Leveraging feature fusion network for automated classification of thyroid tumors from ultrasound medical scans

Abstract

Thyroid cancer, characterized by uncontrolled cellular growth in the thyroid gland, poses significant health risks. This study presents a novel diagnostic model for distinguishing benign and malignant thyroid tumors in ultrasound images by integrating a transferred EfficientNetB0 model with a new parallel deep convolutional neural network (CNN). The methodology involves preprocessing using Anisotropic Diffusion Filtering (ADF) for noise reduction, followed by feature extraction via deep CNNs. A refined classification model, developed through feature selection and dimensionality reduction, is trained and validated using a dataset of 1137 ultrasound images. The proposed system achieves an accuracy of 92.28 % and an F1-score of 92.76 %, demonstrating its effectiveness in assisting clinical diagnosis. Comparative complexity analysis validate its robustness in addition to visual analysis tool (spider graph) that provides additional insights. The results demonstrate the potential of deep learning (DL) models in improving the reliability of thyroid cancer diagnosis, aiding clinicians in decision-making processes and reducing the risk of misdiagnosis.