RouteRelay: Event-Triggered Cross-Layer Route Reuse for Efficient Dynamic Sparse Attention
本文提出RouteRelay方法，通过跨层复用路由元数据减少动态稀疏注意力中的重复计算，提高效率。
0 citationsRead paper
Selected work
Representative Papers
Multi-Contrast Fusion Module: An attention mechanism integrating multi-contrast features for fetal torso plane classification
Fine-grained anatomical structure identification in fetal trunk standard-plane ultrasound images remains challenging due to inherently low contrast and blurred texture. Method: This paper proposes a lightweight multi-contrast fusion module that introduces a novel multi-contrast attention mechanism at the network’s lower layers. By adaptively weighting low-level features extracted under multiple contrast enhancements, the module significantly improves modeling of subtle anatomical details with negligible parameter overhead. It operates directly on raw ultrasound data to enhance feature representation of clinically critical regions. Contribution/Results: Evaluated on a fetal trunk standard-plane dataset, the method achieves substantial improvements in classification accuracy—particularly for key structures including the heart, spine, and stomach bubble—thereby enhancing diagnostic consistency and reliability. The approach demonstrates clear clinical utility in automated fetal anatomy assessment.
0 citationsRead paper
Multi-Center Study on Deep Learning-Assisted Detection and Classification of Fetal Central Nervous System Anomalies Using Ultrasound Imaging
To address low detection accuracy, high diagnostic burden on clinicians, and elevated misdiagnosis rates in prenatal ultrasound screening for fetal central nervous system (CNS) malformations, this study develops the first multi-center deep learning model covering the entire gestational period for automated detection and classification of four canonical CNS anomalies: anencephaly, encephalocele, holoprosencephaly, and spina bifida. Methodologically, we integrate a ResNet-based architecture, multi-center collaborative training, and class activation mapping (CAM) for interpretable lesion localization. Our contribution is the first demonstration of robust, gestational-week-agnostic CNS anomaly classification with intrinsic interpretability. The model achieves 94.5% patient-level accuracy and 99.3% AUROC. A retrospective reader study demonstrates significant improvements in radiologists’ diagnostic accuracy and efficiency, alongside substantial reduction in misdiagnosis rates.
0 citationsRead paper
Recent publications
Latest Papers
RouteRelay: Event-Triggered Cross-Layer Route Reuse for Efficient Dynamic Sparse Attention
本文提出RouteRelay方法，通过跨层复用路由元数据减少动态稀疏注意力中的重复计算，提高效率。
0 citationsRead paper
Multi-Contrast Fusion Module: An attention mechanism integrating multi-contrast features for fetal torso plane classification
Fine-grained anatomical structure identification in fetal trunk standard-plane ultrasound images remains challenging due to inherently low contrast and blurred texture. Method: This paper proposes a lightweight multi-contrast fusion module that introduces a novel multi-contrast attention mechanism at the network’s lower layers. By adaptively weighting low-level features extracted under multiple contrast enhancements, the module significantly improves modeling of subtle anatomical details with negligible parameter overhead. It operates directly on raw ultrasound data to enhance feature representation of clinically critical regions. Contribution/Results: Evaluated on a fetal trunk standard-plane dataset, the method achieves substantial improvements in classification accuracy—particularly for key structures including the heart, spine, and stomach bubble—thereby enhancing diagnostic consistency and reliability. The approach demonstrates clear clinical utility in automated fetal anatomy assessment.
0 citationsRead paper
Multi-Center Study on Deep Learning-Assisted Detection and Classification of Fetal Central Nervous System Anomalies Using Ultrasound Imaging
To address low detection accuracy, high diagnostic burden on clinicians, and elevated misdiagnosis rates in prenatal ultrasound screening for fetal central nervous system (CNS) malformations, this study develops the first multi-center deep learning model covering the entire gestational period for automated detection and classification of four canonical CNS anomalies: anencephaly, encephalocele, holoprosencephaly, and spina bifida. Methodologically, we integrate a ResNet-based architecture, multi-center collaborative training, and class activation mapping (CAM) for interpretable lesion localization. Our contribution is the first demonstration of robust, gestational-week-agnostic CNS anomaly classification with intrinsic interpretability. The model achieves 94.5% patient-level accuracy and 99.3% AUROC. A retrospective reader study demonstrates significant improvements in radiologists’ diagnostic accuracy and efficiency, alongside substantial reduction in misdiagnosis rates.
0 citationsRead paper