Hybrid deep learning and ensemble learning approach for high accuracy thyroid disease classification

Abstract

Thyroid disease is a common endocrine disorder affecting the thyroid gland, a small butterfly-shaped organ at the base of the neck. According to the World Health Organization (WHO), nearly one billion people worldwide are affected by thyroid-related conditions. Conventional diagnostic methods, such as thyroid scans and function tests, are often costly, time-consuming, and complex for clinicians to interpret. To overcome these limitations, this study introduces a novel temporal conditional-Markov random field (TC MRF) framework for early detection and classification of thyroid disease. The multi-modality images computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound (US) are collected from the ImageNet database and preprocessed using contrast stretching adaptive Gaussian star (CSAGS) filter to improve image clarity. The enhanced images are then processed over a convolutional neural network (CNN) for feature extraction. These features are classified using a random forest (RF) model to determine whether the thyroid condition is normal or abnormal. The proposed TC MRF achieves a classification accuracy of 98.27% and F1-score of 96.05%. The TC-MRF enhances the total accuracy range of 6.30%, 4.11%, and 5.36% better than naive Bayes, multilayer perceptron (MLP), and decision tree, respectively.