This is a repository that includes all the codes used to build, train and test our framework. We participated in Accurate Automated Spinal Curvature Estimation MICCAI 2019 Challenge. Our method achieved SMAPE score of 25.69 on test dataset. Details about the challenge is available at AASCE 2019 Grand Challenge. We presented our paper at MICCAI 2019 Workshop -- SpinEva2019. Workshop
Here, we propose a novel framework to estimate vertebra landmarks. This framework has two separately trained networks.
- Object Detector to Predict each vertebra as a single object.
- Landmark Estimator to find landmark location in each vertebra.
The overall network estimates 68 vertebrae landmarks (4 corners for each vertebra) to form a spinal curve. The vertebra detection and estimation is improved by applying some post processing that include outliers rejection and curve smoothing.Finally, we also compute 3 Cobb Angles (MT, PT, TL/L) using the slope of the vertebra landmarks.
If you want to use this code, refer to this folder structure:
- Object_Detection : Faster-RCNN implementation. Please visit Luminoth for more details on training and testing. We used their implementation.
- Landmark_Detection : Use codes here to train local landmark detector. It implements DenseNet in keras.
- CobbAngle_Calculation : It consisits of MATLAB code to smoothen estimated landmarks, and to calculate Cobb Angle from them.
TestAll.py: Use this for inference. It combines Object Detection, Landmark Detection and outlier rejection in a pipeline to produce overall landmarks.- Jupyter_Notebooks: Utility scripts written in notebook
We are working on further improvement of this framework.