Video-based ultrasound diagnosis via difference-guided multi-frame modeling.
Authors
Affiliations (5)
Affiliations (5)
- Department of Information and Artificial intelligence, Yangzhou University, Yangzhou, Jiangsu, China. Electronic address: [email protected].
- Department of Ultrasound, The Affiliated Hospital of Yangzhou University, Yangzhou University, Yangzhou, China. Electronic address: [email protected].
- Department of Information and Artificial intelligence, Yangzhou University, Yangzhou, Jiangsu, China. Electronic address: [email protected].
- Department of Information and Artificial intelligence, Yangzhou University, Yangzhou, Jiangsu, China. Electronic address: [email protected].
- Department of Radiology, The Affiliated Hospital of Yangzhou University, Yangzhou University, Yangzhou, China. Electronic address: [email protected].
Abstract
Ultrasound diagnosis plays a crucial role in clinical practice due to its non-invasive nature, real-time imaging capability, and wide applicability in detecting various lesions. However, existing AI-assisted methods typically treat ultrasound data similarly to static imaging modalities such as MRI and CT, overlooking that ultrasound examinations are inherently acquired in video form with rich temporal and structural information. To address this limitation, we propose a difference-guided multi-frame modeling method for ultrasound diagnosis based on video. Specifically, a difference-guided key-frame selection strategy is first introduced to identify the most informative frames by measuring inter-frame variations, thereby reducing redundancy and suppressing noise in ultrasound video sequences. The selected key frames are then modeled through a multi-frame learning architecture to capture long-range dependencies and contextual interactions, followed by an attention-based feature fusion mechanism for robust case-level prediction. Extensive experiments on three ultrasound video datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches across multiple evaluation metrics, validating its effectiveness and robustness in ultrasound diagnosis.