Brain Tumors Detection and Segmentation in MRI




Publications:


 

Abstract (ISBI 2014):


Automatic detection and segmentation of brain tumors in 3D MR neuroimages can significantly aid early diagnosis, surgical planning, and follow-up assessment. However, due to diverse location and varying size, primary and metastatic tumors present substantial challenges for detection. We present a fully automatic, unsupervised algorithm that can detect single and multiple tumors from 3 to 28,079 mm^3 in volume. Using 20 clinical 3D MR scans containing from 1 to 15 tumors per scan, the proposed approach achieves between 87.84% and 95.30% detection rate and an average end-to-end running time of under 3 minutes. In addition, 5 normal clinical 3D MR scans are evaluated quantitatively to demonstrate that the approach has the potential to discriminate between abnormal and normal brains.

 

Method:


3D Blob Filtering: Brain tumors are usually blob-like structures, and the presence of it often corrupts the approximate bilateral symmetry of the brain. Therefore, we detect brain tumors by performing 3D Laplacian-of-Gaussian filtering over a given brain MRI scan to extract 3D blobs as our initial tumor candidates.


Equation 1. 1D separable form for 3D Laplacian of Gaussian (LoG) filter. f is the input 3d volume, delta g is 1D LoG, and g is a Gaussian in the x, y, or z dimension.


Blob Shape Pruning: 3D blob detector could also pick up structures such as blood vessels, ventricles, skull plates, etc. Since these tend to be more elongated in shape, we measure the local shape of all detected 3D blobs by computing their 3D affine adaptations, and discard blobs with highly elliptical shapes from the tumor candidate pool.


Equation 2. The 3D structure tensor of a 3D blob, composed of the gradient information I along the x, y, and z dimension. The shape information is the eigenvalues, computed by applying eigenvalue decomposition to this matrix.


Bilateral Symmetry-based Pruning: Normal human brains exhibit an approximate bilateral symmetry, and the presence of brain tumors often break this symmetry. Therefore, blobs that have a bilateral match can be discarded as normal blob-like structures of the brain. This is done by compare all blobs to its bilaterally symmetrical location of the brain, with respect to the Midsaggital Plane (MSP), using Earth Mover's Distance as a metric.

Figure 1. Intermediate and final results for a sample case that has 5 tumors. A) representative 2D slices of the original 3D MRI scan, 5 tumors are circled in red. B) initially detected 3D blobs, 27561 total. C) after shape compactness pruning: 3,631 blobs remaining. D) after bilateral symmetry-based pruning: 1,452 blobs remaining. E) after final thresholding, 11 blobs remaining. The smallest tumor in this case has a volume of 12 mm^3.

 

Sample Results and Videos (Click on the images for video):

   videos require Xvid MPEG-4 codec, can be downloaded from http://www.xvidmovies.com/codec/


   2-tumor-brain (case 15) automatic detection and segmentation result:



   8-tumor-brain (case 11) automatic detection and segmentation result:



   5-tumor-brain (case 17, Figure 1) automatic detection and segmentation result: