Research Projects

Methodology

• Medical image segmentation of vertebral regions
• Vertebral ROI extraction and preprocessing
• Deep learning classification with ordinal regression
• Continuous-stage prediction from discrete labels

Key Contributions

• Proposed a continuous-perspective formulation for maturation staging
• Developed an end-to-end pipeline from ROI extraction to stage prediction
• Demonstrated improved ordering consistency over standard classification

Technologies

Methodology

• Noise robustness evaluation across visual perturbations
• Mechanistic interpretability of transformer representations
• Steering vector analysis for controllable behavior
• Table understanding and structured-data reasoning

Key Contributions

• Identified systematic failure modes under visual noise
• Analyzed internal representations via steering vectors

Technologies

Methodology

• Multimodal model fine-tuning on medical image-text pairs
• Report generation with clinical accuracy constraints
• Evaluation against radiologist-authored references

Technologies