Pancreatic Cyst Classification

Project overview

During my undergraduate research, under the mentorship of Dr. Mingren Shen and Dr. Dane Morgan, I led a three-student team developing models to classify pancreatic cysts as potentially cancerous or noncancerous. The clinical goal was to identify harmful cysts with high sensitivity while remaining specific enough to avoid unnecessary surgery.

Method and challenges

We worked with CT radiomics data from the University of Wisconsin School of Medicine. Because the cohort consisted of patients who had undergone surgery, it was naturally skewed toward severe cases. We explored resampling methods, chose metrics appropriate for class imbalance, and examined whether influential features aligned with medically relevant characteristics.

Outcome

I trained an XGBoost model and coordinated the team’s work and schedule. The research resulted in a second-author publication in Abdominal Radiology.