Background: Regularized regression methods such as principal component or partial least squares regression perform well in learning tasks on high dimensional spectral data, but cannot explicitly eliminate irrelevant features. The random forest classifier with its associated Gini feature importance, on the other hand, allows for an explicit feature elimination, but may not be optimally adapted to spectral data due to the topology of its constituent classification trees which are based on orthogonal splits in feature space.

Knowledge of left atrial (LA) anatomy is important for atrial fibrillation ablation guidance, fibrosis quantification and biophysical modelling. Segmentation of the LA from Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) images is a complex problem. This manuscript presents a benchmark to evaluate algorithms that address LA segmentation. The datasets, ground truth and evaluation code have been made publicly available through the http://www.cardiacatlas.org website. This manuscript also reports the results of the Left Atrial Segmentation Challenge (LASC) carried out at the STACOM'13 workshop, in conjunction with MICCAI'13. Thirty CT and 30 MRI datasets were provided to participants for segmentation. Each participant segmented the LA including a short part of the LA appendage trunk and proximal sections of the pulmonary veins (PVs). We present results for nine algorithms for CT and eight algorithms for MRI. Results showed that methodologies combining statistical models with region growing approaches were the most appropriate to handle the proposed task. The ground truth and automatic segmentations were standardised to reduce the influence of inconsistently defined regions (e.g., mitral plane, PVs end points, LA appendage). This standardisation framework, which is a contribution of this work, can be used to label and further analyse anatomical regions of the LA. By performing the standardisation directly on the left atrial surface, we can process multiple input data, including meshes exported from different electroanatomical mapping systems.

Abstract. In his original paper on random forests, Breiman proposed two different decision tree ensembles: one generated from "orthogonal" trees with thresholds on individual features in every split, and one from "oblique" trees separating the feature space by randomly oriented hyperplanes. In spite of a rising interest in the random forest framework, however, ensembles built from orthogonal trees (RF) have gained most, if not all, attention so far.In the present work we propose to employ "oblique" random forests (oRF) built from multivariate trees which explicitly learn optimal split directions at internal nodes using linear discriminative models, rather than using random coefficients as the original oRF. This oRF outperforms RF, as well as other classifiers, on nearly all data sets but those with discrete factorial features. Learned node models perform distinctively better than random splits. An oRF feature importance score shows to be preferable over standard RF feature importance scores such as Gini or permutation importance. The topology of the oRF decision space appears to be smoother and better adapted to the data, resulting in improved generalization performance. Overall, the oRF propose here may be preferred over standard RF on most learning tasks involving numerical and spectral data.

Dynamic contrast-enhanced magnetic resonance (DCE-MR) imaging can be used to study microvascular structure in vivo by monitoring the abundance of an injected diffusible contrast agent over time. The resulting spatially resolved intensity-time curves are usually interpreted in terms of kinetic parameters obtained by fitting a pharmacokinetic model to the observed data. Least squares estimates of the highly nonlinear model parameters, however, can exhibit high variance and can be severely biased. As a remedy, we bring to bear spatial prior knowledge by means of a generalized Gaussian Markov random field (GGMRF). By using information from neighboring voxels and computing the maximum a posteriori solution for entire parameter maps at once, both bias and variance of the parameter estimates can be reduced thus leading to smaller root mean square error (RMSE). Since the number of variables gets very big for common image resolutions, sparse solvers have to be employed. To this end, we propose a generalized iterated conditional modes (ICM) algorithm operating on blocks instead of sites which is shown to converge considerably faster than the conventional ICM algorithm. Results on simulated DCE-MR images show a clear reduction of RMSE and variance as well as, in some cases, reduced estimation bias. The mean residual bias (MRB) is reduced on the simulated data as well as for all 37 patients of a prostate DCE-MRI dataset. Using the proposed algorithm, average computation times only increase by a factor of 1.18 (871 ms per voxel) for a Gaussian prior and 1.51 (1.12 s per voxel) for an edge-preserving prior compared to the single voxel approach (740 ms per voxel).

A publicly available standardized framework for the evaluation of (semi)automatic methods for CAC identification in cardiac CT is described. An evaluation of five (semi)automatic methods within this framework shows that automatic per patient CVD risk categorization is feasible. CAC lesions at ambiguous locations such as the coronary ostia remain challenging, but their detection had limited impact on CVD risk determination.

We describe the optimal high-level postprocessing of single-voxel (1)H magnetic resonance spectra and assess the benefits and limitations of automated methods as diagnostic aids in the detection of recurrent brain tumor. In a previous clinical study, 90 long-echo-time single-voxel spectra were obtained from 52 patients and classified during follow-up (30/28/32 normal/non-progressive tumor/tumor). Based on these data, a large number of evaluation strategies, including both standard resonance line quantification and algorithms from pattern recognition and machine learning, were compared in a quantitative evaluation. Results from linear and non-linear feature extraction, including ICA, PCA and wavelet transformations, and also the data from resonance line quantification were combined systematically with different classifiers such as LDA, chemometric methods (PLS, PCR), support vector machines and ensemble methods. Classification accuracy was assessed using a leave-one-out cross-validation scheme and the area under the curve (AUC) of the receiver operator characteristic (ROC). A regularized linear regression on spectra with binned channels reached 91% classification accuracy compared with 83% from quantification. Interpreting the loadings of these regressions, we find that lipid and lactate signals are too unreliable to be used in a simple machine rule. Choline and NAA are the main source of relevant information. Overall, we find that fully automated pattern recognition algorithms perform as well as, or slightly better than, a manually controlled and optimized resonance line quantification.

Despite its diagnostic value and technological availability, 1 H NMR spectroscopic imaging (MRSI) has not found its way into clinical routine yet. Prerequisite for the clinical application is an automated and reliable method for the diagnostic evaluation of MRS images. In the present paper, different approaches to the estimation of tumor probability from MRSI in the prostate are assessed. Two approaches to feature extraction are compared: quantification (VARPRO, AMARES, QUEST) and subspace methods on spectral patterns (principal components, independent components, nonnegative matrix factorization, partial least squares). Linear as well as nonlinear classifiers (support vector machines, Gaussian processes, random forests) are applied and discussed. Quantification-based approaches are much more sensitive to the choice and parameterization of the quantification algorithm than to the choice of the classifier. Furthermore, linear methods based on magnitude spectra easily achieve equal performance and also allow for biochemical interpretation in combination with subspace methods. Nonlinear methods operating directly on magnitude spectra achieve the best results but are less transparent than the linear methods. Clinical studies have shown significant diagnostic value of 1 H magnetic resonance spectroscopic imaging (MRSI) for the detection of tumorous tissue in the prostate (1-6). Despite the promising results of these and other studies, the integration of MRSI in the clinical routine remains difficult, mainly because of the complexity and effort associated with the evaluation of the acquired data.Two basic approaches to the evaluation of MRSI can be distinguished: the quantification-based approach and the pattern recognition (PR) approach. Quantification aims at estimating relative metabolite concentrations as accurately as possible. For that purpose, the most likely parameter estimate for a given signal model is usually determined with a nonlinear least squares (NLS) approach. However, quantification may fail for various reasons. In particular, in the presence of artifacts and severe noise the NLS objective can have many local optima and the result becomes very sensitive to the choice of initial values. Prior knowledge about the expected signal shape can help to alleviate these problems (7), but it also leads to an estimation bias and can be harmful in unanticipated cases where the employed prior knowledge is inadequate. A subsequent statistical analysis that gains diagnostic information from the spectral data relies on these parameter estimates and therefore inherits all problems associated with the quantification.Pattern recognition approaches do not require an explicit quantification step. Although the same methods and classifiers (e.g., logistic regression (8), artificial neural networks (8), support vector machines (SVM) (9), etc.) can be used for both quantified signals and spectral patterns, only methods applied to the latter will be referred to as pattern recognition approaches in accordance with, for example,...

Besides the diagnostic evaluation of a spectrum, the assessment of its quality and a check for plausibility of its information remains a highly interactive and thus time-consuming process in MR spectroscopic imaging (MRSI) data analysis. In the automation of this quality control, a score is proposed that is obtained by training a machine learning classifier on a representative set of spectra that have previously been classified by experts into evaluable data and nonevaluable data. In the first quantitative evaluation of different quality measures on a test set of 45,312 long echo time spectra in the diagnosis of brain tumor, the proposed pattern recognition (using the random forest classifier) separated high-and low-quality spectra comparable to the human operator (area-under-the-curve of the receiver-operator-characteristic, AUC >0.993), and performed better than decision rules based on the signal-to-noise-ratio (AUC <0.934) or the estimated Cramé r-Rao-bound on the errors of a spectral fitting (AUC <0.952). This probabilistic assessment of the data quality provides comprehensible confidence images and allows filtering the input of any subsequent data processing, i.e., quantitation or pattern recognition, in an automated fashion. It thus can increase robustness and reliability of the final diagnostic evaluation and allows for the automation of a tedious part of MRSI data analysis.