Health Informatics & Vision Foundation Models
My current research focuses on trustworthy machine learning and medical-image analysis for healthcare. As a Research Assistant, I contribute to explainable and uncertainty-aware analysis of panoramic radiographs, while my undergraduate thesis evaluates parameter-efficient adaptation of vision foundation models.
Undergraduate Thesis (Ongoing)
Parameter-Efficient Adaptation of Vision Foundation Models for Multi-Class Dental Radiograph Diagnosis
Description
To address the computational costs and representational gaps of applying Vision Foundation Models to dental radiography, this study leverages the DentIRO dataset to analyze embedding-space failure modes. It subsequently benchmarks parameter-efficient fine-tuning methods against full fine-tuning, evaluating model accuracy, robustness, and interpretability under data-scarce conditions.
Objective
This study evaluates frozen Vision Foundation Models (DINOv2, BiomedCLIP, MedSAM) via linear probing to quantify baseline performance gaps on subtle pathologies like caries. It benchmarks parameter-efficient fine-tuning strategies (LoRA, BitFit, VPT) against full fine-tuning while utilizing layer-depth ablations to isolate where localized dental decay features are best learned. Finally, the framework validates model robustness under label-scarce regimes (10–50% data) and confirms clinical alignment using transformer-adapted Grad-CAM heatmaps.
Motivation
Full fine-tuning of Vision Foundation Models on dental radiographs is computationally expensive and constrained by scarce data. This work addresses these domain gaps using parameter-efficient fine-tuning to achieve scalable, accurate, and interpretable dental diagnosis.
Core Research Pillars
Parameter-Efficient Vision Foundation Models
Evaluating and fine-tuning foundation models (DINOv2, BiomedCLIP, MedSAM) with LoRA, BitFit, and VPT for data-scarce medical imaging domains.
Pediatric Morphology & Uncertainty-Aware AI
Developing morphological alignment modules (PMAM) and uncertainty-quantified frameworks for explainable multi-disease detection in pediatric radiography.
Health Informatics & Hematological ML
Developing benchmarked, calibrated, and interpretable machine learning pipelines for complex clinical classification and disease detection.
Manuscripts & Conference Submissions
Includes 1 Health Conference & 3 NLP Conference Papers (Under Review)PMAM: A Pediatric Morphology Alignment Module for Explainable, Uncertainty-Aware Multi-Disease Detection in Panoramic Radiographs
Investigates a pediatric morphology alignment module (PMAM) to address anatomical growth variations and quantify predictive uncertainty for explainable, trustworthy multi-disease detection in dental panoramic radiographs.
Parameter-Efficient Adaptation of Vision Foundation Models for Multi-Class Dental Radiograph Diagnosis
Leverages the DentIRO dataset to analyze embedding-space failure modes of Vision Foundation Models in dental radiography, benchmarking parameter-efficient fine-tuning (LoRA, BitFit, VPT) against full fine-tuning under data-scarce conditions with Grad-CAM interpretability.
Detecting Compensated Microcytosis: A Benchmarked and Explainable Machine Learning Approach to Tri-Class Hematological Classification
A benchmarked and explainable machine learning approach to tri-class hematological classification for identifying compensated microcytosis with transparent biomarker feature attribution.
Bridging the Dialect Gap: Enhancing Bangla Dialect-to-English Machine Translation with Regional Metadata and Romanized Standard Bangla
Enhances regional Bangla dialect-to-English machine translation by integrating regional metadata injection and romanized standard Bangla intermediate representations.
Explainable Neural Machine Translation for Bangla Regional Dialect Normalization: A Multi-Model Comparative Study with Attention-Based Interpretability
A comprehensive multi-model comparative evaluation for normalizing Bangla regional dialects into standard Bangla, integrating attention-based interpretability and systematic error analysis.
Explainable and Parameter-Efficient Identification of Ethnic Languages in Shared Bengali Script under Low-Resource Conditions
Investigates parameter-efficient adaptation and explainability for ethnic and indigenous languages written in shared Bengali script under severe low-resource conditions.
Methodological Continuity: From Low-Resource NLP to Clinical AI
My experimental methodology originated in low-resource Natural Language Processing, where lack of standardized benchmarks required building custom corpora, handling dialect variations, and establishing rigorous multi-model evaluation frameworks.
In my current research at the Health Informatics Research Lab and in my undergraduate thesis, I translate these principles to clinical AI—where data scarcity demands parameter-efficient adaptation (LoRA, BitFit, VPT) and diagnostic stakes require rigorous feature attribution (SHAP, Grad-CAM).