GazeCausalNet: Causality-inspired gaze-guided region intervention for breast ultrasound diagnosis.
Authors
Affiliations (6)
Affiliations (6)
- College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, Jiangsu, China.
- Department of Ultrasound, The Affiliated People's Hospital of Jiangsu University, Zhenjiang, Jiangsu, China.
- School of Biomedical Engineering and the Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, 200240, China.
- College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, Jiangsu, China. Electronic address: [email protected].
- Department of Ultrasound, Huashan Hospital, Fudan University, Shanghai, 200040, Shanghai, China; Shanghai Institute of Medical Imaging, Shanghai, 200032, Shanghai, China. Electronic address: [email protected].
- College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, 211106, Jiangsu, China. Electronic address: [email protected].
Abstract
Deep learning has advanced breast ultrasound analysis, yet spurious correlations and domain shifts still impede cross-setting generalization. Because most models optimize dataset-dependent associations rather than stable, diagnosis-relevant representations, they may be vulnerable to contextual and acquisition-related factors tied to imaging protocols and acquisition artifacts. To address this, we propose GazeCausalNet, a causality-inspired gaze-guided region intervention framework that uses expert gaze trajectories to identify gaze-attended regions enriched in diagnostic evidence. Based on Linear Optimal Transport (LOT), GazeCausalNet generates smooth and spatially coherent counterfactual-style samples supporting controlled same-class contextual perturbations and opposite-class semantic-transition perturbations. We further introduce counterfactual gradient alignment, comprising background-stability alignment to reduce excessive sensitivity along constructed contextual perturbation directions and semantic-transition alignment to amplify sensitivity to diagnostically meaningful changes. Experiments on public and private breast ultrasound datasets achieve 97.08% and 95.83% AUC, respectively, supporting improved robustness and clinically informed interpretability without claiming formal causal-effect identification.