Fairness in Cervical Spine MRI Segmentation

🚧 Work in Progress — This project is in early development

Investigating algorithmic bias in automated medical image segmentation using the Duke CSpineSeg dataset.

Overview

This project examines fairness in cervical spine MRI segmentation, focusing on how machine-generated labels ("Silver Standard") may introduce or amplify bias compared to expert annotations ("Gold Standard") across age and demographic groups.

Key Objectives

• Fairness Audit: Quantify performance disparities across patient subgroups
• Label Bias Analysis: Compare expert vs. automated annotation quality
• Label Classification: Develop methods to distinguish annotation sources

Methods

• Dataset: Duke CSpineSeg (vertebral bodies & intervertebral discs)
• Architecture: 3D nnU-Net baseline with potential transformer variants
• Infrastructure: MLOps pipeline for reproducible experiments

References

Based on recent work examining bias in medical imaging datasets (MAMA-MIA) and the "Biased Ruler" phenomenon in automated labeling systems.

---

Student: Linus Juni

Supervisors: Aditya Parikh, Aasa Feragen

Institution: DTU Compute