Selected work

Representative Papers

The nature of mathematical models

Feb 11, 2025

Existing mathematical modeling lacks a rigorous, unambiguous ontological foundation, hindering a unified characterization of the mapping between models and real-world phenomena. This paper introduces, for the first time, an axiomatic definition of mathematical models grounded in Hilbert-space operator theory: a model is formalized as a computable operator acting on random variables, systematically unifying theoretical derivation, experimental implementation, and statistical identification. We further establish a geometric correspondence between the model manifold and the prediction surface, exposing intrinsic structural properties and the fundamental nature of model computability. This framework fills a critical gap in the formal ontology of modeling, providing a unified mathematical foundation for interdisciplinary model construction. It significantly enhances the logical rigor of theoretical inference and the reliability of empirical validation.

1 citationsRead paper

Do Stack Overflow Answer Edits Occur Beyond Java? A Replication on Python and JavaScript

Aug 08, 2026

This study investigates the prevalence of answer editing on Stack Overflow in the context of Python and JavaScript and evaluates its impact on code improvement in open-source projects. Extending prior work focused on Java, the authors replicate the methodology using the SOTorrent dataset, a Siamese-based code clone detection tool, and GitHub project popularity tiers to analyze these two widely used languages for the first time. The findings reveal that 41.25% of accepted Python answers and 39.10% of accepted JavaScript answers have been edited—approximately 2.5 times higher than the rate observed for Java. Moreover, edited code snippets show significantly increased reuse in high-popularity projects, with the effect being statistically significant for Python. These results demonstrate the cross-language generality of answer editing and suggest its heightened potential for enhancing code quality.

0 citationsRead paper

An Agentic AI Framework Overcomes Fundamental Limitations of Large Language Models for Glaucoma Detection from Fundus Photography

Aug 07, 2026

This study addresses critical limitations of existing large language models in glaucoma detection—namely hallucination, low accuracy, and inconsistent outputs—by introducing the first multi-agent collaborative framework tailored for glaucoma screening. The proposed approach employs a three-stage pipeline of assessment, tool invocation, and reflective integration, synergistically combining multimodal reasoning through language models (Gemini 2.5 Flash, GPT-5.4 mini) and specialized vision models (QAModel, SwinV2-Tiny, SegFormer-B0). This framework substantially enhances diagnostic performance: classification accuracy reaches 88%, matching that of ophthalmologists and surpassing baseline methods by 16–47 percentage points; cup-to-disc ratio estimation error is reduced by 15–50%, with correlation to expert ratings improving from weak to moderate–strong (r = 0.59–0.84); and inter-run consistency is markedly improved (κ = 0.96), effectively mitigating overdiagnosis and stochasticity inherent in single-model approaches.

0 citationsRead paper

Pulse Graphs: Prime-Activated Boolean Dynamics on Directed Graphs

Jul 11, 2026

This study investigates synchronous Boolean dynamical systems on finite acyclic directed graphs governed by a prime activation rule—where a node becomes active at the next time step if and only if the number of its active in-neighbors is prime. Introducing prime-counting logic into Boolean networks for the first time, the work combines combinatorial constructions, maximal-length affine feedback shift registers, mean-field approximations, and a prime–Poisson mapping to characterize system dynamics. Key contributions include establishing that the maximum attractor period \(L(n)\) for an \(n\)-node system satisfies \(2^{n-3}-1 \leq L(n) \leq 2^n-1\), deriving exact update rules and attractor classifications for complete graphs, and uncovering a non-degenerate fold bifurcation in sparse random graphs at a critical average degree \(c^* \approx 3.825\), indicative of a pronounced phase transition.

0 citationsRead paper

From Classification to Localization and Clinical Validation: Large-Scale Development of a Deep Learning System for Thoracic Disease Detection on Chest Radiographs in Thailand

Jul 10, 2026

This study addresses the critical shortage of radiologists in Thailand and Southeast Asia, which severely limits the capacity for interpreting chest X-rays (CXRs), by proposing Inspectra CXR v5—the first end-to-end deep learning system deployed in a large-scale Thai clinical setting that integrates multi-label disease classification with weakly supervised lesion localization. Built upon a DenseNet-121 backbone, the system introduces an innovative Attend-and-Compare module and a probabilistic class activation map (PCAM) aggregation layer, trained on over 870,000 report-annotated CXRs. It achieves an average in-house AUROC of 0.994 and demonstrates robust generalization across 13 hospitals with an AUROC of 0.970. Lesion localization accuracy reaches 77.9%, while radiologist evaluations confirm high consistency in both classification (93.6%) and localization (94.7%), with a system usability score of 89.

0 citationsRead paper

Recent publications

Latest Papers

Do Stack Overflow Answer Edits Occur Beyond Java? A Replication on Python and JavaScript

Aug 08, 2026

This study investigates the prevalence of answer editing on Stack Overflow in the context of Python and JavaScript and evaluates its impact on code improvement in open-source projects. Extending prior work focused on Java, the authors replicate the methodology using the SOTorrent dataset, a Siamese-based code clone detection tool, and GitHub project popularity tiers to analyze these two widely used languages for the first time. The findings reveal that 41.25% of accepted Python answers and 39.10% of accepted JavaScript answers have been edited—approximately 2.5 times higher than the rate observed for Java. Moreover, edited code snippets show significantly increased reuse in high-popularity projects, with the effect being statistically significant for Python. These results demonstrate the cross-language generality of answer editing and suggest its heightened potential for enhancing code quality.

0 citationsRead paper

An Agentic AI Framework Overcomes Fundamental Limitations of Large Language Models for Glaucoma Detection from Fundus Photography

Aug 07, 2026

This study addresses critical limitations of existing large language models in glaucoma detection—namely hallucination, low accuracy, and inconsistent outputs—by introducing the first multi-agent collaborative framework tailored for glaucoma screening. The proposed approach employs a three-stage pipeline of assessment, tool invocation, and reflective integration, synergistically combining multimodal reasoning through language models (Gemini 2.5 Flash, GPT-5.4 mini) and specialized vision models (QAModel, SwinV2-Tiny, SegFormer-B0). This framework substantially enhances diagnostic performance: classification accuracy reaches 88%, matching that of ophthalmologists and surpassing baseline methods by 16–47 percentage points; cup-to-disc ratio estimation error is reduced by 15–50%, with correlation to expert ratings improving from weak to moderate–strong (r = 0.59–0.84); and inter-run consistency is markedly improved (κ = 0.96), effectively mitigating overdiagnosis and stochasticity inherent in single-model approaches.

0 citationsRead paper

Pulse Graphs: Prime-Activated Boolean Dynamics on Directed Graphs

Jul 11, 2026

This study investigates synchronous Boolean dynamical systems on finite acyclic directed graphs governed by a prime activation rule—where a node becomes active at the next time step if and only if the number of its active in-neighbors is prime. Introducing prime-counting logic into Boolean networks for the first time, the work combines combinatorial constructions, maximal-length affine feedback shift registers, mean-field approximations, and a prime–Poisson mapping to characterize system dynamics. Key contributions include establishing that the maximum attractor period \(L(n)\) for an \(n\)-node system satisfies \(2^{n-3}-1 \leq L(n) \leq 2^n-1\), deriving exact update rules and attractor classifications for complete graphs, and uncovering a non-degenerate fold bifurcation in sparse random graphs at a critical average degree \(c^* \approx 3.825\), indicative of a pronounced phase transition.

0 citationsRead paper

From Classification to Localization and Clinical Validation: Large-Scale Development of a Deep Learning System for Thoracic Disease Detection on Chest Radiographs in Thailand

Jul 10, 2026

This study addresses the critical shortage of radiologists in Thailand and Southeast Asia, which severely limits the capacity for interpreting chest X-rays (CXRs), by proposing Inspectra CXR v5—the first end-to-end deep learning system deployed in a large-scale Thai clinical setting that integrates multi-label disease classification with weakly supervised lesion localization. Built upon a DenseNet-121 backbone, the system introduces an innovative Attend-and-Compare module and a probabilistic class activation map (PCAM) aggregation layer, trained on over 870,000 report-annotated CXRs. It achieves an average in-house AUROC of 0.994 and demonstrates robust generalization across 13 hospitals with an AUROC of 0.970. Lesion localization accuracy reaches 77.9%, while radiologist evaluations confirm high consistency in both classification (93.6%) and localization (94.7%), with a system usability score of 89.

0 citationsRead paper

RADIANT-PET: Reasoning-Augmented PET/CT Lesion Segmentation with Large Language Models and Reinforcement Learning

Jun 23, 2026

This work addresses the challenge of false positives in PET/CT lesion segmentation caused by physiological uptake and imaging artifacts. The authors propose a novel paradigm that integrates voxel-level segmentation with lesion-level reasoning via a large language model (LLM). Initially, permissive candidate regions are generated and converted into structured textual descriptions, which are then evaluated by the LLM in conjunction with multimodal clinical context—such as radiology reports—to determine lesion validity. The approach innovatively leverages the LLM for anatomical semantic reasoning and employs Group Relative Policy Optimization, a reinforcement learning strategy, to enhance the model’s consistency and accuracy. Evaluated on the AutoPET and OSU test sets, the method significantly outperforms purely image-based approaches, demonstrating the greatest improvements in false positive suppression and clinical consistency when radiology reports are available.

0 citationsRead paper