Impact of age on the efficacy of oxaliplatin in the preoperative chemoradiotherapy and adjuvant chemotherapy of rectal cancer: a post hoc analysis of the CAO/ARO/AIO-04 phase III trial.

Background: The German rectal cancer trial CAO/ARO/AIO-04 has shown a significant benefit in 3-year disease-free survival (DFS) of adding oxaliplatin to a standard preoperative 5-fluorouracil (5-FU)-based chemoradiotherapy (CRT) and adjuvant chemotherapy in patients with locally advanced rectal cancer. The use of oxaliplatin as adjuvant treatment in elderly patients with colon cancer is controversial. We therefore investigated the impact of age on clinical outcome in the CAO/ARO/AIO-04 phase III trial. Patients and methods: We carried out a post hoc analysis of the CAO/ARO/AIO-04 phase III trial evaluating primary and secondary end points according to age. Patient and tumor characteristics, NCI CTC adverse events grades 3-4 (version 3.0), dose intensities as well as survival and recurrence data were analyzed in three specified age groups (<60, 60-70, and ≥70 years). The influence of age as a continuous variable on DFS was modeled using a subpopulation treatment effect pattern plot (STEPP) analysis. Results: A total of 1232 patients were assessable. With the exception of Eastern Cooperative Oncology Group status (P < 0.001), no differences in patient and tumor characteristics were noticed between age groups. Likewise, toxicity pattern, dose intensities of CRT and surgical results were similar in all age groups. After a median follow-up of 50 months, in patients aged <60 years a significant benefit of adding oxaliplatin to 5-FU-based CRT and adjuvant chemotherapy was observed for local (P = 0.013) and systemic recurrences (P = 0.023), DFS (P = 0.011), and even overall survival (OS; P = 0.044). The STEPP analysis revealed improved hazard ratios for DFS in patients aged 40-70 years compared with elderly patients treated with oxaliplatin. Conclusion: The addition of oxaliplatin significantly improved DFS and OS in younger patients aged <60 years with advanced rectal cancer. Patients aged ≥70 years had no benefit. Clinical Trials Number: NCT00349076.

Neoadjuvant rectal score as individual-level surrogate for disease-free survival in rectal cancer in the CAO/ARO/AIO-04 randomized phase III trial.

Background: Surrogate end points in rectal cancer after preoperative chemoradiation are lacking as their statistical validation poses major challenges, including confirmation based on large phase III trials. We examined the prognostic role and individual-level surrogacy of neoadjuvant rectal (NAR) score that incorporates weighted cT, ypT and ypN categories for disease-free survival (DFS) in 1191 patients with rectal carcinoma treated within the CAO/ARO/AIO-04 phase III trial. Patients and methods: Cox regression models adjusted for treatment arm, resection status, and NAR score were used in multivariable analysis. The four Prentice criteria (PC1-4) were used to assess individual-level surrogacy of NAR for DFS. Results: After a median follow-up of 50 months, the addition of oxaliplatin to fluorouracil-based chemoradiotherapy (CRT) significantly improved 3-year DFS [75.9% (95% confidence interval [CI] 72.30% to 79.50%) versus 71.3% (95% CI 67.60% to 74.90%); P = 0.034; PC 1) and resulted in a shift toward lower NAR groups (P = 0.034, PC 2) compared with fluorouracil-only CRT. The 3-year DFS was 91.7% (95% CI 88.2% to 95.2%), 81.8% (95% CI 78.4% to 85.1%), and 58.1% (95% CI 52.4% to 63.9%) for low, intermediate, and high NAR score, respectively (P < 0.001; PC 3). NAR score remained an independent prognostic factor for DFS [low versus high NAR: hazard ratio (HR) 4.670; 95% CI 3.106-7.020; P < 0.001; low versus intermediate NAR: HR 1.971; 95% CI 1.303-2.98; P = 0.001] in multivariable analysis. Notwithstanding the inherent methodological difficulty in interpretation of PC 4 to establish surrogacy, the treatment effect on DFS was captured by NAR, supporting satisfaction of individual-level PC 4. Conclusion: Our study validates the prognostic role and individual-level surrogacy of NAR score for DFS within a large randomized phase III trial. NAR score could help oncologists to speed up response-adapted therapeutic decision, and further large phase III trial data sets should aim to confirm trial-level surrogacy.

Detecting treatment-subgroup interactions in clustered data with generalized linear mixed-effects model trees.

Identification of subgroups of patients for whom treatment A is more effective than treatment B, and vice versa, is of key importance to the development of personalized medicine. Tree-based algorithms are helpful tools for the detection of such interactions, but none of the available algorithms allow for taking into account clustered or nested dataset structures, which are particularly common in psychological research. Therefore, we propose the generalized linear mixed-effects model tree (GLMM tree) algorithm, which allows for the detection of treatment-subgroup interactions, while accounting for the clustered structure of a dataset. The algorithm uses model-based recursive partitioning to detect treatment-subgroup interactions, and a GLMM to estimate the random-effects parameters. In a simulation study, GLMM trees show higher accuracy in recovering treatment-subgroup interactions, higher predictive accuracy, and lower type II error rates than linear-model-based recursive partitioning and mixed-effects regression trees. Also, GLMM trees show somewhat higher predictive accuracy than linear mixed-effects models with pre-specified interaction effects, on average. We illustrate the application of GLMM trees on an individual patient-level data meta-analysis on treatments for depression. We conclude that GLMM trees are a promising exploratory tool for the detection of treatment-subgroup interactions in clustered datasets.

Challenges for defining minimal clinically important difference (MCID) after spinal cord injury.

STUDY DESIGN: This is a review article. OBJECTIVES: This study discusses the following: (1) concepts and constraints for the determination of minimal clinically important difference (MCID), (2) the contrasts between MCID and minimal detectable difference (MDD), (3) MCID within the different domains of International Classification of Functioning, disability and health, (4) the roles of clinical investigators and clinical participants in defining MCID and (5) the implementation of MCID in acute versus chronic spinal cord injury (SCI) studies. METHODS: The methods include narrative reviews of SCI outcomes, a 2-day meeting of the authors and statistical methods of analysis representing MDD. RESULTS: The data from SCI study outcomes are dependent on many elements, including the following: the level and severity of SCI, the heterogeneity within each study cohort, the therapeutic target, the nature of the therapy, any confounding influences or comorbidities, the assessment times relative to the date of injury, the outcome measurement instrument and the clinical end-point threshold used to determine a treatment effect. Even if statistically significant differences can be established, this finding does not guarantee that the experimental therapeutic provides a person living with SCI an improved capacity for functional independence and/or an increased quality of life. The MDD statistical concept describes the smallest real change in the specified outcome, beyond measurement error, and it should not be confused with the minimum threshold for demonstrating a clinical benefit or MCID. Unfortunately, MCID and MDD are not uncomplicated estimations; nevertheless, any MCID should exceed the expected MDD plus any probable spontaneous recovery. CONCLUSION: Estimation of an MCID for SCI remains elusive. In the interim, if the target of a therapeutic is the injured spinal cord, it is most desirable that any improvement in neurological status be correlated with a functional (meaningful) benefit.

Boosting - an unusual yet attractive optimiser.

This editorial is part of a For-Discussion-Section of Methods of Information in Medicine about the papers "The Evolution of Boosting Algorithms - From Machine Learning to Statistical Modelling" and "Extending Statistical Boosting - An Overview of Recent Methodological Developments", written by Andreas Mayr and co-authors. It preludes two discussed reviews on developments and applications of boosting in biomedical research. The two review papers, written by Andreas Mayr, Harald Binder, Olaf Gefeller, and Matthias Schmid, give an overview on recently published methods that utilise gradient or likelihood-based boosting for fitting models in the life sciences. The reviews are followed by invited comments by experts in both boosting theory and applications.

Histone macroH2A isoforms predict the risk of lung cancer recurrence.

Lung cancer is the leading cause of cancer deaths. Despite optimal diagnosis and early treatment, many patients die of recurrent disease. There are no sufficiently useful biomarkers to predict the risk of tumor recurrence. Here, we show that expression of histone macroH2A1.1 and macroH2A2 predicts lung cancer recurrence, identifying these histone variants as a novel tool for an improved risk stratification of cancer patients. Moreover, macroH2A isoforms are highly expressed in cells undergoing senescence, a known antitumor mechanism, suggesting macroH2A1.1 may be a useful biomarker for senescent cells in tumors.

Weight estimation by three-dimensional ultrasound imaging in the small fetus.

OBJECTIVES: To improve birth weight estimation in fetuses weighing

Flexible modeling of malignant melanoma survival.

OBJECTIVES: This paper compares the diagnostic capabilities of flexible ensemble methods modeling the survival time of melanoma patients in comparison to the well established proportional hazards model. Both a random forest type algorithm for censored data as well as a model combination of the proportional hazards model with recursive partitioning are investigated. METHODS: Benchmark experiments utilizing the integrated Brier score as a measure for goodness of prediction are the basis of the performance assessment for all competing algorithms. For the purpose of comparing regression relationships represented by the models under test, we describe fitted conditional survival functions by a univariate measure derived from the area under the curve. Based on this measure, we adapt a visualization technique useful for the inspection and comparison of model fits. RESULTS: For the data of malignant melanoma patients the predictive performance of the competing models is on par, allowing for a fair comparison of the fitted relationships. Newly introduced MODplots visualize differences in the fitting structure of the underlying models. CONCLUSION: The paper provides a framework for comparing the predictive and diagnostic performance of a parametric, a non-parametric and a combined approach.

Anatomical accuracy of interactive and automated rigid registration between X-ray CT and FDG-PET.

AIM: Comparison of anatomical accuracy of software-based interactive (IRR) and automated rigid registration (ARR) of separately acquired CT and FDG-PET data sets. PATIENTS, METHODS: Independently acquired PET and helical CT data from 22 tumour patients were registered manually using the Syngo advanced Fusion VC20H tool. IRR was performed separately for the thorax and the abdomen using physiological FDG uptake in several organs as a reference. In addition, ARR was performed with the commercially available software tool Mirada 7D on all of the patients. For both methods, the distances between the representation of 53 malignant lesions on PET and CT were measured in X-, Y-, and Z-direction with reference to a common coordinate system (X-, Y-, Z-distances). RESULTS: The percentage of lesions misregistered by less than 1.5 cm was in X-direction 91% for IRR and 89% for ARR; in Y-direction 85% and 68%; in Z-direction 72% and 51%, respectively. The average X-, Y- and Z-distances for IRR ranged from 0.58 +/- 0.55 cm (X-direction) to 1.17 +/- 1.66 cm (Z-direction). For ARR, the average X-, Y- and Z-distances varied between 0.66 +/- 0.61 cm (X-direction) and 1.81 +/- 1.37 cm (Z-direction). Mixed effects analysis of the absolute X-, Y- and Z-distances revealed a significantly better alignment for IRR compared to ARR in Z-direction (p < 0.01). Lesion size and localization either in thorax or abdomen had no significant influence on the accuracy of registration. CONCLUSION: For the majority of malignant lesions, manual image registration with the possibility to separately align different body segments was more accurate than the automated approach. Current software for ARR does not reach the anatomical accuracy reported for PET/CT hybrid scanners.

Evaluation of carotid artery stenosis with multisection CT and MR imaging: influence of imaging modality and postprocessing.

BACKGROUND AND PURPOSE: We prospectively evaluated the influence of different imaging techniques (time-of-flight MR angiography [TOF-MRA], contrast-enhanced MR angiography [CE-MRA], multisection CT angiography [CTA]) and postprocessing methods (maximum intensity projection [MIP], multiplanar reformation [MPR]) on carotid artery stenosis grading. MATERIALS AND METHODS: Fifty patients (34 men, 16 women) with symptomatic stenosis of the internal carotid artery were examined with a 16-section spiral CT and a 1.5T MR unit. Two MRA techniques were applied: 3D-TOF and CE-MRA. MPR was used for postprocessing with all modalities; MIP was used only with MRA. Four readers measured and calculated the percentage diameter stenosis independently according to NASCET criteria. The Wilcoxon test was used to measure interobserver variability, and the Friedman test was used to test the null-hypothesis of equality of the modalities. RESULTS: The hypothesis for global equality was rejected (P < .001). TOF-MRA and CTA assessed with MPR showed the highest concordance (difference, 0.6%; confidence interval [CI], -3.0, 4.3%), and CE-MRA with MIP and CTA showed the lowest concordance in stenosis grading (difference, 7.0%; CI, 3.4, 10.6%). MPR resulted in lower degrees of stenosis than MIP for both MRA sequences, although not statistically significant (CE, -3.0%; CI, -6.6, 0.6%; TOF, -2.2%; CI, -5.8, 1.4%). When only studies with good or excellent image quality were considered, the differences decreased, but the trends remained. CONCLUSION: Stenosis grading is dependent on the examination method and postprocessing technique. CTA and TOF-MRA evaluated with MPR revealed highest concordance.

Association between split selection instability and predictive error in survival trees.

OBJECTIVES: To evaluate split selection instability in six survival tree algorithms and its relationship with predictive error by means of a bootstrap study. METHODS: We study the following algorithms: logrank statistic with multivariate p-value adjustment without pruning (LR), Kaplan-Meier distance of survival curves (KM), martingale residuals (MR), Poisson regression for censored data (PR), within-node impurity (WI), and exponential log-likelihood loss (XL). With the exception of LR, initial trees are pruned by using split-complexity, and final trees are selected by means of cross-validation. We employ a real dataset from a clinical study of patients with gallbladder stones. The predictive error is evaluated using the integrated Brier score for censored data. The relationship between split selection instability and predictive error is evaluated by means of box-percentile plots, covariate and cutpoint selection entropy, and cutpoint selection coefficients of variation, respectively, in the root node. RESULTS: We found a positive association between covariate selection instability and predictive error in the root node. LR yields the lowest predictive error, while KM and MR yield the highest predictive error. CONCLUSIONS: The predictive error of survival trees is related to split selection instability. Based on the low predictive error of LR, we recommend the use of this algorithm for the construction of survival trees. Unpruned survival trees with multivariate p-value adjustment can perform equally well compared to pruned trees. The analysis of split selection instability can be used to communicate the results of tree-based analyses to clinicians and to support the application of survival trees.

Retrospective interactive rigid fusion of (18)F-FDG-PET and CT. Additional diagnostic information in melanoma patients.

AIM: This study investigates whether interactive rigid fusion of routine PET and CT data improves localization, detection and characterization of lesions compared to separate reading. For this purpose, routine PET and CT scans of patients with metastases from malignant melanoma were used. PATIENTS, METHODS: In 34 patients with histologically confirmed malignant melanoma, FDG-PET and spiral CT were performed using clinical standard protocols. For all of these patients, gold standard was available. Clinical and radiological follow-up identified 82 lesions as definitely pathological. Two board-certified nuclear medicine physicians and two board-certified radiologists analyzed PET and CT images independently from each other. For each patient up to 32 anatomical regions (24 lymph node regions, 8 extranodular regions) were systematically classified. Discordant areas were interactively analyzed in manually and rigidly registered images using a commercially available fusion tool. No side-by-side reading was performed. RESULTS: Image fusion disclosed that the evaluation of the PET images alone led to a mislocalization in 26 of 91 focally FDG enhancing lesions. The overall sensitivities of PET, CT, and image fusion were 85, 88, and 94%, respectively; the overall specificities of PET, CT and image fusion were 98, 95 and 100%, respectively. Image fusion exhibited statistically significant higher specificity values as compared with CT. Ten definitely malignant sites were false-negative in CT, but could be detected by PET. On the other hand, twelve metastases were false-negative in PET, but could be detected by CT. These included two lesions, which had a clear correlate on the PET image when the fused images were evaluated. On the whole, registration of the PET and CT images yielded additional diagnostic information in 44% of the definitely malignant lesions. CONCLUSION: Retrospective image fusion of independently obtained PET and CT data is particularly valuable in exactly localizing foci of abnormal FDG uptake and improves the detection of metastases of malignant melanoma.

Bone-implant interface shear modulus and ultimate stress in a transcortical rabbit model of open-pore Ti6Al4V implants.

This experimental study on laser-textured implants aimed to evaluate periimplant bone elasticity and ultimate stress of the bone-implant interface in a rabbit femur model. After randomization, two cylindrical Ti6Al4V samples (3.5 mm wide, 5.5 mm long) were transcortically implanted in each femur of 15 female New Zealand White Rabbits. Polished implants had been laser-textured with 100, 200, and 300 microm diameter pores, and another corundum blasted implant was additionally textured with 200 microm pores. Twelve weeks into the experiment, a modified push-out test was performed. The median shear modulus indicating the elasticity of the periimplant bone was 41.12 MPa for the proximal implant location and 25.38 MPa for the distal, without evidence for significant differences between implant types. Taking into account the median ultimate shear stress for 200 microm implants with and without corundum blasting, no significant difference could be demonstrated. However, for blasted 200 microm implants a statistically significant (p<0.025) relative gain in ultimate shear stress of 41% and 17% was proven in comparison with 100 and 300 microm implants, respectively. Non-blasted 200 microm implants reached 48% relative gain in respect of 100 microm samples.

[Anatomical accuracy of lesion localization. Retrospective interactive rigid image registration between 18F-FDG-PET and X-ray CT]. / Genauigkeit der Lokalisierung maligner Herde mittels retrospektiver, interaktiver starrer Bildfusion von FDG-PET und CT.

UNLABELLED: The aim of this study was to evaluate the anatomical accuracy and reproducibility of retrospective interactive rigid image registration (RIR) between routinely archived X-ray computer tomography (CT) and positron emission tomography performed with 18F-deoxyglucose (FDG-PET) in oncological patients. METHODS: Two observers registered PET and CT data obtained in 37 patients using a commercially available image fusion tool. RIR was performed separately for the thorax and the abdomen using physiological FDG uptake in several organs as a reference. One observer performed the procedure twice (O1a and O1b), another person once (O2). For 94 malignant lesions, clearly visible in CT and PET, the signed and absolute distances between their representation on PET and CT were measured in X-, Y-, and Z-direction with reference to a coordinate system centered in the CT representation of each lesion (X-, Y-, Z-distances). RESULTS: The mean differences of the signed and absolute distances between O1a, O1b, and O2 did not exceed 3 mm in any dimension. The absolute X-, Y-, and Z-distances ranged between 0.57 +/- 0.58 cm for O1a (X-direction) and 1.12 +/- 1.28 cm for O2 (Z-direction). When averaging the absolute distances measured by O1a, O1b, and O2, the percentage of lesions misregistered by less than 1.5 cm was 91% for the X-, 88% for the Y-, and 77% for the Z-direction. The larger error of fusion determined for the remaining lesions was caused by non-rigid body transformations due to differences in breathing, arm position, or bowel movements between the two examinations. Mixed effects analysis of the signed and absolute X-, Y-, and Z-distances disclosed a significantly greater misalignment in the thorax than in the abdomen as well as axially than transaxially. CONCLUSION: The anatomical inaccuracy of RIR can be expected to be <1.5 cm for the majority of neoplastic foci. Errors of alignment are bigger in the thorax and in Z-direction, due to non-rigid body transformations caused, e.g., by breathing.

Automated segmentation of the optic nerve head for diagnosis of glaucoma.

Glaucoma is the second most common cause of blindness worldwide. Low awareness and high costs connected to glaucoma are reasons to improve methods of screening and therapy. A well-established method for diagnosis of glaucoma is the examination of the optic nerve head using scanning-laser-tomography. This system acquires and analyzes the surface topography of the optic nerve head. The analysis that leads to a diagnosis of the disease depends on prior manual outlining of the optic nerve head by an experienced ophthalmologist. Our contribution presents a method for optic nerve head segmentation and its validation. The method is based on morphological operations, Hough transform, and an anchored active contour model. The results were validated by comparing the performance of different classifiers on data from a case-control study with contours of the optic nerve head manually outlined by an experienced ophthalmologist. We achieved the following results with respect to glaucoma diagnosis: linear discriminant analysis with 27.7% estimated error rate for automated segmentation (aut) and 26.8% estimated error rate for manual segmentation (man), classification trees with 25.2% (aut) and 22.0% (man) and bootstrap aggregation with 22.2% (aut) and 13.4% (man). It could thus be shown that our approach is suitable for automated diagnosis and screening of glaucoma.

Attenuation correction of SPECT images based on separately performed CT: Effect on the measurement of regional uptake values.

AIM: A new software approach uses separately acquired CT images for attenuation correction after retrospective fusion with the SPECT data. This study evaluates the effect of this CT-based attenuation correction on indium-111-pentetreotide-SPECT images. METHODS: Indium-111-pentetreotide-SPECT imaging using a dual-head gamma camera e.cam (Siemens Medical Solutions, Erlangen, Germany) as well as separate spiral computed tomography (CT) was performed in 13 patients. After fusion of SPECT and CT data, the bilinear attenuation coefficients were calculated for each pixel in the CT image volume using their Hounsfield unit values and attenuation-corrected images were reconstructed iteratively (OSEM 2D). Regions of interest (ROIs) were drawn on 24 suspicious foci and background, and target to background ratios were calculated for corrected (TBAC) and uncorrected (TBNAC) images. The shortest distance from the centre of the lesion to the surface of the body (DS) was measured on the corresponding CT slice. Furthermore, ROIs were drawn over the rim and the centre of the liver. Ratios of hepatic count rates for corrected (LRAC) and uncorrected (LRNAC) images were also compared. RESULTS: In lesions located more centrally, TBAC was up to 52% higher, whereas in peripherally located lesions, TBAC was up to 63% lower than TBNAC. The TBAC/TBNAC quotient was linearly correlated with DS. In the liver, attenuation correction resulted in a 35% increase of LRAC compared with LRNAC. CONCLUSIONS: Attenuation correction of SPECT images performed by separately acquired CT data is quick and simple. It improves the contrast between target and background for lesions located more centrally in the body and improves homogeneity of the visualisation of tracer uptake in the liver.

Joint analysis of the NACP-REP1 marker within the alpha synuclein gene concludes association with alcohol dependence.

Various studies have linked alcohol dependence phenotypes to chromosome 4. One candidate gene is NACP (non-amyloid component of plaques), coding for alpha synuclein. Recently, it has been shown that alpha synuclein mRNA is increased in alcohol-dependent patients within withdrawal state. This increase is significantly associated with craving, especially obsessive craving. On the basis of these observations, the present study analysed two polymorphic repeats within the NACP gene. We found highly significant longer alleles of NACP-REP1 in alcohol-dependent patients compared with healthy controls (Kruskal-Wallis test, chi(2)=99.5; df=3, P<0.001). In addition, these lengths significantly correlate with levels of expressed alpha synuclein mRNA (chi(2)=8.83; df=2, P=0.012). The present results point to a novel approach for a genetic determination of craving, a key factor in the genesis and maintenance not only of alcoholism but also of addiction in general.

[Web-based training in radiology - student course in the Virtual University of Bavaria]. / Web-basiertes Training in der Radiologie - Studentenkurs in der Virtuellen Hochschule Bayern (VHB).

PURPOSE: The ninth version of the licensing regulation for medical doctors (Approbation Regulation (AR)) sets a benchmark in terms of practical experience, interdigitation of preclinical and clinical studies, interdisciplinary approach, economic efficiency, independence of students, added new teaching and learning modalities, and ongoing evaluation of the progress of the medical students. It is the aim to implement these major points of the AR in a model course for diagnostic radiology and radiation protection within the scope of the Virtual University of Bavaria and test them in practice. MATERIALS AND METHODS: In cooperation with residents and board certified radiologists, students developed the virtual course "Web-Based Training (WBT) Radiology" in diagnostic radiology and radiation protection for students in the first clinical semester. A representative target group taken from the student body was asked about the options to get access to the World Wide Web (Internet), and the satisfaction concerning configuration and content of the newly developed program. A comparison was made between the results of the final examination taken by students who made use of the virtual course in addition to conventional lessons and taken by students who did not subscribe to the virtual course and exclusively relied on conventional lessons. In addition, a pilot study was conducted in the winter semester 2002/03, which compared students taking either the traditional lessons or the new virtual course on the Internet. RESULTS: The virtual course-model had test results with a positive trend. All targeted students had Internet access. Constructive criticism was immediately implemented and contributed to rapid optimization. The learning success of the additive or alternative virtual course was in no way less than the learning success achieved with the conventional course. CONCLUSION: The learning success as measure of quality in teaching and the acceptance by students and teachers justify the continuation of this course model and its expansion. Besides enabling the learning in small study groups; the course "WBT Radiology" might not only help implementing the major points of the new AR but might also complement any deficiencies in the current education. Economic aspects may encourage their implementations.

Simulation based analysis of automated, classification of medical images.

OBJECTIVES: The ability of various classifiers to discriminate between normal and glaucomatous eyes based on features derived from automated analysis of laser scanning images of the eye background is investigated. METHODS: To compare the classifiers without over-optimization for a given dataset, we use a simulation model to create topography images. We designed three different simulation setups as model of extreme situations and medical subgroups. RESULTS: Neither linear nor tree-based classifiers are ideal for all setups. The most robust performance is obtained by a combination of both, so-called Double-Bagging. Classification of real data from a case-control study shows best results with Double-Bagging. All results obtained with the analysis method extracting features automatically are worse than those obtained by the same classifiers but with features derived from an analysis method that requires intervention of a physician. CONCLUSIONS: Robust classification results for classification of laser scanning images obtained with the Heidelberg Retina Tomograph are achieved by combined classifiers. The examined automated procedure causes an increased misclassification error compared to the established clinical routine requiring an expert physician's intervention.