A lightweight GhostYOLO framework with GhostDynamic Module and multi-scale attention for thyroid nodule diagnosis.

Ultrasound is the primary tool for thyroid nodule assessment, but diagnostic accuracy and efficiency require improvement. This study aims to develop a high-performance, computationally efficient deep learning model for thyroid nodule detection. We propose GhostYOLO, a novel framework based on YOLOv11. It incorporates a GhostDynamic Module to reduce complexity and a Multi-Scale Attention (MSAttention) mechanism to improve feature extraction across varying nodule sizes. The model was trained and evaluated on a large internal dataset of 12,385 ultrasound images from 3,140 patients and validated on an external public dataset of 8,500 images from 842 cases. On the internal test set, GhostYOLO achieved a mean Average Precision (mAP) of 67.2% and an F1-Score of 86.3%. External validation yielded a mAP50 of 74.4% (mAP: 48.9%) and an F1-Score of 79.5%. The model demonstrated high efficiency with only 9.3 million parameters and 20.2 GFLOPs. GhostYOLO offers a robust and computationally efficient solution for automated thyroid nodule detection, showing strong potential to aid in clinical diagnosis.