Data analyst who builds the whole pipeline.

My contribution

• Co-developed the 3D disease prototype (prototype.py): anatomical weighting, deformation logic and tissue colouring.
• Led the Blender evaluation. I tested materials, lighting and export, identified Blender's limitations, and drove the team's decision to move to a programmatic PyVista + Trimesh pipeline.
• Integrated the audit logger into the pipeline, so every case records its extracted CT values, progression score, disease stage and model outputs.
• Fixed mesh instability in severe cases by recalibrating the deformation constants, so the 3D mesh stopped tearing at high progression values.
• Researched explainable AI and medical visualisation for the risks and significance analysis, and co-wrote the final report and presentation.

Team stack

• Python
• OpenCV
• PyVista
• Trimesh
• Blender
• JSON audit logs

Key decisions

• Blender → PyVista (my evaluation): Blender looked more realistic, but material export problems, environment dependency and manual steps made it unusable for an automated, reproducible pipeline.
• Language model → imaging signals: we first planned to drive the model from radiology reports, but the language model's outputs weren't reproducible, and clinical use needs every result to be traceable.

I replicated DP-MLM (Findings of ACL 2024), a method that uses a masked language model (RoBERTa) to rewrite text with a formal differential-privacy guarantee. I then applied it to new data: AG News subsets and a synthetic dataset containing personal information (PII).

• Python
• PyTorch
• Hugging Face Transformers
• scikit-learn
• NLTK

Result

At ε = 8, 67% of tokens were rewritten while TF-IDF similarity stayed at 0.86. Downstream classifier accuracy fell from 0.42 to 0.29 on a small test split, which measures the cost of the privacy.