PMC

Preprocessing flow chart: Input image is the original image. Eventually, the output image will be fed into the registration/decomposition framework.

Illustration of image appearance variability on a selection of images from each (evaluation) database. From top to bottom: IBSR, LPBA40, BRATS and TBI.

Example BRATS image with its decomposition result in atlas space. (a) Input image after pre-processing; (b) quasi-normal image L + M; (c) non-brain image S; (d) pathology image T.

Example coronal slice of (a) an IBSR MR brain image, (b) the corresponding original IBSR brain segmentation (i.e., union of white matter, gray matter and CSF) and (c) the refined brain segmentation result.

Box plot results for the TBI dataset. Our PCA model shows the best evaluation scores. BET, BEaST, MASS and ROBEX also perform reasonably well. BSE and CNN exhibit inferior performance on this dataset.

Box plot results for the BRATS tumor dataset. BSE and CNN fail on this dataset. BEaST also fails when applied directly to the BRATS dataset due to spatial normalization failures. We therefore show results for BEaST* here, our modification which uses the affine registration of the PCA model first. BET shows better performance, but also exhibits outliers. ROBEX, BEaST*, MASS, and our PCA model work well on this dataset. Overall our model exhibits the best performance scores.

Box plot results for the LPBA40 normal dataset. All seven methods work well on this dataset. Our PCA model has the best Dice and surface distances. ROBEX, BEaST, MASS, BET and BSE show similar performance, but BET exhibits larger variance and BSE exhibits two outliers indicating failure. The CNN model shows overall slightly worse performance than the other methods.

Examples of 3D volumes of average errors for the normal IBSR and LPBA40 datasets, as well as for the pathological BRATS and TBI datasets. For IBSR/BRATS, we show results for BEaST*. Images and their brain masks are first affinely aligned to the atlas. At each location we then calculate the proportion of segmentation errors among all the segmented cases of a dataset (both over- and under-segmentation errors). Lower values are better (a value of 0 indicates perfect results over all images) and higher values indicate poorer performance (a value of 1 indicates failure on all cases). Clearly, BSE and CNN struggle with the BRATS dataset whereas our PCA method shows good performance across all datasets.

Box plot results for the IBSR normal dataset. We show the results from seven methods: PCA, RBX (ROBEX), BST* (BEaST*), MAS (MASS), BET, BSE and CNN. Due to the poor results of MASS and CNN, and the outliers of BSE on this dataset, we limit the range of the plots for better visibility. On each box, the center line denotes the median, and the top and the bottom edge denote the 75th and 25th percentile, respectively. The whiskers extend to the most extreme points that are not considered outliers. The outliers are marked with ‘+’ signs. In addition, we mark the mean with green ‘*’ signs. ROBEX, BET, and BSE show similar performance, but BSE exhibits two outliers. MASS works well on most images, but fails on many cases. BEaST fails on the original images. We therefore show the BEaST* results using the initial affine registration of our PCA model. BEaST* performs well with high Dice scores and low surface distances, but with low mean values. CNN performs poorly on this dataset. Our PCA model has similar performance to BEaST* but with higher mean values. Both methods perform better than other methods on the Dice scores and surface distances.