Enhancing Thyroid Cancer Diagnosis through a Resilient Deep Learning Ensemble Approach

Abstract

Thyroid nodules affect over half of the world's population, making it one of the most serious concerns for medical professionals, particularly pathologists. Fine needle aspiration (FNA) is a technique that examines cell samples for abnormalities or malignant features, with category IV being one of the most dangerous cancers. Deep learning, a dominant branch of artificial intelligence, emerges as a promising solution to revolutionize thyroid nodule diagnosis to achieve unprecedented diagnostic performance. The primary contribution of this paper is to introduce various deep learning models, including Resnet50, DensNet121, VGG16, MobileNetV2, and Vision Transformer (ViT), for image classification, particularly on a thyroid dataset sources from the National Cancer Institute-Cairo University. The primary focus is on determining the most suitable deep learning classifier. The research further explores ensemble learning techniques to enhance performance, minimize errors, and prevent overfitting in the trained models. Results indicate that Resnet50 achieved a notable F1-score of 92%, surpassing existing state-of-the-art methods on the same dataset. Additionally, a meta-learning ensemble fusion method demonstrated additional superior performance, surpassing both baseline models and other ensemble methods, with an increased f1-score nf 3.078%,