Project Week 25/NeedleSegmentation

Back to Projects List

Key Investigators

• Paolo Zaffino (Magna Graecia University, Italy)
• Salvatore Scaramuzzino (Magna Graecia University/ASL Vercelli, Italy)
• Maria Francesca Spadea (Magna Graecia University, Italy)
• Guillaume Pernelle (remote) (Imperial College, London, UK)
• Alireza Mehrtash (remote) (Brigham and Women's Hospital, Harvard Medical School, USA)
• Tina Kapur (Brigham and Women's Hospital, Harvard Medical School, USA)

Project Description

NeedleFinder is a tool for segmentation of needles from MR scans which requires manual initialization of the tip of the needle. It has been tested extensively on MR-guided gynecologic brachytherapy data, and preliminarily on MR-guided prostate biopsy data. In this project, we aim to eliminate this reliance on manual interaction and develop a completely automatic strategy to segment the needles. We have tested a CNN approach that provides good results, even if a post processing step must be implemented in order to remove some noise and to refine the obtained segmentations.

Objective	Approach and Plan	Progress and Next Steps
Refine/clear the segmentations coming from the CNN algorithm. Figure out how to transfer MRIs to a server hosting the CNN code and get back the results.	Clustering and morphological filters for data cleaning. Talk with someone from the core team to figure out how to remotely process the data.	A bridge between data coming by the unet and the already existing NeedleFinder module was implemented. Now we are able to use the segmentations generated by the unet to initialize the NeedleFinder algorithm. The process is completely automatic. We had an interesting discussion with Steve about DeepInfer.

Illustrations

Red volume: ground truth, yellow circles: needle coming from NeedleFinder, labels: segmentation coming from unet

Background and References

1. NeedleFinder website
2. Model-based Catheter Segmentation in MRI-images