Change Point--Aware Evaluation and Re-Calibration of PPG-Based Blood Pressure Estimation
本文针对PPG血压估计在血压快速波动期间的性能下降问题，提出基于时间序列变化点检测的评估框架，并引入由变化点触发的目标重校准方法以提高鲁棒性。
0 citationsRead paper
Selected work
Representative Papers
Bidirectional Fusion Guided by Cardiac Patterns for Semi-Supervised ECG Segmentation
This work addresses the scarcity of annotated data in electrocardiogram (ECG) segmentation by proposing CardioMix, a framework that introduces a cardiac rhythm–guided bidirectional CutMix strategy to enable physiologically consistent data augmentation between labeled and unlabeled samples. By explicitly modeling the temporal patterns of cardiac electrical activity, CardioMix ensures that augmented samples maintain physiological plausibility. Designed as a plug-and-play module, it seamlessly integrates with prevailing semi-supervised segmentation algorithms. Evaluated on the SemiSegECG multi-dataset benchmark, CardioMix consistently outperforms existing CutMix variants across varying label proportions, achieving robust and state-of-the-art segmentation performance.
0 citationsRead paper
Task-Agnostic Noisy Label Detection via Standardized Loss Aggregation
This study addresses the pervasive issue of label noise in medical imaging datasets, primarily caused by inter-observer variability and ambiguous samples. To tackle this challenge, the authors propose the Standardized Loss Aggregation (SLA) framework, which iteratively aggregates standardized fold-level validation losses across repeated cross-validation runs. By jointly characterizing the frequency and magnitude of performance deviations, SLA generates statistically stable and interpretable noise scores without requiring task-specific adaptations. Experimental results on public fundus image datasets demonstrate that SLA consistently outperforms existing baselines across varying noise levels, exhibiting particularly faster convergence and superior performance in low-noise regimes, thereby significantly enhancing data annotation quality.
0 citationsRead paper
A Comparative Study of Machine Learning and Deep Learning for Out-of-Distribution Detection
This study addresses the challenge of efficiently and accurately detecting out-of-distribution (OOD) samples in visually constrained medical imaging scenarios to enhance AI system reliability. The authors systematically evaluate the OOD detection performance of both traditional machine learning and deep learning approaches on a dataset comprising over 60,000 multi-resolution fundus and non-fundus images. Experimental results demonstrate that both method families achieve near-perfect AUROC (1.000) and accuracy (0.999–1.000) on in-distribution and out-of-distribution validation sets. Notably, lightweight machine learning models attain comparable detection accuracy while significantly reducing inference latency, highlighting their practical utility in resource-constrained clinical applications.
0 citationsRead paper
Automated Structured Radiology Report Generation with Rich Clinical Context
Existing SRRG methods neglect clinical context, leading to temporal hallucinations—such as fabricated medical history—and undermining report reliability. To address this, we propose C-SRRG, the first framework that systematically integrates multi-view X-ray images, clinical indications, imaging parameters, and prior medical history into a unified multimodal clinical context. Leveraging a multimodal large language model, C-SRRG jointly models radiographic imagery and structured clinical text to ensure temporally consistent, clinically grounded reasoning. This design inherently suppresses temporal hallucinations at their source, significantly improving diagnostic logical coherence and report accuracy. On standard benchmarks, C-SRRG comprehensively outperforms state-of-the-art methods, yielding higher-quality, clinically aligned reports. To foster reproducibility and advancement in clinically aware automated radiology reporting, we publicly release our code, dataset, and model weights.
0 citationsRead paper
Recent publications
Latest Papers
Change Point--Aware Evaluation and Re-Calibration of PPG-Based Blood Pressure Estimation
本文针对PPG血压估计在血压快速波动期间的性能下降问题，提出基于时间序列变化点检测的评估框架，并引入由变化点触发的目标重校准方法以提高鲁棒性。
0 citationsRead paper
Bidirectional Fusion Guided by Cardiac Patterns for Semi-Supervised ECG Segmentation
This work addresses the scarcity of annotated data in electrocardiogram (ECG) segmentation by proposing CardioMix, a framework that introduces a cardiac rhythm–guided bidirectional CutMix strategy to enable physiologically consistent data augmentation between labeled and unlabeled samples. By explicitly modeling the temporal patterns of cardiac electrical activity, CardioMix ensures that augmented samples maintain physiological plausibility. Designed as a plug-and-play module, it seamlessly integrates with prevailing semi-supervised segmentation algorithms. Evaluated on the SemiSegECG multi-dataset benchmark, CardioMix consistently outperforms existing CutMix variants across varying label proportions, achieving robust and state-of-the-art segmentation performance.
0 citationsRead paper
Task-Agnostic Noisy Label Detection via Standardized Loss Aggregation
This study addresses the pervasive issue of label noise in medical imaging datasets, primarily caused by inter-observer variability and ambiguous samples. To tackle this challenge, the authors propose the Standardized Loss Aggregation (SLA) framework, which iteratively aggregates standardized fold-level validation losses across repeated cross-validation runs. By jointly characterizing the frequency and magnitude of performance deviations, SLA generates statistically stable and interpretable noise scores without requiring task-specific adaptations. Experimental results on public fundus image datasets demonstrate that SLA consistently outperforms existing baselines across varying noise levels, exhibiting particularly faster convergence and superior performance in low-noise regimes, thereby significantly enhancing data annotation quality.
0 citationsRead paper
A Comparative Study of Machine Learning and Deep Learning for Out-of-Distribution Detection
This study addresses the challenge of efficiently and accurately detecting out-of-distribution (OOD) samples in visually constrained medical imaging scenarios to enhance AI system reliability. The authors systematically evaluate the OOD detection performance of both traditional machine learning and deep learning approaches on a dataset comprising over 60,000 multi-resolution fundus and non-fundus images. Experimental results demonstrate that both method families achieve near-perfect AUROC (1.000) and accuracy (0.999–1.000) on in-distribution and out-of-distribution validation sets. Notably, lightweight machine learning models attain comparable detection accuracy while significantly reducing inference latency, highlighting their practical utility in resource-constrained clinical applications.
0 citationsRead paper
Automated Structured Radiology Report Generation with Rich Clinical Context
Existing SRRG methods neglect clinical context, leading to temporal hallucinations—such as fabricated medical history—and undermining report reliability. To address this, we propose C-SRRG, the first framework that systematically integrates multi-view X-ray images, clinical indications, imaging parameters, and prior medical history into a unified multimodal clinical context. Leveraging a multimodal large language model, C-SRRG jointly models radiographic imagery and structured clinical text to ensure temporally consistent, clinically grounded reasoning. This design inherently suppresses temporal hallucinations at their source, significantly improving diagnostic logical coherence and report accuracy. On standard benchmarks, C-SRRG comprehensively outperforms state-of-the-art methods, yielding higher-quality, clinically aligned reports. To foster reproducibility and advancement in clinically aware automated radiology reporting, we publicly release our code, dataset, and model weights.
0 citationsRead paper