Bardia Khosravi, MD, MPH, MHPE

My research centers on developing artificial intelligence solutions that address fundamental challenges in medical imaging, with a sustained focus on generative AI, bias mitigation, and clinical translation. Throughout my work, I have maintained a consistent vision: creating AI systems that enhance radiological practice while ensuring equity and explainability in clinical deployment.

Core Research Philosophy: Generative AI for Medical Imaging

At the heart of my research lies the development and application of generative artificial intelligence for medical imaging challenges. Beginning with my post-doctoral fellowship at Mayo Clinic, I recognized that traditional supervised learning approaches in medical AI were limited by data scarcity, privacy constraints, and representation bias. This insight led me to pioneer the use of generative adversarial networks (GANs) and later diffusion models for creating high-fidelity synthetic medical images, particularly in orthopedic imaging.

My work on synthetic pelvis radiographs demonstrated a paradigm shift: AI models could be trained effectively on synthetic data without compromising performance while completely eliminating patient privacy concerns. This foundational work has evolved into more sophisticated applications, including few-shot biomedical image segmentation using diffusion models, where high-quality results can be achieved with as few as 20 training images. These advances address the critical challenge of deploying AI in specialized medical domains where large labeled datasets are unavailable.

Clinical Translation and Practical Applications

My research consistently emphasizes real-world clinical utility. Working closely with orthopedic surgeons, I have developed several deployed clinical tools, including automated implant detection systems (THAAID) and patient-specific hip arthroplasty dislocation risk calculators. These projects demonstrate my ability to translate complex AI methodologies into clinically actionable tools that directly impact patient care.

The dislocation risk calculator exemplifies my approach to explainable AI in clinical practice, providing both accurate predictions and interpretable reasoning for surgical decision-making. Similarly, my work on automated radiographic analysis pipelines and anatomical landmark detection tools addresses routine clinical workflow challenges while maintaining the interpretability that clinicians require.

Addressing Bias and Ensuring Equity

A significant thread throughout my research addresses the critical issue of bias in medical AI systems. Through systematic research published in Radiology: Artificial Intelligence, I have developed comprehensive frameworks for identifying and mitigating bias in data handling, model development, and performance evaluation. This work extends beyond technical solutions to address fundamental questions about representation and fairness in medical AI.

My research on racial disparities in imaging datasets using generative AI provides practical tools for the research community to build more inclusive AI models. By developing methods to analyze and visualize demographic biases in medical imaging datasets, this work contributes to building more equitable healthcare AI systems.

Evolution Toward Multimodal AI Systems

My current research represents a natural evolution of my generative AI expertise toward multimodal systems that integrate imaging with clinical text. The development of HIPAA-compliant open-source large language models for radiology report annotation addresses practical deployment challenges while maintaining privacy compliance. This work provides systematic analysis of model size, prompting strategies, and pathology complexity, offering practical guidance for implementing language models in clinical radiology workflows.

The integration of vision-language models (VLMs) for radiology represents the current frontier of my research. These multimodal systems combine sophisticated image analysis with natural language understanding, enabling more comprehensive and contextually aware medical image interpretation. This direction builds directly on my foundational work in generative AI while addressing the growing need for AI systems that can reason about both visual and textual medical information.

Methodological Contributions and Open Science

Throughout my research program, I have maintained a strong commitment to methodological rigor and reproducible research. My contribution to the RIDGE framework for medical image segmentation model evaluation provides standardized metrics for assessing model performance across diverse clinical scenarios. Additionally, my development of MIDeL.org, a free educational platform for teaching deep learning in medical imaging, reflects my commitment to democratizing AI education in healthcare.

My technical contributions include multiple open-source packages for medical image generation and labeling, as well as sophisticated anonymization tools that address privacy concerns while enabling large-scale AI research. These tools use advanced algorithms to automatically remove identifying information from medical images, facilitating dataset creation while maintaining patient privacy.

Future Vision

My research continues to evolve toward AI systems that not only match human performance in medical image interpretation but provide comprehensive, explainable, and contextually aware analyses. The integration of generative AI with multimodal reasoning represents a particularly promising direction, enabling AI systems that can both generate synthetic training data and provide sophisticated clinical insights.

The ultimate goal of my research program is to develop AI systems that enhance rather than replace human expertise in radiology, providing tools that make radiologists more efficient and accurate while ensuring equitable access to high-quality medical imaging interpretation across diverse patient populations. This vision maintains continuity with my foundational work in generative AI while addressing the evolving needs of modern radiology practice.