ResTAC-Net: A Residual Texture-Attention Convolutional Network for Robust Thyroid Nodule Classification from Ultrasound Image

Abstract

Thyroid cancer is a type of endocrine malignancy that is increasing rapidly worldwide. Prompt and accurate diagnosis is necessary to optimize patient outcomes. Ultrasound is frequently employed for evaluating thyroid nodules because it is non-invasive, cost-effective, and provides real-time imaging. However, interpretation accuracy often relies on the radiologist's experience, which can result in inter-observer variability. This paper presents a fully automated deep learning system for more precise classification of thyroid nodules. The approach incorporates advanced image preprocessing techniques and a custom residual convolutional neural network (CNN), trained from the initial stages to address variability challenges. The utilized dataset is extensive, comprising 5,195 thyroid nodule images from two public databases, the Digital Database of Thyroid Image (DDTI) and the Thyroid Nodule Ultrasound Image (TNUI) dataset. Both datasets include benign and malignant cases verified by pathology. The proposed network integrates residual learning modules to enhance feature extraction, as well as adaptive feature aggregation, spatial attention mechanisms, and multi-scale convolutional operations to improve performance. Preprocessing steps include standardizing the region of interest, enhancing contrast, normalizing intensity, and reducing speckle noise to facilitate efficient feature learning by the model. Experimental results indicate high effectiveness, with an overall accuracy of 97.3 percent, sensitivity of 96.8 percent, specificity of 97.7 percent, F1 score of 97.1 percent, Matthew’s correlation coefficient of 0.94, and an AUC of 0.99. Comparative analysis demonstrates that the proposed custom residual CNN outperforms established deep learning architectures under similar conditions. The suggested framework provides a robust and reliable computer-aided diagnostic tool that can support clinicians in thyroid cancer screening, improve diagnostic precision, and reduce inter-observer variability in ultrasound radiation.