Anomaly Segmentation and Classification in Thyroid Ultrasound Images Using YOLOv8-GSO-Inception-V4

In recent years, deep learning models have become increasingly prevalent in medical imaging, particularly in assisting radiologists with diagnostic tasks such as thyroid ultrasound interpretation. Among these, the You Only Look Once (YOLO) series and Inception-based architectures have demonstrated remarkable performance in computer vision applications, including object detection, segmentation, and classification. This paper proposes a unified deep learning pipeline that integrates YOLOv8 for segmentation and a GSO-optimized Inception-V4 (GSO-Inception-V4) for classification to enable automated thyroid nodule analysis in ultrasound images. Specifically, YOLOv8 is employed to segment nodule regions with high spatial precision, while GSO-Inception-V4 classifies the extracted nodules as benign or malignant, thus providing a comprehensive and efficient diagnostic solution. The proposed system was trained and evaluated using the publicly available Thyroid Digital Image Database (TDID). Experimental results show that the segmentation model achieved a precision of 97.37%, recall of 97.37%, mAP50 of 98.9%, confirming the accuracy and robustness of YOLOv8 for thyroid nodule localization. For classification, the GSO-Inception-V4 model attained 98.76% accuracy, 99% precision, 99.5% sensitivity, and an F1-score of 99.25%, particularly excelling in identifying malignant nodules. These results demonstrate the model's strong potential for clinical deployment and highlight its value as a practical tool for enhancing diagnostic accuracy in thyroid ultrasound analysis.