi-model AI system designed to train and deliver multiple deep learning models for the classification of focal lesions in thyroid, using ultrasound images. Leveraging a dataset of 484 images, we trained a diverse array of 1300 models encompassing advanced convolutional neural networks, including ResNet, DenseNet and VGG architectures. To minimize random errors, the training dataset was randomly sampled 20 times. The primary objective was to enhance the diagnostic accuracy in distinguishing benign from malignant thyroid nodules through automated analysis. The performance of our models was rigorously assessed, demonstrating promising results with an average area under the curve of 0.86 and sensitivity of 0.85. These findings highlight the significant potential of integrating deep learning techniques with ultrasound imaging to improve the classification of thyroid nodules.

Recommended Citation

Rafało, M. & Żyłka, A. (2024). Multi-Model Deep Learning Framework for Thyroid Cancer Classification Using Ultrasound Imaging. In B. Marcinkowski, A. Przybylek, A. Jarzębowicz, N. Iivari, E. Insfran, M. Lang, H. Linger, & C. Schneider (Eds.), Harnessing Opportunities: Reshaping ISD in the post-COVID-19 and Generative AI Era (ISD2024 Proceedings). Gdańsk, Poland: University of Gdańsk. ISBN: 978-83-972632-0-8. https://doi.org/10.62036/ISD.2024.78

Paper Type

Short Paper

DOI

10.62036/ISD.2024.78

Download

Share

COinS
 

Multi-Model Deep Learning Framework for Thyroid Cancer Classification Using Ultrasound Imaging

This study presents the development and evaluation of a novel, multi-model AI system designed to train and deliver multiple deep learning models for the classification of focal lesions in thyroid, using ultrasound images. Leveraging a dataset of 484 images, we trained a diverse array of 1300 models encompassing advanced convolutional neural networks, including ResNet, DenseNet and VGG architectures. To minimize random errors, the training dataset was randomly sampled 20 times. The primary objective was to enhance the diagnostic accuracy in distinguishing benign from malignant thyroid nodules through automated analysis. The performance of our models was rigorously assessed, demonstrating promising results with an average area under the curve of 0.86 and sensitivity of 0.85. These findings highlight the significant potential of integrating deep learning techniques with ultrasound imaging to improve the classification of thyroid nodules.