3D Multimodal k-means and Morphological Operations (3DMKM) Segmentation of Brain Tumors from MR Images

Abstract

Tumor segmentation algorithms can aid in prognosis and treatment, and are a better alternative to manual segmentation. This study combined thresholding, morphological operations and k-means segmentation to create a new algorithm called 3D multimodal k-means and morphological operations algorithm (3D-MKM) for segmenting tumors. This algorithm used the fast spoiled gradient (FSPGR), T2 weighted fast spin echo (T2-FSE), T2 weighted fluid-attenuated inversion recovery (T2-FLAIR) and contrast enhanced FSPGR (C-FSPGR) as input images. It adjusted the histograms of each sequence to highlight the tumor regions, then performed a thresholding on the T2FLAIR scan to obtain the region of interest (ROI) mask containing the tumor, edema and surrounding tissue. A multichannel view of the ROI was then made by combining the images from different sequences. The multichannel ROI was then segmented by the k-means algorithm into clusters. Next, the clusters were assembled into the enhancing tumor, non-enhancing tumor and edema masks, and further refined using morphological operations. The 3D-MKM algorithm was tested on 9 datasets. It demonstrated promising results in segmenting the entire lesion, with a Sørensen-Dice similarity coefficient of $0.88 \pm 0.05$ and a Hausdorff distance of $12.08 \pm 7.07$ mm from ground truth.