Henri A Vrooman PhD, Presenter: Nothing to Disclose

Chris A. Cocosco PhD, Abstract Co-Author: Nothing to Disclose

Rik Stokking, Abstract Co-Author: Nothing to Disclose

Monique M. Breteler PhD, Abstract Co-Author: Nothing to Disclose

Wiro Niessen PhD, Abstract Co-Author: Nothing to Disclose

k-Nearest-Neighbor (kNN) classification based on prior training data has been successfully applied to classify brain tissue. Unfortunately, training on manually labeled subjects is required and needs to be redone if protocols change. Both conventional kNN classification and a new method, in which training is automated using an atlas, were evaluated for segmenting gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF).

From 12 subjects, low-resolution T2 and PD scans and a high-resolution HASTE scan were used as feature sets. Scans were registered to the high-resolution data. Intensity non-uniformity correction and rescaling were performed to achieve similar feature significance. For the conventional kNN method, manual segmentations were used for training and classifications were done by a leave-one-out strategy. The performance as a function of the number of samples per tissue (22000-220000) and k (1-100) was studied. For fully automated training, scans were registered to a probabilistic brain atlas. Randomly, 7500 initial training samples were selected per tissue based on a threshold (T=0.6) on the probability maps and processed to keep the most reliable samples. Classification results were validated by measuring the percentage overlap (SI).

SI between observers were CSF:91.2%, GM:90.3%, and WM:90.6%. For conventional kNN classification, varying the number of training samples did not result in significant differences (p<0.05). Increasing k gave significantly better results. With pruning, there is an overestimation of GM at the expense of CSF at higher thresholds. The difference between the conventional method (k=45) and the observers was not significantly larger than interobserver variability for all tissue types. The automated method performed equal to the observers for WM (89.4%) and less for CSF (85.6%) and GM (87.9%).

Conventional kNN classification achieves highly accurate results and may replace manual segmentation. Atlas-based kNN classification is close to interobserver variability for GM and CSF and performs not statistically different for WM, showing its potential for fully automated segmentation, without the need of laborious manual training.

C.A.C.: Chris Cocosco works for Philips Research Hamburg, Division Technical Systems, Roentgenstrasse 24-26, 22335 Hamburg, Germany

k-Nearest-Neighbor (kNN) classification based on prior training data has been successfully applied to classify brain tissue. Unfortunately, training on manually labeled subjects is required and needs to be redone if protocols change. Both conventional kNN classification and a new method, in which training is automated using an atlas, were evaluated for segmenting gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF).

From 12 subjects, low-resolution T2 and PD scans and a high-resolution HASTE scan were used as feature sets. Scans were registered to the high-resolution data. Intensity non-uniformity correction and rescaling were performed to achieve similar feature significance. For the conventional kNN method, manual segmentations were used for training and classifications were done by a leave-one-out strategy. The performance as a function of the number of samples per tissue (22000-220000) and k (1-100) was studied. For fully automated training, scans were registered to a probabilistic brain atlas. Randomly, 7500 initial training samples were selected per tissue based on a threshold (T=0.6) on the probability maps and processed to keep the most reliable samples. Classification results were validated by measuring the percentage overlap (SI).

SI between observers were CSF:91.2%, GM:90.3%, and WM:90.6%. For conventional kNN classification, varying the number of training samples did not result in significant differences (p<0.05). Increasing k gave significantly better results. With pruning, there is an overestimation of GM at the expense of CSF at higher thresholds. The difference between the conventional method (k=45) and the observers was not significantly larger than interobserver variability for all tissue types. The automated method performed equal to the observers for WM (89.4%) and less for CSF (85.6%) and GM (87.9%).

Conventional kNN classification achieves highly accurate results and may replace manual segmentation. Atlas-based kNN classification is close to interobserver variability for GM and CSF and performs not statistically different for WM, showing its potential for fully automated segmentation, without the need of laborious manual training.

C.A.C.: Chris Cocosco works for Philips Research Hamburg, Division Technical Systems, Roentgenstrasse 24-26, 22335 Hamburg, Germany

Vrooman, H, Cocosco, C, Stokking, R, Breteler, M, Niessen, W, Multi-Spectral MRI Brain Tissue Classification Using the k-Nearest-Neighbor Algorithm: Training on Manually Labeled Subjects versus Automatic Atlas-based Training on a Single Subject.  Radiological Society of North America 2005 Scientific Assembly and Annual Meeting, November 27 - December 2, 2005 ,Chicago IL. http://archive.rsna.org/2005/4416515.html