Self-Adaptive 2D-3D Ensemble of Fully Convolutional Networks for Medical Image Segmentation

The proposed 2D-3D FCN ensemble is constructed in two phases as shown in Fig. 1. In Phase I, the 2D FCN and 3D FCN architectures are adapted to the specific dataset using a Multiobjective Evolutionary based Algorithm (MEA algorithm) presented in our previous work [24]. This is performed by dividing the dataset into 5 folds and selecting a fold at random to define the 2D and 3D FCN architectures. In Phase II, the optimal 2D FCN and 3D FCN architectures are trained with each of the 5 folds from the training dataset and subsequently averaging the softmax probability maps of the 2D and 3D FCNs.