Predicting Collision Cross Sections with GRACE: Geometric Residual Adduct Conditioning via Early-fusion
为解决碰撞截面预测难题，提出GRACE模型，通过早期融合几何残差加合物调节方法，利用3D结构信息和加合物条件改善预测准确性。
0 citationsRead paper
Selected work
Representative Papers
Random Hazard Forests
本文提出随机风险森林（RHF）方法，通过非参数风险似然估计来解决临床数据中不规则时间和不同测量安排下的个体化风险预测问题。
0 citationsRead paper
Quantum Kernel Estimation for the Discovery of Early Lung Cancer Detection
研究使用量子-经典混合机器学习方法，通过DNA片段组学和甲基化分析来提高早期肺癌检测的准确性，采用量子核方法有效捕捉非线性cfDNA结构。
0 citationsRead paper
Autonomous Skeletal Landmark Localization towards Agentic C-Arm Control
This study addresses the frequent need for manual intervention in C-arm repositioning during emergency procedures, which often delays treatment. The authors propose a novel approach that leverages a fine-tuned multimodal large language model (MLLM) to autonomously localize skeletal landmarks by integrating synthetic and real X-ray data during training. Incorporating clinical feedback enables dynamic C-arm navigation. Evaluated on two datasets, the method achieves landmark localization accuracy comparable to state-of-the-art deep learning models while demonstrating superior reasoning and spatial awareness capabilities. These attributes facilitate error correction and sequential adjustments, substantially enhancing procedural automation in fluoroscopic guidance.
0 citationsRead paper
Automated C-Arm Positioning via Conformal Landmark Localization
This study addresses the high radiation exposure and low efficiency associated with manual C-arm positioning in fluoroscopy-guided surgery by proposing a fully automated method for 3D anatomical landmark localization. Methodologically, we design an end-to-end deep network that takes a single X-ray image as input and directly predicts a 3D displacement vector to the target anatomical landmark. To enhance anatomical plausibility, we introduce skeletal pose regularization and a probabilistic loss function; further, we integrate conformal prediction to quantify both aleatoric and epistemic uncertainty, yielding well-calibrated 3D confidence regions. Trained on DeepDRR-synthesized data, our approach achieves sub-centimeter accuracy (mean error <6 mm) across multiple network architectures, with empirically validated coverage of predicted intervals. To the best of our knowledge, this is the first work to jointly incorporate conformal prediction and anatomical constraints for autonomous C-arm navigation—significantly improving system safety, reliability, and clinical applicability.
0 citationsRead paper
Recent publications
Latest Papers
Predicting Collision Cross Sections with GRACE: Geometric Residual Adduct Conditioning via Early-fusion
为解决碰撞截面预测难题，提出GRACE模型，通过早期融合几何残差加合物调节方法，利用3D结构信息和加合物条件改善预测准确性。
0 citationsRead paper
Random Hazard Forests
本文提出随机风险森林（RHF）方法，通过非参数风险似然估计来解决临床数据中不规则时间和不同测量安排下的个体化风险预测问题。
0 citationsRead paper
Quantum Kernel Estimation for the Discovery of Early Lung Cancer Detection
研究使用量子-经典混合机器学习方法，通过DNA片段组学和甲基化分析来提高早期肺癌检测的准确性，采用量子核方法有效捕捉非线性cfDNA结构。
0 citationsRead paper
Autonomous Skeletal Landmark Localization towards Agentic C-Arm Control
This study addresses the frequent need for manual intervention in C-arm repositioning during emergency procedures, which often delays treatment. The authors propose a novel approach that leverages a fine-tuned multimodal large language model (MLLM) to autonomously localize skeletal landmarks by integrating synthetic and real X-ray data during training. Incorporating clinical feedback enables dynamic C-arm navigation. Evaluated on two datasets, the method achieves landmark localization accuracy comparable to state-of-the-art deep learning models while demonstrating superior reasoning and spatial awareness capabilities. These attributes facilitate error correction and sequential adjustments, substantially enhancing procedural automation in fluoroscopic guidance.
0 citationsRead paper
Automated C-Arm Positioning via Conformal Landmark Localization
This study addresses the high radiation exposure and low efficiency associated with manual C-arm positioning in fluoroscopy-guided surgery by proposing a fully automated method for 3D anatomical landmark localization. Methodologically, we design an end-to-end deep network that takes a single X-ray image as input and directly predicts a 3D displacement vector to the target anatomical landmark. To enhance anatomical plausibility, we introduce skeletal pose regularization and a probabilistic loss function; further, we integrate conformal prediction to quantify both aleatoric and epistemic uncertainty, yielding well-calibrated 3D confidence regions. Trained on DeepDRR-synthesized data, our approach achieves sub-centimeter accuracy (mean error <6 mm) across multiple network architectures, with empirically validated coverage of predicted intervals. To the best of our knowledge, this is the first work to jointly incorporate conformal prediction and anatomical constraints for autonomous C-arm navigation—significantly improving system safety, reliability, and clinical applicability.
0 citationsRead paper