Image fusion for dynamic contrast enhanced magnetic resonance imaging
 Thorsten Twellmann^{1}Email author,
 Axel Saalbach^{1},
 Olaf Gerstung^{1},
 Martin O Leach^{2} and
 Tim W Nattkemper^{1}
DOI: 10.1186/1475925X335
© Twellmann et al; licensee BioMed Central Ltd. 2004
Received: 04 June 2004
Accepted: 19 October 2004
Published: 19 October 2004
Abstract
Background
Multivariate imaging techniques such as dynamic contrastenhanced magnetic resonance imaging (DCEMRI) have been shown to provide valuable information for medical diagnosis. Even though these techniques provide new information, integrating and evaluating the much wider range of information is a challenging task for the human observer. This task may be assisted with the use of image fusion algorithms.
Methods
In this paper, image fusion based on Kernel Principal Component Analysis (KPCA) is proposed for the first time. It is demonstrated that a priori knowledge about the data domain can be easily incorporated into the parametrisation of the KPCA, leading to taskoriented visualisations of the multivariate data. The results of the fusion process are compared with those of the wellknown and established standard linear Principal Component Analysis (PCA) by means of temporal sequences of 3D MRI volumes from six patients who took part in a breast cancer screening study.
Results
The PCA and KPCA algorithms are able to integrate information from a sequence of MRI volumes into informative gray value or colour images. By incorporating a priori knowledge, the fusion process can be automated and optimised in order to visualise suspicious lesions with high contrast to normal tissue.
Conclusion
Our machine learning based image fusion approach maps the full signal space of a temporal DCEMRI sequence to a single meaningful visualisation with good tissue/lesion contrast and thus supports the radiologist during manual image evaluation.
Background
Today, much effort is spent on enhancing the capabilities of the imaging techniques e.g. increasing the spatial and temporal resolution. In contrast to these improvements in image acquisition, much less effort has been spent on effective visualisation methods. Even though several approaches for detection and classification of suspicious lesions in DCEMRI data of the breast have been proposed (e.g. [4–8]), it is still common practice for the huge amount of data to be analysed manually using simple operations such as subtraction images of two volumes.
Obviously, these images can only comprise a small fraction of the information which is commonly spread over all volumes of the sequences. As a consequence, analysing multivariate images in radiology remains a time consuming and challenging task which potentially can be alleviated by the application of image fusion techniques.
Image fusion
Image fusion methods have been an area of research for several decades. According to Genderen & Pohl [9, 10], image fusion 'is the combination of two or more different images to form a new image by using a certain algorithm' e.g. integration of a large number of multivariate images from a remote sensing process into one image. Because Genderen & Pohl already stated PCA as a standard technique for image fusion in remote sensing, we adopt the more general definition of the term image fusion from the remote sensing community. Whereas in the medical imaging community the meaning of the term image fusion is commonly restricted to fusion of multimodal images, the definition of this term used in this article also includes multivariate images such as multispectral or multitemporal images.
Pattern recognition methods such as artificial neural networks (ANN) have gained much attention from the remote sensing community [11–15]. From the point of view of pattern recognition, the problem of image fusion is strongly related to the task of dimension reduction: Ignoring the spatial order of the patterns x, the image data is an unordered set of patterns that forms a data distribution in the data space and image fusion or dimension reduction corresponds to a mapping
Wellknown algorithms such as Principal Component Analysis (PCA) [16] or Self Organising Maps [17] have been successfully applied for various tasks of multispectral or multitemporal image fusion [11–15]. It is important to note that these methods are not bounded with limitations on the dimensionality of . Hence, they are especially suited if is high dimensional.
In this work, we investigate the application of machine learning algorithms to medical image fusion. We compare the results of the standard linear PCA with it's nonlinear extension, the so called Kernel PCA (KPCA) which was proposed by Schölkopf et al. in 1998 [18]. Our empirical observations are presented and discussed by means of DCEMRI data sets from a breast cancer screening study [19]. Image material presented in this paper is also provided online in original size (PNG format) [20].
Methods
In the following, we briefly describe the theoretical background of the linear PCA and nonlinear KPCA algorithms and their application to the task of image fusion. Both methods determine a set of projection directions, referred to as principal directions (PDs), by optimising a certain criterion. The mapping M is defined by a subset of all possible PDs. Projecting each pattern x _{ p }on to one of these PDs associates each spatial position p with a new scalar value (the principal component) of which integrates information from the different components of x _{ p }, respectively. The resulting 3D image can be visualised as a gray value image or using perceptually optimised colour scales [21, 22]. Alternatively, the low dimensional representation of the patterns can be displayed as RGB composite images, if M is defined by a set of three PDs.
Principal component analysis
Principal Component Analysis is one of the most frequently used dimension reduction method. Suppose the data are given by the set Γ = {x _{ i }}, x _{ i }∈ , 0 ≤ i ≤ N, PCA is a transformation in a new coordinate system of uncorrelated and orthogonal principal axes ξ∈ , ξ = 1 which can be derived from the eigenvectors of the covariance matrix
by solving the eigenvalue equation
λ ξ= C ξ (2)
for λ ≥ 0 and ξ∈ \ {0}. The first eigenvector ξ _{1} (the one with the largest eigenvalue λ _{1}) maximises the variance . Therefore, the set of the first n ≤ d eigenvectors or PDs carry more variance than any other n orthogonal projections.
Kernel principal component analysis
In recent years, kernel based methods have been the object of much research effort within the machine learning community. The concept of a subset of kernel methods is based on the combination of wellknown linear algorithms such as Principal Component Analysis or Fisher Discriminant Analysis with nonlinear kernel functions [23, 24]. While the application of these functions allows more powerful nonlinear solutions, the kernelised algorithms retain most properties of their linear versions.
Consider a nonlinear function
which maps the examples x ∈ Γ to some feature space [25]. Furthermore, assume that the mapped data are centred in . In order to perform the PCA in , one has to find the eigenvectors ξ of the covariance matrix
Nλ α= K α (4)
where α denotes the column vector of coefficients α^{(1)},...,α^{(N)}describing the dual form of the eigenvector by
and K is the symmetric Gram matrix with elements
K _{ ij }= K(x _{ i }, x _{ j }) = Φ(x _{ i }), Φ(x _{ j }) . (6)
Normalising α _{ k }corresponding to the kth eigenvalue λ _{ k }of K ensures λ _{ k } α _{ k }, α _{ k } = 1. Now, principal components can be extracted in by projecting an example x on ξ _{ k }using
with bandwidth parameter σ or the Polynomial Kernel of degree d
K(x _{ i }, x _{ j }) = x _{ i }, x _{ j } ^{ d } (9)
 1.
Calculate the Gram matrix K of Γ using a suitable parameterised kernel function.
 2.
Transform K according
 3.
Calculate the eigenvector expansion coefficients α _{ k }, i.e. the eigenvectors of and normalise them.
 4.
Extract principal components using (7).
Compression vs. discrimination
Application of both image fusion techniques leads to a set of up to d PDs in case of PCA and up to N PDs in case of KPCA. In general, a compact visualisation of the complete data as a single image is desired. In this case, inspection of the fused image based on the PD corresponding to the first (largest) eigenvalue is optimal in terms of a general compression scheme: The projection on this PD retains most of the total data variance and leads to a reconstruction with least mean square error. Nevertheless, image fusion is commonly employed with a well defined intention e.g. in order to detect a specific phenomenon such as bushfires in multitemporal satellite images [11] or (as in this work) tumour lesions in DCEMRI data. In addition to the general compression characteristics, the fused image has to show taskspecific discriminative properties which do not necessarily reflect the total data variance. In this case, using a PD corresponding to one of the following eigenvalues may lead to more discriminative visualisations. If the image data are fused by KPCA, an additional degree of freedom can be exploited. In addition to the index of the selected PD, the type and parameterisation of the kernel K can be varied leading to alternative mappings to the feature space, changing the characteristic of the fusion image.
Experiments
In the following, the fusion results of both methods are discussed and illustrated with DCEMRI sequences from six cases (referred to as S _{1},...,S _{6}) which were taken during the the MARIBS breast screening study [19]. Each sequence consists of seven 3D MRI volumes of the female breast, recorded with a separation of 90 sec using a standardised protocol (A fast spoiled gradient echo sequence (FLASH) with TR = 12 ms, TE = 5 ms, ip angle = 35°, FOV = 340 mm and coronal slice orientation). Before recording the third volume, a gadoliniumbased contrast agent was administered with a bolus injection. Therefore, each spatial position p in the 256 × 128 × 64 (1.33 mm × 1.33 mm × 2.5 mm) sized volume is associated with a pattern , d = 7 describing the temporal signal kinetic of the local tissue.
The images were manually evaluated by an expert who marked voxels of tumour with a cursor on an evaluation device. Below, the kinetic signals of the marked tumour voxels are labelled '+'. Signals corresponding to voxels of the complement of the marked region are labelled ''.
For this kind of data, experiments of Lucht et al. [5] suggest recording a much longer temporal sequence of 28 images which makes the need for efficient fusion techniques evident.
Evaluation criteria
 1.
The marked region should be visualised with high contrast compared to unmarked regions in order to facilitate detection of kinetic signals which are similar to the marked signals.
 2.
The fusion image should follow the first criteria without time consuming manual manipulation by the observer (e.g. tuning of transfer functions such as windowing).
Following the first criteria, the purpose of the visualisation is specified implicitly by the voxel labels. In the present work, the expert marked regions of tumour tissues. Thus, optimal fusion images of an image sequence display locations of cancerous kinetic signals with high contrast to normal signals.
Next to the visualisation of the fusion images as gray value and RGB images, both methods are evaluated by means of a receiveroperatingcharacteristic (ROC) analysis [27, 28]. To this end, pixel intensities of the fusion images are interpreted as confidence values for the existence of suspicious signals and are compared with the expert label as ground truth. The ROC analysis objectively measures the applicability of the fusion images for the task of lesion detection.
However, no conclusion can be drawn about how well other tissue types are distinguishable in the fusion images, i.e. how well the information of the entire signal space is represented.
Preprocessing
For numerical reasons, the voxel value range of each volume sequence is individually normalised to [0; 1]. In order to preserve the signal kinetics, the individual minimal and maximal intensity value is determined simultaneously on all d image volumes of each sequence. To ensure this normalisation is robust with respect to single outlier values, the values are calculated based on an application of a 3 × 3 × 3 median filter.
Since about 66% of each volume is covered by background, all images sequences are preprocessed with a full automatic tissue/background separation method. The histogram of the sum of local intensity differences (sod) feature
individually calculated for each sequence, has a bimodal shape and shows a clear separable maximum for the background voxels. The optimal threshold separating background from tissue can be computed automatically [29]. The resulting binary masks are postprocessed with a morphological closing operator [30] to ensure closed masks for the regions of tissue.
Adaptation
In order to automate and optimise the fusion process, a priori knowledge about the phenomenon to be visualised, given by the expert label, is used to find a suitable parameterisation of the algorithms as described in detail in the following section. In practice, these labels are not available for new image sequences. Thus, the algorithms have to be adapted on a small number of image sequences, e.g. from a subgroup of cases of a screening study, which were manually evaluated by a human expert and can be subsequently applied to the data of an arbitrary number of unseen cases.
To assure the experimental setup reflects the circumstances of a practical application, the data sets Γ used for adaptation consist of marked tissue signals from only five of the six image sequences and the sixth unseen image sequence is used for the evaluation of the algorithm's capabilities. This setup is repeated six times, each time using a different image sequence for evaluation. In case of KPCA, using all kinetic signals from the five image sequences is prohibitive due to the computational and memory complexity. Therefore, the KPCA is adapted with a reduced data set Γ consisting of all signals of the marked tumour regions and an equal number of signals randomly selected from nontumour regions.
Parameter selection
with class specific mean μ^{±} and variance v^{±}. The Fisher score is commonly used for ranking components x^{(k)}of a set {(x, y)} of binary labelled (y = ±) examples according to their discriminative power. In a similar manner, the score can be evaluated for different PDs on a random subset of the training set Γ utilising the corresponding principal component values with their associated expert label and thus can be interpreted as a measure for the first evaluation criteria. Furthermore, the sign of the PCA/KPCA based PDs can be adjusted in order to obtain a high value for the average intensity of tumour voxels causing tumour lesions to appear as bright regions.
Thereby, the a priori knowledge of which region of the five image sequences used for adaptation should be visualised with high contrast can be utilised for selecting proper parameterisations which lead to discriminative visualisations tailored to the given task.
Fusion
Results
Fusion results for the sequences S _{1},...,S _{6} based on the PCA algorithm are shown in the lower 2 × 2 block of Fig. 4, Fig. 5, Fig. 6, Fig. 7, Fig. 8 and Fig. 9. For all six sequences, the fusion image I _{1} based on the PD with the leading eigenvalue does not lead to discriminative visualisations. The tumour lesions appear with the same intensity as fatty tissue, while glandula tissue is displayed as dark areas (S _{3}, S _{4}). In contrast to I _{1}, the discriminative power of I _{2} is obviously much greater for all six image sequences. The display of the tumour lesions (high intensity values) differs significantly from areas of glandular tissues, blood vessels (medium intensity values) and fatty tissue (low intensity values). The contrast between tumour lesion and the surrounding tissue decreases in I _{3} of S _{2}, S _{3} and S _{5}. Additionally, the surrounding tissue is displayed less detailed (S _{1}, S _{2}, S _{4}, S _{5}). According to the weak discriminative characteristic of I _{1} and I _{3}, the tumour lesions are coloured with shadings of green or cyan in the corresponding I _{ RGB }.
Fusion images based on KPCA are shown in the upper 2 × 2 block of Fig. 4, Fig. 5, Fig. 6, Fig. 7, Fig. 8 and Fig. 9. For S _{1}, S _{2}, S _{3} and S _{4}, image I _{1} displays the tumour lesion with high contrast to the surrounding tissues. Adipose tissue appears in I _{1} and I _{2} with mediumin tensity. In I _{2} of S _{4} and S _{3}, glandular tissue can be observed in addition to the tumour. These areas appear dark in I _{1}. The fraction of glandular tissue regions in I _{2} of S _{1} and S _{2} is much smaller, since the tumour is located near the chest muscle where the breast mostly consists of fatty tissue and blood vessels.
An interesting detail can be observed in I _{3} of S _{4}. The image clearly shows a ring structure as part of or around the tumour lesion. At positions inside the ring which are displayed with high intensity values in I _{1} and I _{2}, the temporal kinetic patterns show a fast uptake with a following constant or slightly decreasing concentration of the contrast agent. In contrast to the signals inside the ring, all signals corresponding to the ring structure in I _{3} show as teadily increasing concentration. In all composite images I _{ RGB }except for S _{5}, the tumour lesions are coloured white and can be easily discriminated from fatty tissue (shadings of blue to purple) and glandular tissue (shadings of blue to green). For image S _{5}, only I _{1} shows a discriminative characteristic. The tumour is displayed as a small cluster of high intensity values in the lower right area of the right breast, next to the implant.
Area under ROC curve values for fusion images I _{1}, I _{2} and I _{3} for series S _{1},...,S _{6} based on KPCA, PCA and PCA trained with the same reduced training set as KPCA. For each AUC value, the pixel intensities of the fusion images are interpreted as confidence values indicating the existence of suspicious signals at the corresponding positions. The largest AUC value for each case is printed bold.
AreaUnderROCCurve  

KPCA  PCA  PCA (reduced set)  
Sequence  I _{1}  I _{2}  I _{3}  I _{1}  I _{2}  I _{3}  I _{1}  I _{2}  I _{3} 
S _{1}  0.950  0.972  0.879  0.539  0.993  0.633  0.772  0.972  0.692 
S _{2}  0.918  0.945  0.728  0.727  0.993  0.712  0.852  0.948  0.547 
S _{3}  0.995  0.998  0.710  0.520  0.997  0.926  0.799  0.997  0.747 
S _{4}  0.996  0.985  0.259  0.926  0.999  0.919  0.992  0.985  0.963 
S _{5}  0.959  0.966  0.904  0.693  0.997  0.925  0.814  0.964  0.344 
S _{6}  0.994  0.986  0.706  0.926  0.999  0.919  0.785  0.986  0.802 
The influence of the bandwidth σ on the fusion characteristic is illustrated in Fig. 10. For small values of the bandwidth σ only a small fraction of the tumour lesion appears with high intensities. If the bandwidth is chosen according to the maximum of the Fisher score, the lesion is visualised with high contrast to the surrounding tissue. In the shown example, the Fisher criterion decreases along with the contrast of the visualisation for further increasing bandwidth values.
Discussion
The results shown in the preceding section indicate that fusion of DCEMRI data by PCA or KPCA leads to compact and meaningful visualisations. Lesions are correctly displayed as bright regions or with specific colouring and can be easily discriminated from surrounding tissue. Once a small subgroup of cases is evaluated, the obtained secondary information in the form of labelled tumour areas is utilised for automation of the data processing and presentation: (i) The sign of the PD is selected in a way that tumour lesions always appear with high intensities. (ii) The parametrisation of the kernel function of the KPCA is optimised in such a way that the fusion images show the desired discriminative characteristics. Thus, both evaluation criteria stated in the section Evaluation criteria are accomplished.
The ROC analysis of the KPCA based fusion images indicates that the fusion images I _{2} show superior discriminative characteristics for four of six cases (S _{1}, S _{2}, S _{3} and S _{5}). However, selection of a suitable kernel parametrisation leads to comparable AUC values for I _{1}. For fusion images corresponding to PDs with smaller eigenvalues, KPCA based images still show more details than those based on PCA, if the bandwidth value is chosen according to the maximum of the Fisher score. Figure 11 shows the KPCA based (left column) and the PCA based (right column) fusion images I _{4}, I _{5} and I _{6} for sequence S _{4}. While KPCA distributes the total data variance on N PDs, the PCA method uses only d PDs. Therefore, the PCA based fusion images I _{4}, I _{5} and I _{6} typically contain a large fraction of high frequent noise. It is important to note that the fusion images based on KPCA are not necessarily uncorrelated, if each image is calculated using PDs with different bandwidth values, and therefore may display redundant information. In five of six cases, RGB visualisations based on KPCA show the tumour lesion as white regions which are easy to discriminate from other tissue types. In contrast to subtraction images which also allow detection of lesions with high sensitivity (see e.g. [4]), the fusion images I _{ RGB }provide a more comprehensive display of the data. A single subtraction image displays only the information of a two dimensional subspace of the signal space , i.e. the information of two manually selected components of the signal vector. Without further manipulation of the transfer function and after selection of two suitable components, a subtraction image commonly shows the lesion as a cluster of high intensity values and other types of tissue are not displayed or indistinguishable. The fusion images are low dimensional representations of the entire signal space. Thus, the RGB composite images I _{ RGB }based on PCA or KPCA clearly display the lesion in combination with glandular or fatty tissue and major blood vessels.
In consideration of the fact that both methods are able to fuse the multitemporal DCEMRI to single meaningful images which do not only show the lesion with high intensities, but also other types of tissue such as fatty or glandular tissue, the standard linear PCA seems to be most suitable for the given signal domain because of it's low computation time and superior detection performance. Only for PCA, the three fusion images can be calculated for a complete volume in a reasonable time and without delaying the diagnostic process. According to the ROC analysis, the introduction of nonlinearity by the kernel function did not improve the discriminative properties of the fusion images, but visual appraisal of the RGB composite images based on KPCA suggest a more comprehensive display of the different types of tissue. It is an open question whether fusion images of other data domains with more complex or higher dimensional signals might benefit more obviously from the nonlinearity of KPCA.
Conclusion
In this paper, we have demonstrated the integration of distributed information from DCEMRI image sequences to meaningful visualisations by means of PCA and KPCA. Both methods were able to accentuate the regions marked by the expert as important in image sequences blinded to automatic analyses. By the employment of taskspecific information, the parametrisation of the KPCA algorithm was optimised in order to accentuate the relevant characteristics of the visualisation.
List of abbreviations
 PCA :

Principal Component Analysis
 KPCA :

Kernel Principal Component Analysis
 PD :

Principal Direction
 DCEMRI :

Dynamic Contrast Enhanced Magnetic Resonance Imaging
Declarations
Acknowledgements
The authors would like to thank the advisory staff, radiologists, geneticists and surgeons of the MARIBS Breast Screening Study:
Principal Investigator: M.O. Leach. Manchester: C. Boggis, S. Reaney, M. Wilson, J. Hawnaur*, J. Adams, D.G.R. Evans, A. Shenton. Sutton &St George's: P. Kessar*, G. Brown, J. Murfitt, R. Eeles*, R.S. Houlston, A. ArdenJones, K. Bishop, S. Gray, F. Lennard, S. Shanley, N. Rahman, R. Houlston. Newcastle: J. Potterton, A. Coulthard, L. McLean, M. McElroy, W. Wotherspoon, F. Douglas. Cambridge: R. Warren, A.K. Dixon, P. Britton, R. Sinnatamby, A. Freeman, J. Mackay, J. Rankin. Edinburgh: J. Walsh, B.B. Muir, A. Murray, C.M. Steel, E. Anderson, J.M. Dixon. Leeds/Sheffield/Hull: L.W. Turnbull, G. Hall, P. Kneeshaw, A. Hubbard, G. Liney, C. Chu, O. Quarrell, J. Kumar. Guy's &St Thomas, London: S. Rankin, S. McWilliams, J. Pemberton, A. Jones, E. Sanderson, S.V. Hodgson. Aberdeen: F.J. Gilbert, G. Needham, H. Deans, K. Duncan, L. Gomersall, G. Iyengar, N. Haites, H. Gregory. Southampton: C. Rubin, M. Briley, S. Hegarty, G. Michaels, M. Reddy, D.M. Eccles. Dundee: D.G. Sheppard, J. Rehman, A. Cook, C. Walker, D. Goudie. Bristol: N. Slack, I. Lyburn, A. Jones, E. Kutt, J. Basten, M. Shere, S. Cawthorn, Z. Rayter. Birmingham: C.P. Walker, S. Bradley, M. Wallis, T. Cole, C. McKeown, J. Morton Glasgow: L. Wilkinson, J. Litherland, C. Cordiner, R. Davidson Liverpool: G. Whitehouse, D. Ritchie, F. White, I. Ellis Belfast: G. Crothers, McAllister, P. Morrison Northwick Park: W. Teh, B Shah, J. Paterson Portsmouth: P. Gordon, P. Buxton, J. Domsan UCH London: M. HallCraggs Barnet: G. Kaplan. Mount Vernon: A.R. Padhani, K. Raza. Study staff: L. Pointon, R. Hoff, D. Thompson, K. Chan, M. Khazen, C. Hayes, R. Fowler, M. Sydenham, E. Moore, J. Anderson, C. Levesley. Advisory Group members (not mentioned elsewhere): J. Brown (Bristol, MRC HSRC), D. Easton (Cambridge), L. Walker (Hull), S. Moss (Sutton).
Authors’ Affiliations
References
 Orel S, Schnall M: MR Imaging of the Breast for the Detection, Diagnosis, and Staging of Breast Cancer. Radiology 2001, 220: 13–30.View ArticleGoogle Scholar
 Kinkel K, Hylton N: Challenges to Interpretation of Breast MRI. Journal of Magnetic Resonance Imaging 2001, 13: 821–829. 10.1002/jmri.1117View ArticleGoogle Scholar
 Kuhl C K, Mielcareck P, Klaschik S, Leutner C, Wardelmann E, Gieseke J, Schild H H: Dynamic breast MR imaging: Are signal intensity time course data useful for differential diagnosis of enhancing lesions? Radiology 1999, 211: 101–10.View ArticleGoogle Scholar
 Twellmann T, Saalbach A, Müller C, Nattkemper T, Wismüller A: Detection of Suspicious Lesions in Dynamic ContrastEnhanced MRI Data. In Proc of EMBC 2004, 26th Annual Int Conf of the IEEE Engineering in Medicine and Biology Society 2004.Google Scholar
 Lucht R, Knopp M, Brix G: Classification of signaltime curves from dynamic MR mammography by neural networks. Magnetic Resonance Imaging 2001, 19: 51–57. 10.1016/S0730725X(01)002223View ArticleGoogle Scholar
 Jacobs M, Barker P, Bluemke D, Maranto C, Arnold C, Herskovites E, Bhujwalla Z: Benign and Malignant Breast Lesions: Diagnosis with Multiparametric MR Imaging. Radiology 2003, 229: 225–232.View ArticleGoogle Scholar
 Kelcz F, FurmanHaran E, Grobgeld D, Degani H: Clinical Testing of HighSpatialResolution Parametric ContrastEnhanced MR Imaging of the Breast. A J R oentgenol 2002, 179: 1485–92.View ArticleGoogle Scholar
 Tofts P: Modeling Tracer Kinetics in Dynamic GdDTPA MR Imaging. Journal of Magnetic Resonance Imaging 1997, 7: 91–101.View ArticleGoogle Scholar
 Pohl C, van Genderen J: Image fusion: Issues, techniques and applications. In In Proceedings EARSeL Workshop Edited by: van Gendern J, Cappellini V. 1994.Google Scholar
 Pohl C, van Genderen J: Multisensor image fusion in remote sensing: concepts, methods and applications. Int Journal of Remote Sensing 1998,19(5):823–854. 10.1080/014311698215748View ArticleGoogle Scholar
 Richards J, Milne A: Mapping Fire Burns and Veetation Regeneration Using Principal Component Analysis. In Proceedings Int Geoscience and Remote Sensing Symposium 1983.Google Scholar
 Manduca A: Multispectral medical image visualization with selforganizing maps. In In Proceedings ICIP94 (Cat. No. 94CH35708). Volume 1. Los Alamitos, CA, USA: IEEE Comput Soc Press; 1994:633–7.Google Scholar
 Villmann T, Merenyi E, Hammer B: Neural Maps in remote sensing image anlysis. Neural Networks 2003,16(3–4):389–403. 10.1016/S08936080(03)000212View ArticleGoogle Scholar
 Merenyi E: The challenges in spectral image analysis: An introduction and review of ANN approaches. In Proceedings of the European Symposium on Artificial Neural Networks (ESANN) 1999 1999.Google Scholar
 Richards J: Remote Sensing Digital Image Analysis – An introduction. Springer; 1993.View ArticleGoogle Scholar
 Jolliffe I: Principal Component Analysis. Springer; 1986.View ArticleGoogle Scholar
 Ritter H, Martinetz T, Schulten K: Neural Computation and SelfOrganizing Maps. AddisonWesley; 1992.Google Scholar
 Schölkopf B, Smola A, Müller K: Nonlinear Component Analysis as a Kernel Eigenvalue Problem. Neural Computation 1998,10(5):1299–1319. 10.1162/089976698300017467View ArticleGoogle Scholar
 Brown J, Buckley D, Coulthard A, Dixon A, Dixon J, Easton D, Eeles R, Evans D, Gilbert F, Graves M, Hayes C, Jenkins J, Jones A, Keevil S, Leach M, Liney G, Moss S, Padhani A, Parker G, LJ P, Ponder B, Redpath T, Sloane J, Turnbull L, Walker L, Warren R: Magnetic resonance imaging screening in women at genetic risk of breast cancer: imaging and analysis protocol for the UK multicentre study. UK MRI Breast Screening Study Advisory Group. Magn Reson Imaging 2000,18(7):765–76. 10.1016/S0730725X(00)001673View ArticleGoogle Scholar
 Image Fusion for Dynamic Contrast Enhanced Magnet Resoncance Imaging – Auxiliary Material [http://www.techfak.unibielefeld.de/ags/ani/projects/ImageFusionForDCEMRI]
 Ware C: Color Sequences for Univariate Maps: Theory, Experiments and Principles. IEEE Computer Graphics & Applications 1988,8(5):41–49. 10.1109/38.7760View ArticleGoogle Scholar
 Ware C: Information Visualization: Perception for Design. Morgan Kaufmann; 2000.Google Scholar
 Mika S, Schölkopf B, Smola A, Müller K, Scholz M, Rätsch G: Kernel PCA and DeNoising in Feature Spaces. In In Advances in Neural Information Processing Systems 11. Edited by: Kearns M, Solla SA, Cohn D. MIT Press; 1999.Google Scholar
 Mika S, Rätsch G, Weston J, Schölkopf B, Müller K: Fisher Discriminant Analysis with Kernels. In In Neural Networks for Signal Processing IX. Edited by: Hu Y, Larsen J, Wilson E, Douglas S. IEEE; 1999:41–48.Google Scholar
 Schölkopf B, Mika S, Burges C, Knirsch , Müller K, Rätsch G, Smola AJ: Input space vs. feature space in kernelbased methods. IEEE Trans on Neural Networks 1999,10(5):1000–1017. 10.1109/72.788641View ArticleGoogle Scholar
 Schölkopf B, Smola A, Müller K: Advances in Kernel Methods – Support Vector Learning. MIT Press; 1999.Google Scholar
 Fawcett T: ROC Graphs: Notes and Practical Considerations for Researchers. Tech rep HP Labs 2003.Google Scholar
 Hanley JA: Receiver operating characteristic methodology: the state of the art. CRC critical reviews in diagnostic imaging 1989, 29: 307–335.Google Scholar
 Otsu N: A threshold selection method from gray level histograms. IEEE Trans Systems Man and Cybernetics 1979, 9: 62–66.View ArticleGoogle Scholar
 Sonka M, Hlavac V, Boyle R: Image Processing, Analysis, and Machine Vision. Brooks/Cole; 1998.Google Scholar
 LAPACK – Linear Algebra PACKage [http://www.netlib.org/lapack/]
Copyright
This article is published under license to BioMed Central Ltd. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.