Our approach is three fold:
(1) CellProfiler for segmentation and feature extraction + Random Forest for classification
(2) Graph partitioning for segmentation + SPADE for feature extraction + Random Forest for classification
(3) Graph partitioning for segmentation + Deep Learning for patch level classification + majority vote
Through holdout validation, each of the above methods exhibited similar accuracy, but made different mistakes. The impression is the prediction accuracies among the three approaches are (3)>(1)>(2).
Two sets of results are submitted to the challenge. One is the ensemble of the three approaches. The other is just approach (3).