Presenter Information

Yukang ShenFollow

Location

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

Streaming Media

Document Type

Event

Start Date

24-11-2025 4:00 PM

Description

Vision-Language-Action (VLA) systems are beginning to support everyday clinical workflows. Deltoid intramuscular injection is a representative task, but progress is limited by data scarcity, privacy constraints, and the cost of expert annotation. Recent text-to-image (T2I) models make large-scale data synthesis possible, yet ensuring anatomical correctness, diversity, and label quality remains difficult. To address this gap, we propose a Synthetic Data Engine tailored for medical perception, integrating cold-start filtering, controlled T2I generation, CLIP-based quality checks, and iterative segmentation training. We further introduce an anthropometry-grounded formulation of injection safety that produces interpretable safe-zone guidance. Experiments show that synthetic data can effectively bootstrap deltoid-segmentation performance and support reliable injection-area perception.

Share

COinS
 
Nov 24th, 4:00 PM

GRM-1245 A Synthetic Data Engine for Explainable Injection-Area Perception

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

Vision-Language-Action (VLA) systems are beginning to support everyday clinical workflows. Deltoid intramuscular injection is a representative task, but progress is limited by data scarcity, privacy constraints, and the cost of expert annotation. Recent text-to-image (T2I) models make large-scale data synthesis possible, yet ensuring anatomical correctness, diversity, and label quality remains difficult. To address this gap, we propose a Synthetic Data Engine tailored for medical perception, integrating cold-start filtering, controlled T2I generation, CLIP-based quality checks, and iterative segmentation training. We further introduce an anthropometry-grounded formulation of injection safety that produces interpretable safe-zone guidance. Experiments show that synthetic data can effectively bootstrap deltoid-segmentation performance and support reliable injection-area perception.