Automatic Cardiac Disease Assessment on cine-MRI via Time-Series Segmentation and Domain Specific Features
Introduction
Winner of the 2017 MICCAI ACDC (Automatic Cardiac Diagnostic Challenge). The method uses an ensemble of 2 CNNs for segmentation and an ensemble of MLPs and a random forest for diagnostic.
Segmentation
As shown in Fig.2, they use an ensemble of a 2D and a 3D unet. These are usual unets but with a feature map combination on the deconvolution side.
Classification
Given the segmentation map obtained by their method, they extract the following cardiac features:
These features are then used to train 50 MLPs and a random forest of 1000 trees. The output of these models are then combined for prediction.
Results
The got very good results with only 4 erroneous diagnosis
and the best overall segmentation accuracy
Code
Segmentation code is available here