Fakrul Islam Tushar

About Fakrul Islam Tushar

I am Fakrul Islam Tushar, PhD, an Assistant Research Professor and researcher in…

Experience

1. Graduate Research Assistant

   Duke Center For Virtual Imaging Trials

   Aug 2021 — Present · Durham, NC, US

   AI-Powered Clinical Trial Emulation: Led a cross-disciplinary team to develop the first in silico emulationof the National Lungs Screening Trial, integrating simulated patients and AI readers to replicate end-to-end screening, matching clinical performance while reducing trial duration from 10 years to hours andsaving ≈ $300M. Resulted in multiple first-author publications, open-access releases, award recognitions, and an invited talk. Tushar et al, Medical Image Analysis (2025), RSNA(2024); VITM (2024); SPIE (2024); Project-page: fitushar.github.io/VLST.github.io/.• AI–Human Data Curation, Task-Aware Pretraining, and Benchmarking: Co-led the DLCS dataset cu-ration via human–AI collaboration. open-access AI benchmarking efforts, benchmarked foundation models (self-supervised and transfer learning), and proposed SWS++, a task-relevant pretraining strategy requiring no additional data. Achieved state-of-the-art AUCs (up to 0.90) across public datasets. Tushar et al, arXiv (2024); Wang, Tushar et al, Radiology: Artificial Intelligence (2025).Dataset: zenodo.org/records/13••••69; Code: github.com/fitushar/AI-in-Lung-Health-Benchmarking-Detection-and-Diagnostic-Models-Across-Multiple-CT-Scan-Datasets.• Simulation for Data-Limited AI: Demonstrated that anatomy-informed simulated data im-proves AI model performance in low-data settings, outperforming current state-of-the-art by +10% in detection and +2–9% in classification and segmentation tasks. Tushar et al, arXiv (2025). Code: github.com/fitushar/SYN-LUNGS.• Consensus-Based Labeling with Domain Adaptation and VLM Verification: Developed an inference-time framework combining ensemble consensus, affine calibration, and unsupervised domain adaptation to generate high-confidence pseudo-labels. Applied cost-aware thresholding and used LLaVA-Med to add an extra layer of validation. Project-page: fitushar.github.io/ReFINELung.github.io/• Generative Model for Population-Scale Synthetic Data (on-going)

Education

• Duke University

  Medical Imaging, Biomedical/Medical Engineering

  2019 — 2019

• Université Bourgogne Europe

  Master's in Medical Imaging And Applications , Biomedical/Medical Engineering

  2017 — 2018

• Jatrabari Ideal High School

  Secondary School Certificate (SSC), Science Group

  2010 — 2010

• Universitat de Girona

  Master's in Medical Imaging And Applications, Biomedical/Medical Engineering

  2018 — 2019

• Dania College

  Higher Secondary Certificate (HSC), Science Group

  2010 — 2012

• Università degli Studi di Cassino e del Lazio Meridionale

  Master's in Medical Imaging And Applications, Biomedical/Medical Engineering

  2018 — 2018

• Duke University Pratt School of Engineering

  Doctor of Philosophy - PhD, Electrical and Computer Engineering

• American International University-Bangladesh

  Bachelor of Science, Electrical and Electronics Engineering

Skills

• Teaching
• Time Management
• Public Speaking
• Social Media
• Powerpoint
• Microsoft Excel
• Research
• Microsoft Powerpoint
• Marketing
• Strategic Planning
• Public Relations
• Microsoft Office
• Event Planning
• Event Management
• Proteus
• Ni Multisim
• Microsoft Word
• Team Leadership
• Social Networking
• Leadership
• Teamwork
• Management
• Photography