Machine Learning Approach for Thyroid Nodule Classification Using CNN

Abstract

Thyroid nodule classification is a crucial aspect of clinical diagnosis, requiring automated and precise classification methods to differentiate between benign and malignant cases. This study develops a MATLAB-based computer-aided diagnosis (CAD) system to classify thyroid nodules using ultrasound images. The dataset comprises 2,450 ultrasound images from the PERSIAN cohort (Iran, 2018–2020), along with 3,538 images from the Algerian Ultrasound Images Thyroid Dataset (AUITD) and 99 cases from the DDTI Thyroid Ultrasound Images database. Image preprocessing was conducted using median filtering and contrast enhancement, followed by feature extraction utilizing Gray-Level Co-occurrence Matrix (GLCM) and Discrete Wavelet Transform (DWT). Instead of conventional Support Vector Machine (SVM) classification, a novel hybrid deep learning model integrating Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks was implemented to capture spatial and sequential dependencies within ultrasound images. Experimental results demonstrated that the proposed CNN-LSTM model outperformed SVM, achieving an accuracy of 94.2% compared to SVM's 86.5%, with a sensitivity of 92.8% and specificity of 95.1%. Additionally, an F1-score of 0.93 and an AUC-ROC of 0.96 indicate the robustness of the proposed system in distinguishing benign from malignant nodules. The findings suggest that this method can improve clinical decision-making by reducing unnecessary biopsies and enhancing the accuracy of ultrasound-based thyroid nodule assessment.