Random forest classification for hippocampal segmentation in 3D MR images

Main goal of this paper is a detailed analysis of the performances of Random Forest algorithm in the field of automated hippocampalsegmentation using 3D MR Images. Fifty-six T1-weighted whole brain MR images were included in the study, together with the related manually segmented bilateral hippocampi (mask). Firstly, the relationship between manual and automated segmentations of hippocampus was explored using a number of standard metrics. For left (right) hemisphere the Dice's coefficient obtained by RF was 70.6% (68.4%). The structural complexity of 3D MR images is twofold. The amount of voxels per image is huge and the numbers of hippocampus and background voxels are strongly imbalanced. In order to overcome these two limitations, we propose two simple strategies: one consists of filtering the input data using the logical OR of the masks of training images, followed by the RF classification task, the other is constituted by learning the RF classifier plane by plane. Using both strategies, the segmentation performances of RF improve significantly and Dice's coefficients increases up to 79.1% (77.4%) for left (right) sides.