• No results found

Automation of spectroscopic scoring of prostate MRSI by a canonical correlation  analysis based classification

N/A
N/A
Protected

Academic year: 2021

Share "Automation of spectroscopic scoring of prostate MRSI by a canonical correlation  analysis based classification"

Copied!
1
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

Automation of spectroscopic scoring of prostate MRSI by a canonical correlation 

analysis based classification

M. De Vos, P. Pels, T. Laudadio, S. Van Huffel, J. Kurhanewicz, S. J. Nelson ESAT, Katholieke Universiteit Leuven, Heverlee­Leuven, Belgium, Radiology, University of California, San Francisco, San Francisco, CA, United States. Introduction:  A recent study showed high inter­observer agreement by using a standardized system for  scoring spectroscopic imaging data of the prostate[1]. This study explores the feasibility of replacing  the time expensive human interpretation  of the MRSI  exams  by a  statistical tissue segmentation  technique based on Canonical Correlation Analysis (CCA) [2]. Methods: MRSI was performed on a 1.5T whole body MR scanner (Signa, GE) using an endorectal  receiver coil. The MRSI data sets (n=2508 voxels) were analyzed by an experienced MRS reader on a standardized five category scale, and these scores were used as golden standard for statistics. All  datasets were analyzed using CCA, which is a multivariate generalization of cross­correlation in order  to include spatial information. In this study 3 versions of the spatial model of CCA were used: a single  voxel (SV CCA), a 2­dimensional (Axial CCA), and a 3­dimensional model (3D CCA). The spectral  models of CCA were based on scored  in vivo  spectra. A subgroup of 4 patients with a large number  of malignant voxels was selected to verify the influence of spatial distribution of malignant tissue.  Descriptive   statistics are shown in Table 1. Specificity ranged from 87% to 90% for all methods,  indicating a good classification of healthy tissue. Sensitivity ranges from 65% to 71% due to a larger  number of false  negatives i.e. malignant spectra classified as healthy by the CCA classification. The  3D CCA method takes into account more spatial information leading to higher sensitivity (90%) and  specificity (71%). Because in the subgroup a higher number of malignant voxels is present, the Axial  CCA and 3D CCA benefit more from the spatial information. Also the accuracy of the 3D CCA (86%)  is higher due to the increased incorporation of spatial information. The accuracy of the CCA methods is limited by two factors. Firstly, no anatomical MRI information was incorporated in the analysis;  information that is used by the spectroscopist to achieve a higher scoring sensitivity. Secondly, the CCA methods incorporated spectra from central gland voxels, which are excluded when the data are  scored with knowledge of morphologic information provided by T2 MRI. Conclusion:   CCA based methods provide a means of classifying prostate MRSI data with high  accuracy and specificity. Sensitivity could be increased by increasing the spatial resolution of the  spectral data and by adding morphologic information obtained from MRI.  References [1] Jung,et al. Radiology  2004;233 [2] Laudadio,et al. MRM  2005 ;54(6)

Referenties

GERELATEERDE DOCUMENTEN

Using low rank approximation and incremental eigenvalue algorithm, WMKCCA is applicable to machine learning problems as a flexible model for common infor- mation extraction

In the second step, NNMFSC is applied within the region with abnormal tissue with the purpose of identifying for each voxel within this region its predominant

To make this technique more practical in a clinical environment we propose an automatic MRSI data segmentation using a blind source separation technique (BSS) that provides,

These methods can disentangle mixed tissue voxels in MRSI data acquired from brain tumors, and thus extract representative, tissue-specific spectra (called spectral sources), as

Therefore the interaction between the diastogram and tachogram will be dependent on body position; the canonical cross-loading in standing position was higher than those found in

While the FDA method was not able to resolve the overlapping choline, creatine and polyamine resonances and the TDF method had to omit the polyamines from the model function to

Fast nosological imaging using canonical correlation analysis of brain data obtained.. by two-dimensional turbo

To make this technique more practical in a clinical environment we propose an automatic MRSI data segmentation using a blind source separation technique (BSS) that provides,