Edit this page
Back to Projects List
Interactive Deep Learning for Congenital Heart Disease Segmentation
- Danielle Pace (MIT)
- Adrian Dalca (MIT)
- Polina Golland (MIT)
- Mehdi Hedjazi Moghari (Boston Children’s Hospital)
- Aim: segment all cardiac chambers and great vessels from cardiac MRI, for children with congenital heart disease.
- 20 training cases + large anatomical variability - remains a challenge for automatic segmentation.
- Approach: Integrate some interaction from the user, e.g. scribbles or landmarks.
Approach and Plan
- Already have framework for interactive segmentation. Currently testing using scribbles for aorta segmentation.
- Investigate data augmentation to prevent overfitting - noise / slight intensity changes / small deformations.
- Parameter tuning.
Progress and Next Steps
- Implemented on-the-fly data augmentation, including (1) random affine transformations constrained by a user-specified maximum rotation, translation, scale and shear, and (2) random elastic deformation.
- Currently running trials to measure impact and tune parameters.
Background and References
- HVSMR Challenge Data: (http://segchd.csail.mit.edu)