NA-MIC Project Weeks

Back to Projects List

Automated Segmentation of Pelvic Bones from Diverse CT Datasets

Key Investigators

• Hamid Alavi (University of Twente, The Netherlands)
• Gabrielle Tuijthof (University of Twente, The Netherlands)
• Nico Venrdonschot (University of Twente, The Netherlands)
• Malte Asseln (University of Twente, The Netherlands)

Project Description

This project aims to segment the hip bones, sacrum, and femur from a collection of public CT datasets that vary in the anatomical regions they cover. To handle this heterogeneity, we aim to develop a four-stage workflow implemented in MONAILabel and 3DSlicer.

Stage 1: Preliminary Bone Segmentation – All CT volumes are processed with TotalSegmentator to generate initial segmentations of the target bones. This provides voxel-level information needed to identify relevant scans.

Stage 2: Selection of Relevant CT Volumes – Using the preliminary segmentations, bone volumes are calculated. Scans with near-zero volumes are excluded, scans within reference ranges are retained, and borderline cases are forwarded to human annotators for review.

Stage 3: Detailed Segmentation and Model Selection – Selected scans undergo precise segmentation using multiple state-of-the-art pretrained models. Aleatoric uncertainty is computed on a subset to select the most consistent model, which is then applied to the remaining scans.

Stage 4: Quality Control and Manual Refinement – Segments with high uncertainty are flagged for review. Annotators refine them interactively using MONAILabel tools like DeepEdit and DeepGrow.

Objective

1. A MonaiLabel application equipped with multiple segmentation models.

Approach and Plan

1. Describe specific steps of what you plan to do to achieve the above described objectives.

Progress and Next Steps

1. Describe specific steps you have actually done.

Illustrations

No response

Background and References

No response