Acute Pancreatitis Severity Prediction (CT-based AI)
We developed a deep learning model that predicts acute pancreatitis (AP) severity from admission contrast-enhanced CT (CECT). Trained with self-supervised learning on large unlabeled CTs and fine-tuned on labeled studies, the model outperformed clinical/imaging scores (BISAP, mCTSI) and generalized to an external multicenter registry.
Role & Contributions
- Research collaboration: Worked in Prof. Yiqiu Shen’s lab on CT-based deep learning to predict acute pancreatitis (AP) severity at admission; contributed across data engineering, model training, and integration into the lab’s workflow.
- Large-scale data pipeline: Built Python pipelines (pydicom, pandas) to process 10,000+ DICOM folders, extracting key metadata (patient ID, series info, imaging parameters, diagnosis descriptors) to enable reliable slice selection and study-level tracking.
- Image preprocessing: Implemented robust preprocessing with scikit-image (resizing, filtering, windowing) and OpenCV (intensity normalization), producing consistent 2D axial inputs for ViT-based encoders.
- Pancreas localization: Adapted the MedSAM transformer segmentation pipeline to axial CT slices; generated pancreas masks and integrated them into training to prioritize relevant anatomy and reduce low-signal slices.
- Slice selection strategy: Operationalized mask-guided selection of the most pancreas-relevant slices (top-k aggregation) to improve saliency fidelity and reduce noise during supervision.
- Self-supervised pretraining: Trained Vision Transformer encoders with DINO-v2 on broader unlabeled CT data; implemented linear probing to identify the best pretraining epoch for task-specific fine-tuning, yielding up to ~5% performance gains.
- Training infrastructure: Wrote custom PyTorch dataloaders for DICOM/NIfTI ingestion, on-the-fly augmentations, and series-level batching; streamlined study-level inference without requiring segmentation at test time.
- Model integration: Integrated the SSL-pretrained ViT encoders with the lab’s attention-based aggregation head for MAP/SAP ordinal prediction (two-step formulation: MAP vs non-MAP; SAP vs non-MAP).
- Experimentation & evaluation: Ran controlled experiments comparing pretraining strategies and pancreas-mask usage; monitored AUROC/AUPRC and calibration on held-out sets to guide model choices.
- Documentation & handoff: Documented pipelines, configuration templates, and expected I/O formats to support reproducibility and enable other researchers in the lab to run end-to-end training/inference.
Methods Snapshot
- Data: 10,130 CT studies (8,335 pts) in development; 100 internal test; 518 external (Hungarian registry).
- Learning: Self-supervised pretraining (DINO/DINO-v2) → ViT encoder → attention-based aggregation; ordinal formulation (MAP↔MSAP↔SAP).
- Baselines: Compared to mCTSI and BISAP; statistics via AUROC/AUPRC, bootstrap CIs, DeLong/permutation tests.
Tech Stack
- Python
- PyTorch
- pydicom
- pandas
- scikit-image
- OpenCV
- ViT
- MedSAM
- DINO-v2