Selected work

Representative Papers

SecondOpinion: Anatomy-Aware Gated Reasoning for Efficient Medical Image Analysis

Aug 03, 2026

This work addresses the inefficiency of conventional medical image analysis models that uniformly process all cases regardless of complexity. The authors propose SecondOpinion, a dual-stream framework wherein a primary stream rapidly handles all inputs, while an anatomically guided auxiliary stream is activated only when a learnable gating mechanism—GateKeeper—deems the initial prediction unreliable. Crucially, the gate is explicitly trained as a binary correctness classifier, enabling on-demand invocation of anatomical reasoning. Results on chest X-ray and pelvic fracture datasets demonstrate that the model matches or exceeds state-of-the-art performance, with auxiliary stream activation rates ranging from 9.23% to 45.71%, closely aligned with task difficulty. This approach significantly enhances computational efficiency and task adaptability.

0 citationsRead paper

How Good LLMs Are at Answering Bangla Medical Visual Questions? Dataset and Benchmarking

May 18, 2026

This study addresses the absence of benchmark datasets for Bengali medical visual question answering (MedVQA) by introducing BanglaMedVQA—the first low-resource MedVQA dataset comprising clinically validated image-question-answer triplets in Bengali. Leveraging this dataset, the authors conduct a systematic evaluation of prominent large language models and vision-language models, including GPT-4o mini, Gemini, and Gemma-3. Experimental results reveal that current models exhibit substantially lower performance on Bengali MedVQA compared to English benchmarks, with particularly poor accuracy on complex diagnostic questions. These findings highlight a critical gap in the capability of existing multimodal models to perform fine-grained medical reasoning in low-resource languages. This work establishes essential infrastructure and empirical evidence for advancing multilingual evaluation frameworks in medical artificial intelligence.

0 citationsRead paper

A Scoping Review of Deep Learning for Urban Visual Pollution and Proposal of a Real-Time Monitoring Framework with a Visual Pollution Index

Feb 10, 2026

This study addresses the lack of a unified, automated framework for detecting, evaluating, and standardizing urban visual pollution—a gap exacerbated by fragmented existing research and limited real-time applicability. Following the PRISMA-ScR guidelines, the authors systematically review 26 core studies to assess the performance of mainstream object detection models, including YOLO, Faster R-CNN, and EfficientDet, in identifying visual pollutants. The work proposes a novel integrated framework that combines a standardized pollutant classification system, a cross-city benchmark dataset, a generalizable deep learning model, and a visual pollution index. This framework enables real-time monitoring and quantitative assessment of urban visual pollution, thereby establishing a technical and methodological foundation for advancing sustainable urban aesthetics and enhancing residents’ well-being.

0 citationsRead paper

Route, Retrieve, Reflect, Repair: Self-Improving Agentic Framework for Visual Detection and Linguistic Reasoning in Medical Imaging

Jan 13, 2026

This work addresses the limitations of existing medical imaging systems—typically black-box models performing single-pass inference—by introducing R⁴, a novel self-improving multi-agent framework for medical vision-language tasks. R⁴ integrates four collaborative agents responsible for routing, retrieval, reflection, and repair, enabling dynamic prompt configuration, joint image-text generation, clinical error detection, and constraint-driven iterative refinement. The framework supports explainable inference, self-diagnosis of errors, and joint spatial-linguistic optimization without requiring model fine-tuning. Evaluated on chest X-ray data, R⁴ achieves a 1.7–2.5 point improvement in LLM-as-a-Judge report generation scores and a 2.5–3.5 percentage point gain in weakly supervised detection mAP50, significantly outperforming single vision-language model baselines.

0 citationsRead paper

Mina: A Multilingual LLM-Powered Legal Assistant Agent for Bangladesh for Empowering Access to Justice

Nov 04, 2025

Low-income populations in Bangladesh face persistent barriers to legal aid due to opaque legal language, non-transparent procedures, and prohibitive costs; existing AI tools lack Bangla language support and judicial system adaptation. Method: We propose Mina—the first full-stack legal large language model system designed for low-resource language environments—integrating multilingual embeddings, RAG-enhanced retrieval, chain-of-tool invocation, and context-aware dialogue to deliver a localized, bilingual (Bangla/English) intelligent legal assistant capable of legal information retrieval, reasoning, translation, and document generation. Contribution/Results: Mina achieves the first end-to-end adaptation of legal AI for low-resource languages—from knowledge modeling and inference architecture to real-world deployment. Evaluated on the Bangladesh Bar Council’s three-tier licensing examination, Mina attains 75–80% accuracy—matching or exceeding industry benchmarks—demonstrating robust legal-context comprehension and generation, thereby advancing equitable access to justice.

0 citationsRead paper

Recent publications

Latest Papers

SecondOpinion: Anatomy-Aware Gated Reasoning for Efficient Medical Image Analysis

Aug 03, 2026

This work addresses the inefficiency of conventional medical image analysis models that uniformly process all cases regardless of complexity. The authors propose SecondOpinion, a dual-stream framework wherein a primary stream rapidly handles all inputs, while an anatomically guided auxiliary stream is activated only when a learnable gating mechanism—GateKeeper—deems the initial prediction unreliable. Crucially, the gate is explicitly trained as a binary correctness classifier, enabling on-demand invocation of anatomical reasoning. Results on chest X-ray and pelvic fracture datasets demonstrate that the model matches or exceeds state-of-the-art performance, with auxiliary stream activation rates ranging from 9.23% to 45.71%, closely aligned with task difficulty. This approach significantly enhances computational efficiency and task adaptability.

0 citationsRead paper

How Good LLMs Are at Answering Bangla Medical Visual Questions? Dataset and Benchmarking

May 18, 2026

This study addresses the absence of benchmark datasets for Bengali medical visual question answering (MedVQA) by introducing BanglaMedVQA—the first low-resource MedVQA dataset comprising clinically validated image-question-answer triplets in Bengali. Leveraging this dataset, the authors conduct a systematic evaluation of prominent large language models and vision-language models, including GPT-4o mini, Gemini, and Gemma-3. Experimental results reveal that current models exhibit substantially lower performance on Bengali MedVQA compared to English benchmarks, with particularly poor accuracy on complex diagnostic questions. These findings highlight a critical gap in the capability of existing multimodal models to perform fine-grained medical reasoning in low-resource languages. This work establishes essential infrastructure and empirical evidence for advancing multilingual evaluation frameworks in medical artificial intelligence.

0 citationsRead paper

A Scoping Review of Deep Learning for Urban Visual Pollution and Proposal of a Real-Time Monitoring Framework with a Visual Pollution Index

Feb 10, 2026

This study addresses the lack of a unified, automated framework for detecting, evaluating, and standardizing urban visual pollution—a gap exacerbated by fragmented existing research and limited real-time applicability. Following the PRISMA-ScR guidelines, the authors systematically review 26 core studies to assess the performance of mainstream object detection models, including YOLO, Faster R-CNN, and EfficientDet, in identifying visual pollutants. The work proposes a novel integrated framework that combines a standardized pollutant classification system, a cross-city benchmark dataset, a generalizable deep learning model, and a visual pollution index. This framework enables real-time monitoring and quantitative assessment of urban visual pollution, thereby establishing a technical and methodological foundation for advancing sustainable urban aesthetics and enhancing residents’ well-being.

0 citationsRead paper

Route, Retrieve, Reflect, Repair: Self-Improving Agentic Framework for Visual Detection and Linguistic Reasoning in Medical Imaging

Jan 13, 2026

This work addresses the limitations of existing medical imaging systems—typically black-box models performing single-pass inference—by introducing R⁴, a novel self-improving multi-agent framework for medical vision-language tasks. R⁴ integrates four collaborative agents responsible for routing, retrieval, reflection, and repair, enabling dynamic prompt configuration, joint image-text generation, clinical error detection, and constraint-driven iterative refinement. The framework supports explainable inference, self-diagnosis of errors, and joint spatial-linguistic optimization without requiring model fine-tuning. Evaluated on chest X-ray data, R⁴ achieves a 1.7–2.5 point improvement in LLM-as-a-Judge report generation scores and a 2.5–3.5 percentage point gain in weakly supervised detection mAP50, significantly outperforming single vision-language model baselines.

0 citationsRead paper

Mina: A Multilingual LLM-Powered Legal Assistant Agent for Bangladesh for Empowering Access to Justice

Nov 04, 2025

Low-income populations in Bangladesh face persistent barriers to legal aid due to opaque legal language, non-transparent procedures, and prohibitive costs; existing AI tools lack Bangla language support and judicial system adaptation. Method: We propose Mina—the first full-stack legal large language model system designed for low-resource language environments—integrating multilingual embeddings, RAG-enhanced retrieval, chain-of-tool invocation, and context-aware dialogue to deliver a localized, bilingual (Bangla/English) intelligent legal assistant capable of legal information retrieval, reasoning, translation, and document generation. Contribution/Results: Mina achieves the first end-to-end adaptation of legal AI for low-resource languages—from knowledge modeling and inference architecture to real-world deployment. Evaluated on the Bangladesh Bar Council’s three-tier licensing examination, Mina attains 75–80% accuracy—matching or exceeding industry benchmarks—demonstrating robust legal-context comprehension and generation, thereby advancing equitable access to justice.

0 citationsRead paper