A dual attention mechanism-based key frame extraction method for thyroid dynamic ultrasound: a multicenter study

The incidence of thyroid nodules and thyroid cancer has increased significantly in recent years. Ultrasound imaging, as an important non-invasive diagnostic tool, plays a crucial role in the detection of thyroid diseases. However, traditional ultrasound image diagnosis relies on manual selection of key frames by physicians, which is susceptible to human factors and is inefficient. This paper proposes a deep learning-based key frame extraction system for thyroid ultrasound videos. By combining several advanced deep learning models, the system achieves efficient and accurate automated key frame extraction and benign-malignant classification. The system first uses the feature extraction network of a pre-trained ultrasound classification model to obtain effective features of ultrasound images. Then, a dual attention mechanism extraction network generates a set of five candidate key frames. Next, a YOLOv8-based post-processing network selects the frame with the largest area from the candidates for further analysis. Finally, a ResNet50-based benign-malignant classification network classifies the selected key frame, effectively improving the diagnostic accuracy of thyroid nodules. Experimental results show that the system significantly improves the efficiency of key frame extraction and the accuracy of diagnosis, demonstrating high potential for clinical application.