A Vision-Language Foundation Model for Precise and Comprehensive Brain Tumor Diagnosis from Preoperative Multimodal Data
研究开发了BrainVLM模型，利用MRI等多模态数据自动分类12种脑肿瘤，并提供诊断不确定性量化及报告生成，以改善术前非侵入性脑肿瘤诊断。
0 citationsRead paper
Selected work
Representative Papers
Urban Boundaries, Social Barriers: A Benchmark and Vision-Centric Framework for Mapping Gated Communities and Equity Implications
研究通过构建GBA-GCs基准和MCGC框架，利用多模态数据识别中国大湾区的封闭与开放社区，以解决城市尺度上此类社区识别及公平性分析的问题。
0 citationsRead paper
Primitive-Driven Compositional Forensic Visual Prompting for Open-World Face Anti-Spoofing
本文提出一种基于视觉特征空间的组合法医视觉提示学习框架，以解决开放世界面部反欺骗中的未知攻击问题。
0 citationsRead paper
Minimum Complete MR Subsets under Semantic-Mutation Fault Models: A Support-Set Domination Boundary
This study investigates when mutation testing requires selecting a subset of metamorphic relations (MRs) based on concrete mutants—rather than merely counting fault classes—to satisfy minimal completeness evidence requirements. To this end, the authors propose a “layer-wise relative completeness” criterion and introduce a dominance boundary theory driven by heterogeneity in killing signatures, thereby decoupling MR-specific concerns from conventional fault-class statistics. They define a scope-based fault signature kernel and, leveraging a set cover formulation, greedy approximation, integer linear programming, and SMS rank analysis—augmented with artifact channels and path-witness mechanisms—prove that the Min-MR-Complete problem is NP-hard, establish a logarithmic approximation bound, and provide both exact and approximate solution methods. Path witnesses further validate the efficacy of the boundary theorem under both collapsed and non-collapsed scenarios.
0 citationsRead paper
NOETHER: A Constructive Framework for Metamorphic Pattern Discovery from Operator Algebras
This work addresses three fundamental challenges in metamorphic testing—namely, the origin, closure, and transferability of metamorphic relations (MRs)—by introducing the NOETHER framework. NOETHER combines an upstream eight-module decomposition grounded in operator algebra (encompassing symmetry, order structure, self-adjointness, and related properties) with a downstream CONSTRUCT-MP algorithm to automatically derive a set of MetaPatterns from programs. These MetaPatterns enjoy algebraic closure and polynomial-time decidability, elevating MR induction to a domain-level algebraic abstraction and enabling a deductive, mechanized approach to MR generation. For the first time, this method provides formal theoretical guarantees for both closure and decidability. Empirical validation across reactor physics, equivariant machine learning, and relational query optimization demonstrates systematic reconstruction of known MRs, synthesis of executable MRs, and verification of core predictions, while counterexamples refute the conjecture of absolute completeness and reveal five dimensions for extending the Translate framework.
0 citationsRead paper
Recent publications
Latest Papers
A Vision-Language Foundation Model for Precise and Comprehensive Brain Tumor Diagnosis from Preoperative Multimodal Data
研究开发了BrainVLM模型，利用MRI等多模态数据自动分类12种脑肿瘤，并提供诊断不确定性量化及报告生成，以改善术前非侵入性脑肿瘤诊断。
0 citationsRead paper
Urban Boundaries, Social Barriers: A Benchmark and Vision-Centric Framework for Mapping Gated Communities and Equity Implications
研究通过构建GBA-GCs基准和MCGC框架，利用多模态数据识别中国大湾区的封闭与开放社区，以解决城市尺度上此类社区识别及公平性分析的问题。
0 citationsRead paper
Primitive-Driven Compositional Forensic Visual Prompting for Open-World Face Anti-Spoofing
本文提出一种基于视觉特征空间的组合法医视觉提示学习框架，以解决开放世界面部反欺骗中的未知攻击问题。
0 citationsRead paper
Minimum Complete MR Subsets under Semantic-Mutation Fault Models: A Support-Set Domination Boundary
This study investigates when mutation testing requires selecting a subset of metamorphic relations (MRs) based on concrete mutants—rather than merely counting fault classes—to satisfy minimal completeness evidence requirements. To this end, the authors propose a “layer-wise relative completeness” criterion and introduce a dominance boundary theory driven by heterogeneity in killing signatures, thereby decoupling MR-specific concerns from conventional fault-class statistics. They define a scope-based fault signature kernel and, leveraging a set cover formulation, greedy approximation, integer linear programming, and SMS rank analysis—augmented with artifact channels and path-witness mechanisms—prove that the Min-MR-Complete problem is NP-hard, establish a logarithmic approximation bound, and provide both exact and approximate solution methods. Path witnesses further validate the efficacy of the boundary theorem under both collapsed and non-collapsed scenarios.
0 citationsRead paper
NOETHER: A Constructive Framework for Metamorphic Pattern Discovery from Operator Algebras
This work addresses three fundamental challenges in metamorphic testing—namely, the origin, closure, and transferability of metamorphic relations (MRs)—by introducing the NOETHER framework. NOETHER combines an upstream eight-module decomposition grounded in operator algebra (encompassing symmetry, order structure, self-adjointness, and related properties) with a downstream CONSTRUCT-MP algorithm to automatically derive a set of MetaPatterns from programs. These MetaPatterns enjoy algebraic closure and polynomial-time decidability, elevating MR induction to a domain-level algebraic abstraction and enabling a deductive, mechanized approach to MR generation. For the first time, this method provides formal theoretical guarantees for both closure and decidability. Empirical validation across reactor physics, equivariant machine learning, and relational query optimization demonstrates systematic reconstruction of known MRs, synthesis of executable MRs, and verification of core predictions, while counterexamples refute the conjecture of absolute completeness and reveal five dimensions for extending the Translate framework.
0 citationsRead paper