Selected work

Representative Papers

Advancing AI-Powered Medical Image Synthesis: Insights from MedVQA-GI Challenge Using CLIP, Fine-Tuned Stable Diffusion, and Dream-Booth + LoRA

Feb 28, 2025Conference and Labs of the Evaluation Forum

This work addresses the critical bottleneck in medical diagnosis: the absence of methods for dynamically generating high-fidelity images from clinical text. We propose the first dual-task text-to-image framework tailored for gastrointestinal imaging—comprising Image Synthesis (IS) and Optimal Prompt Generation (OPG). Methodologically, we systematically integrate fine-tuned Stable Diffusion, DreamBooth-based personalization, and LoRA-based low-rank adaptation, coupled with a CLIP text encoder to jointly optimize generation quality, class controllability, and diversity. Evaluated on multi-center data, our approach achieves FID = 0.064 and Inception Score = 2.327, significantly outperforming baseline models. Key contributions include: (1) transcending traditional static image analysis by enabling dynamic, natural-language-driven medical image synthesis; (2) establishing a scalable, high-precision prompt optimization mechanism; and (3) providing a reproducible technical pipeline and standardized evaluation benchmark for clinical-oriented generative AI.

1 citationsRead paper

Property-Specific Recoverability from Contact PPG to Camera rPPG under Heterogeneous Observation Conditions

Aug 27, 2026

研究通过CHROM方法评估了在不同观察条件下从接触式PPG到远程PPG的特定属性恢复能力，发现生理属性在记录中的保留程度和对条件的依赖性存在差异。

0 citationsRead paper

CFD-Guided Detection of Concept Drift in Multimodal Physiologic Signals

Aug 07, 2026

This study addresses the unreliability of cardiovascular signal predictions from wearable devices due to interference from motion, respiration, and posture changes. To mitigate this, the authors propose PECS, a physiological stability framework that innovatively adapts concepts from computational fluid dynamics to detect concept drift and enable interpretable trust routing. PECS achieves this by comparing internal model variations against observable changes in multimodal signals—primarily ECG, supplemented by PPG, with respiratory signals selectively incorporated in ambiguous scenarios. The framework integrates domain-pair selection strategies with multimodal fusion to dynamically guide model decisions. Evaluated on the BIDMC and MIMIC datasets, PECS achieves concept drift classification accuracies of 0.8786 and 0.9560, respectively, demonstrating its effectiveness and highlighting the selective utility of respiratory signals in resolving prediction discrepancies.

0 citationsRead paper

Attractor Domain Theory: A Mathematical Framework for Cardiovascular Attractor Analysis with Wearable Photoplethysmography (PPG) Validation

Jun 20, 2026

This study addresses the lack of a systematic framework linking geometric properties of reconstructed cardiovascular attractors to physiological quantities, which has led to ad hoc feature selection and difficulty interpreting negative results. The authors propose Attractor Domain Theory (ADT), which uniquely partitions attractor information into three non-redundant domains—geometric, ergodic, and variational—each supporting artifact rejection, stability estimation, and hemodynamic inference, respectively. They establish the necessity and sufficiency of this tripartite division through a Domain Sufficiency Theorem. Validated on 176,742 PPG segments using Takens embedding, an SCSI verification framework, and nonlinear dynamical features, the geometric domain yields an AUC of 0.757 (negative predictive value: 0.966) after bias correction. Ablation experiments identify the nonlinear correlation dimension \(C_{NL}\) as critical, with its removal reducing AUC by 0.413.

0 citationsRead paper

Parameter-Efficient VLMs for Gastrointestinal Endoscopy: Medical Image Generation and Clinical Visual Question Answering

May 23, 2026

This study addresses the challenges of limited annotated data, privacy constraints, and high computational costs associated with conventional fine-tuning that hinder the reliability and scalability of AI systems in gastrointestinal endoscopy. To overcome these issues, the authors propose the first dual-pipeline parameter-efficient fine-tuning (PEFT) framework: one pipeline leverages the Florence-2 model for clinical visual question answering (VQA), while the other employs LoRA-based fine-tuning of Stable Diffusion 2.1 to generate high-fidelity synthetic endoscopic images, thereby preserving patient privacy. Evaluated on the Kvasir-VQA dataset, the approach achieves a ROUGE-1 score of 0.92 and BLEU of 0.24; the synthetic images attain a Fréchet Bowel Distance (FBD) of 1450, reduce computational costs by nearly 90%, and demonstrate superior semantic consistency compared to baseline methods such as FLUX and MSDM, significantly enhancing model interpretability and clinical applicability.

0 citationsRead paper

Recent publications

Latest Papers

Property-Specific Recoverability from Contact PPG to Camera rPPG under Heterogeneous Observation Conditions

Aug 27, 2026

研究通过CHROM方法评估了在不同观察条件下从接触式PPG到远程PPG的特定属性恢复能力，发现生理属性在记录中的保留程度和对条件的依赖性存在差异。

0 citationsRead paper

CFD-Guided Detection of Concept Drift in Multimodal Physiologic Signals

Aug 07, 2026

This study addresses the unreliability of cardiovascular signal predictions from wearable devices due to interference from motion, respiration, and posture changes. To mitigate this, the authors propose PECS, a physiological stability framework that innovatively adapts concepts from computational fluid dynamics to detect concept drift and enable interpretable trust routing. PECS achieves this by comparing internal model variations against observable changes in multimodal signals—primarily ECG, supplemented by PPG, with respiratory signals selectively incorporated in ambiguous scenarios. The framework integrates domain-pair selection strategies with multimodal fusion to dynamically guide model decisions. Evaluated on the BIDMC and MIMIC datasets, PECS achieves concept drift classification accuracies of 0.8786 and 0.9560, respectively, demonstrating its effectiveness and highlighting the selective utility of respiratory signals in resolving prediction discrepancies.

0 citationsRead paper

Attractor Domain Theory: A Mathematical Framework for Cardiovascular Attractor Analysis with Wearable Photoplethysmography (PPG) Validation

Jun 20, 2026

This study addresses the lack of a systematic framework linking geometric properties of reconstructed cardiovascular attractors to physiological quantities, which has led to ad hoc feature selection and difficulty interpreting negative results. The authors propose Attractor Domain Theory (ADT), which uniquely partitions attractor information into three non-redundant domains—geometric, ergodic, and variational—each supporting artifact rejection, stability estimation, and hemodynamic inference, respectively. They establish the necessity and sufficiency of this tripartite division through a Domain Sufficiency Theorem. Validated on 176,742 PPG segments using Takens embedding, an SCSI verification framework, and nonlinear dynamical features, the geometric domain yields an AUC of 0.757 (negative predictive value: 0.966) after bias correction. Ablation experiments identify the nonlinear correlation dimension \(C_{NL}\) as critical, with its removal reducing AUC by 0.413.

0 citationsRead paper

Parameter-Efficient VLMs for Gastrointestinal Endoscopy: Medical Image Generation and Clinical Visual Question Answering

May 23, 2026

This study addresses the challenges of limited annotated data, privacy constraints, and high computational costs associated with conventional fine-tuning that hinder the reliability and scalability of AI systems in gastrointestinal endoscopy. To overcome these issues, the authors propose the first dual-pipeline parameter-efficient fine-tuning (PEFT) framework: one pipeline leverages the Florence-2 model for clinical visual question answering (VQA), while the other employs LoRA-based fine-tuning of Stable Diffusion 2.1 to generate high-fidelity synthetic endoscopic images, thereby preserving patient privacy. Evaluated on the Kvasir-VQA dataset, the approach achieves a ROUGE-1 score of 0.92 and BLEU of 0.24; the synthetic images attain a Fréchet Bowel Distance (FBD) of 1450, reduce computational costs by nearly 90%, and demonstrate superior semantic consistency compared to baseline methods such as FLUX and MSDM, significantly enhancing model interpretability and clinical applicability.

0 citationsRead paper

Attractor-Vascular Coupling Theory: Formal Grounding and Empirical Validation for AAMI-Standard Cuffless Blood Pressure Estimation from Smartphone Photoplethysmography

May 11, 2026

This study proposes a novel method for clinical-grade, cuffless blood pressure estimation using only smartphone-based photoplethysmography (PPG). By introducing the Attractor–Vascular Coupling Theory (AVCT), the work establishes, for the first time, a rigorous nonlinear dynamical systems framework linking geometric features of PPG attractors to blood pressure, while predicting a hierarchy of feature importance to enhance model interpretability. The approach integrates Takens’ delay embedding, attractor morphology descriptors, a cardiac stability index (CSI), pulse transit time (PTT), and LightGBM regression. Evaluated on 46 subjects, it achieves mean absolute errors of 2.05/1.67 mmHg for systolic/diastolic blood pressure (correlation coefficients: 0.990/0.991), with 70%/76% of individuals meeting AAMI standards, demonstrating clinical accuracy with only a single-point calibration.

0 citationsRead paper