Training the MitoSegNet Model and Using It with the MitoS Tool
(A) The original training data is comprised 12 raw images and the appendant hand-generated ground truth images. Each image is split into 4 overlapping tiles of equal length. For each tile, a weight map is generated and subsequently all three set of tiles (raw images, ground truth, and weight maps) are augmented 80 times, increasing the size of the training data to 3,840 image tiles. Prior to training, the augmented training data is split into training (80%) and validation data (20%). The pretrained MitoSegNet model can now be used to segment new images of mitochondria.
(B) We performed a cross validation for which 12 separate MitoSegNet models were trained each with 11 images, excluding one image that was later used to test the prediction accuracy against other segmentation methods.