Embedded Files
Research Team
Research Team
Research Mentor:
Dr. Shuying Li
Team members:
Chloe Mick (Primary Presenter)
Hafsa El Harchi
Kayla Croft
Ashton Householder
Mikey Mannor
Abstract/Introduction
Abstract/Introduction
Neurodegeneration is characterized by damage to neural structures, which often leads to impaired function. Early detection is a major challenge due to microstructural changes in neural and vascular tissue that are difficult to detect using traditional human-based analytical methods. This study focuses on using artificial intelligence (AI) to better understand and detect these changes in high-resolution brain-scanning images. To investigate this, the cerebral vasculature and the surroundings were analyzed using QuPath. High-resolution images were manually annotated to create a training dataset for an artificial neural network (ANN). The annotations allowed the AI to learn to detect vasculature across the entire unannotated high-resolution images. Through iterative training, the AI model was designed to accurately recognize and segment micro blood vessels in human brain images at scale, which can be used to identify key vasculature markers in human brain due to neurodegeneration. The end goal is to develop an AI system trained to reliably detect early vascular markers of neurodegeneration in medical imaging, enabling earlier diagnosis and more effective, targeted treatment planning.
Research Questions
Research Questions
What vascular differences exist between healthy and neurodegenerative brain tissue?
Can AI-driven automated vessel detection assist with recognizing patterns associated with neurodegenerative disease?
How reliable is our model when applied to new unclassified images?
Methods
Methods
This study follows a structured workflow to develop and evaluate an AI-based pipeline for detecting blood vessels in human brain tissue images (neuropathy whole slide images).
First, large-scale brain scan images were curated and imported into QuPath AI (our chosen analysis program). Blood vessels within these images were manually annotated, and key cerebral structures were identified to provide heightened accuracy. These annotations were then compiled into a labeled dataset used for model training.
Currently, the AI model is being trained in segmentation using the labeled dataset. The model’s performance is being optimized to improve detection accuracy, followed by validation on a separate set of labeled images to assess reliability and consistency.
Finally, the trained model will be applied to unannotated images to automatically detect blood vessels. The results will be used to compare vascular features between healthy and diseased brain tissues, enabling further analysis of structural differences and supporting the development of research conclusions.
Our Model vs. Generative AI
Our Model vs. Generative AI
What's the Real Difference?
Generative AI Models (e.g., large language models like ChatGPT) are designed to create new content. These models learn from large data sets and generate text, images, and other outputs that esemble the training distribution. These systems are primarily for producing content. They are not optimized for pixel-level accuracy, spatial consistency, or quantitative validation required in biomedical image analysis. (Stryker et al., 2023)
Our model is very different; It is a task specific model used to detect and segment blood vessels in high resolution neural images. Instead of generating new content, the model predicts spatially resolved labels after trained using human-annotated ground truth (e.g., vessel masks generated in QuPath). The goal here is precision, consistency and biological interpretability, not creativity.
Downsides of Generative AI Models in Healthcare
Generative AI has many limitations in the healthcare realm, and reasonably so, since it's intended purpose is not to analyze and label pixel-level medical imaging. However, this is a big deal; It can produce factually incorrect information, lack consistency, and introduce bias based on patterns rather than objective structure. (Holmes et al., 2025; Templin et al., 2024)
How Our Model Works to Avoid These Issues
Our model is trained only on the verified training data we have provided.
It performs detection and segmentation ONLY on existing images, meaning there will be no risk of fabricated structures.
Once trained, it will produce consistent results for the same input, as tested through image repetition.
It focuses specifically on microlevel neural vasculature, which other existing AI models are not trained to handle
Why This Matters
For early detection of neurodegenerative disease, small and subtle structural changes in brain vasculature matter. These require quantitative, high accuracy, and repeatable results. Looking ahead, a specialized model like ours will provide controlled, measurable outputs that can support medical analysis, while generative AI lacks the reliability needed for this type of task.
Generative AI plays an important role in communication, content creation, and accessibility, but our model serves a different purpose by delivering precise, repeatable, and quantitative analysis of real biological structure, which is essential for detecting subtle changes in brain structure and advancing earlier, more reliable medical diagnosis.
Conclusions and Future Study
Conclusions and Future Study
Early Conclusions
AI-based vessel segmentation using the annotated dataset shows promising early performance, demonstrating that the model can begin to identify and map microvasculature in high resolution brain images while supporting more scalable analysis than manual methods
Future Study
More annotations with varying stains
Correlate vessel characteristics (e.g., size, tortuosity, density per tissue area) with disease types and stages
Use attention-based multiple instance learning (MIL) framework to identify disease-specific vascular patterns
Research poster presented at the 2026 Undergraduate Research Forum showing an AI based study on detecting cerebral blood vessels in brain tissue images, including sections on methodology, training data, model outputs, and future research directions.
Acknowledgements
Acknowledgements
We would like to give a special thanks to our advisor, Dr. Shuying Li, for all of her guidance, feedback, and support throughout this project. We could not have asked for a kinder nor more knowledgeable research lead.
Additional thanks to Miami University and the Undergraduate Research Department for providing us with the opportunities and resources to get involved in meaningful research early on in our career.
References
References
Li, S., et al. Age-informed, attention-based weakly supervised learning for neuropathological image assessment. Brain Informatics (2025).
Bankhead, P. et al. QuPath: Open source software for digital pathology image analysis. Scientific Reports (2017).
Holmes, S. A., Faria, V., & Moulton, E. A. (2025). Generative AI in healthcare: challenges to patient agency and ethical implications. Frontiers in Digital Health, 7. https://doi.org/10.3389/fdgth.2025.1524553
Templin, T., Perez, M. W., Sylvia, S., Leek, J., & Nasa Sinnott-Armstrong. (2024). Addressing 6 challenges in generative AI for digital health: A scoping review. PLOS Digital Health, 3(5), e0000503–e0000503. https://doi.org/10.1371/journal.pdig.0000503
Stryker, C., & Scapicchio, M. (2023). What is generative AI? Ibm.com. https://www.ibm.com/think/topics/generative-ai
Writing support tools were used for grammar and clarity. The model did not generate original research findings. All data and claims are based on verified sources.
NACE Career Readiness Competencies
NACE Career Readiness Competencies
Critical Thinking
Critical Thinking
NACE defines critical thinking as the ability to analyze information and make decisions based on context and reasoning.
Through our research, we applied critical thinking by analyzing complex brain tissue images and identifying patterns in vascular structures. We gathered and interpreted data from multiple brain scans, trained an AI model, evaluated the accuracy of said AI classification model, and adjusted our approach when results were inconsistent. This required making evidence-based decisions and clearly justifying our reasoning to improve outcomes.
Teamwork
Teamwork
NACE defines teamwork as building collaborative relationships and working effectively toward shared goals while valuing different perspectives.
We all developed teamwork skills by collaborating with the research team to design and refine our project. We contributed to discussions, shared findings, and incorporated feedback from our advisor and peers. This experience taught us how to divide responsibilities, stay accountable, and work toward a common goal while respecting different ideas and approaches.
Technology
Technology
NACE defines technology as the ability to use and adapt tools to improve efficiency and accomplish goals.
Our research strengthened our technology skills through the use of tools like QuPath for AI-based image analysis. We trained classifiers to detect blood vessels in brain tissue and worked with large datasets to evaluate model performance. This experience improved our ability to apply technology to solve real-world problems and adapt to new tools quickly.
Page updated
Report abuse