Research

Publications and preprints from the lab.

2026

  • AortaGPT: An Interactive Vision-Language System for Aortic Image Analysis

    Gorkem Can Ates, Chang Wang, Harrison Chojnowski, Yu Xin, Hongxu Jiang, Walker R. Ueland, Muhammad Imran, Michael J. Fassler, Brian Fazzone, Jonathan Krebs, Griffin Stinson, Gilbert R. Upchurch Jr., Philip J. Hess Jr., Chang Zhao, Ruogu Fang, Michol A. Cooper, Wei Shao

    Under Review medical imagingvision-languageaortasegmentationLLM

    Accurate aortic diameter measurement is essential for risk stratification and surgical planning. AortaGPT is the first interactive vision-language system that enables clinicians to perform complex image processing, quantitative analysis, and reporting tasks through natural-language prompts. It integrates super-resolution, zone-based segmentation, and automated diameter measurement within a unified conversational interface, orchestrated by a LangGraph-based controller that enforces dependency-aware execution and supports multi-turn interaction.

    Summary → Code

  • Geometry-Aware Implicit Neural Reconstruction of Oblique Micro-Ultrasound Scans

    Harrison Chojnowski, Gorkem Can Ates, Wayne G. Brisbane, Wei Shao

    Under Review medical imagingimplicit neural representationultrasoundcomputer vision

    Micro-ultrasound is a high-resolution, low-cost modality for prostate cancer imaging, but its transrectal acquisition produces oblique slices at irregular angular intervals that lack a consistent anatomical frame of reference. We present a geometry-aware reconstruction framework: a coordinate-based sampling scheme that uses cylindrical acquisition geometry to accurately map each target voxel, combined with a generalizable implicit neural representation trained across thousands of scans to model the continuous intensity field between oblique slices. Our method achieves a 9% relative SSIM improvement over a coordinate-matched trilinear baseline while preserving ultrasound-specific speckle texture and boundary sharpness, and reduces reconstruction time by over 60% compared to per-scene optimization methods.

    Summary →