Semi-automated Lesions Segmentation of Brain Metastases in MRI Images

A semi-automated method based on a U-Net 3+ network, for the segmentation of brain metastases (BM) lesions is proposed and evaluated on Magnetic Resonance (MRI) images from105 patients with brain metastases. We divided the dataset based on the lesions size as small (S, [2.65, 13.26) mm(2)), medium (M, [13.26, 37.11) mm(2)) and large (L, [37.11, 1152.21) mm(2)) BM. The proposed segmentation method was trained and tested separately on each group and to all aforementioned combinations (7 developed models in total). For each group, 875 image patches with at least one lesion each, were extracted from MRI images, with 700 patches used for 5-fold cross validation and 175 patches for testing on a kept-out set using the averaging ensemble of the five trained models. The segmentation results yielded a Dice Similarity Coefficient (DSC) per patch with median (interquartile range(IQR)) as follows: 0.67(0.25), 0.81(0.13), 0.89(0.08), 0.75(0.22), 0.85(0.28), 0.85(0.13), 0.81(0.24) for S, M, L, S&M, S&L, M&L, and S&M&L size groups respectively. The proposed system will form the basis for a computer-assisted decision and disease follow-up support tool to be used by medical experts.