On the role of the tokenizer in ECG transformer models
研究对比了八种ECG信号的分词策略，通过调整Transformer模型输入，发现与心电图形态对齐的分词方法能提高预测性能和内存效率。
0 citationsRead paper
Selected work
Representative Papers
Designing a Robust LLM-Based Evaluation System for Agentic AI in Drug Discovery Through Human Alignment
为解决药物发现中AI助手输出评估难题，本文提出基于LLM的评价框架，通过定义评估维度、验证与人类专家的一致性及优化LLM裁判来提高评估准确性。
0 citationsRead paper
Static analysis-guided agentic AI translation enables Rust as a full stack bioinformatics language
This study addresses the longstanding challenges in bioinformatics stemming from legacy codebases written in outdated languages such as Perl and Fortran, which suffer from poor maintainability, security vulnerabilities, and limited compatibility with modern hardware. To overcome these limitations, this work proposes a novel automated migration framework that synergistically combines static program analysis with AI-driven agents to efficiently refactor legacy tools into memory-safe, high-performance Rust implementations. The approach substantially reduces refactoring costs and advances Rust as a viable full-stack language for bioinformatics development. Demonstrated on the Bascet software, the refactored version achieves an approximately 80-fold reduction in code size, a 10-fold decrease in build time, over 3× improvement in critical performance metrics, and native Windows support—eliminating reliance on containers or Unix-like environments.
0 citationsRead paper
Beyond Isotropic Assumptions: Continuity-Constrained Segmentation and GPU Morphometry for Nanoscale GBM Analysis
This study addresses structural discontinuities in anisotropic microscopy images caused by insufficient axial sampling by proposing an end-to-end, GPU-accelerated segmentation framework that requires no dense 3D annotations. The method trains 3D U-Net or SwinUNETR models directly on raw anisotropic volumes, incorporating random rotation-based data augmentation and a Z-axis continuity loss to enhance inter-slice coherence. Accurate quantification of membrane thickness is achieved through Gaussian consensus fusion combined with a point spread function–corrected, GPU-accelerated ray-surface intersection algorithm. This work presents the first fully automated, image-restoration-free morphometric analysis of nanoscale glomerular basement membranes (GBM), attaining segmentation accuracy comparable to inter-expert agreement, substantially suppressing staircasing artifacts, improving reconstruction smoothness, and successfully detecting disease-associated GBM thickening.
0 citationsRead paper
Prostate biopsy whole slide image dataset from an underrepresented Middle Eastern population
Publicly available whole-slide imaging (WSI) datasets for histopathology are predominantly derived from Western populations, with severe underrepresentation of regions such as the Middle East—where digital pathology infrastructure remains limited—thereby hindering the cross-population generalizability of AI models. Method: We introduce the first prostate needle biopsy WSI dataset from Erbil, Iraq, comprising 339 WSIs from 185 patients, acquired in native formats using Leica, Hamamatsu, and Grundium scanners. All slides were independently annotated by three board-certified pathologists for Gleason score and ISUP grade group, followed by rigorous de-identification. Contribution/Results: This dataset fills a critical gap in Middle Eastern digital prostate pathology. It enables robust cross-scanner evaluation, color normalization research, and multi-expert inter-observer agreement analysis. Released under the CC BY 4.0 license via BioImage Archive, it significantly enhances reproducibility and validation of AI models across diverse global populations and heterogeneous scanning platforms.
0 citationsRead paper
Recent publications
Latest Papers
On the role of the tokenizer in ECG transformer models
研究对比了八种ECG信号的分词策略，通过调整Transformer模型输入，发现与心电图形态对齐的分词方法能提高预测性能和内存效率。
0 citationsRead paper
Designing a Robust LLM-Based Evaluation System for Agentic AI in Drug Discovery Through Human Alignment
为解决药物发现中AI助手输出评估难题，本文提出基于LLM的评价框架，通过定义评估维度、验证与人类专家的一致性及优化LLM裁判来提高评估准确性。
0 citationsRead paper
Static analysis-guided agentic AI translation enables Rust as a full stack bioinformatics language
This study addresses the longstanding challenges in bioinformatics stemming from legacy codebases written in outdated languages such as Perl and Fortran, which suffer from poor maintainability, security vulnerabilities, and limited compatibility with modern hardware. To overcome these limitations, this work proposes a novel automated migration framework that synergistically combines static program analysis with AI-driven agents to efficiently refactor legacy tools into memory-safe, high-performance Rust implementations. The approach substantially reduces refactoring costs and advances Rust as a viable full-stack language for bioinformatics development. Demonstrated on the Bascet software, the refactored version achieves an approximately 80-fold reduction in code size, a 10-fold decrease in build time, over 3× improvement in critical performance metrics, and native Windows support—eliminating reliance on containers or Unix-like environments.
0 citationsRead paper
Beyond Isotropic Assumptions: Continuity-Constrained Segmentation and GPU Morphometry for Nanoscale GBM Analysis
This study addresses structural discontinuities in anisotropic microscopy images caused by insufficient axial sampling by proposing an end-to-end, GPU-accelerated segmentation framework that requires no dense 3D annotations. The method trains 3D U-Net or SwinUNETR models directly on raw anisotropic volumes, incorporating random rotation-based data augmentation and a Z-axis continuity loss to enhance inter-slice coherence. Accurate quantification of membrane thickness is achieved through Gaussian consensus fusion combined with a point spread function–corrected, GPU-accelerated ray-surface intersection algorithm. This work presents the first fully automated, image-restoration-free morphometric analysis of nanoscale glomerular basement membranes (GBM), attaining segmentation accuracy comparable to inter-expert agreement, substantially suppressing staircasing artifacts, improving reconstruction smoothness, and successfully detecting disease-associated GBM thickening.
0 citationsRead paper
Prostate biopsy whole slide image dataset from an underrepresented Middle Eastern population
Publicly available whole-slide imaging (WSI) datasets for histopathology are predominantly derived from Western populations, with severe underrepresentation of regions such as the Middle East—where digital pathology infrastructure remains limited—thereby hindering the cross-population generalizability of AI models. Method: We introduce the first prostate needle biopsy WSI dataset from Erbil, Iraq, comprising 339 WSIs from 185 patients, acquired in native formats using Leica, Hamamatsu, and Grundium scanners. All slides were independently annotated by three board-certified pathologists for Gleason score and ISUP grade group, followed by rigorous de-identification. Contribution/Results: This dataset fills a critical gap in Middle Eastern digital prostate pathology. It enables robust cross-scanner evaluation, color normalization research, and multi-expert inter-observer agreement analysis. Released under the CC BY 4.0 license via BioImage Archive, it significantly enhances reproducibility and validation of AI models across diverse global populations and heterogeneous scanning platforms.
0 citationsRead paper