Claudia Beleites

Sample size planning for classification models

In biospectroscopy, suitably annotated and statistically independent samples (e.g. patients, batches, etc.) for classifier training and testing are scarce and costly. Learning curves show the model performance as function of the training sample size and can help to determine the sample size needed to train good classifiers. However, building a good model is actually not enough: the performance must also be proven. We discuss learning curves for typical small sample size situations with 5-25 independent samples per class. Although the classification models achieve acceptable performance, the learning curve can be completely masked by the random testing uncertainty due to the equally limited test sample size. In consequence, we determine test sample sizes necessary to achieve reasonable precision in the validation and find that 75-100 samples will usually be needed to test a good but not perfect classifier. Such a data set will then allow refined sample size planning on the basis of the achieved performance. We also demonstrate how to calculate necessary sample sizes in order to show the superiority of one classifier over another: this often requires hundreds of statistically independent test samples or is even theoretically impossible. We demonstrate our findings with a data set of ca. 2550 Raman spectra of single cells (five classes: erythrocytes, leukocytes and three tumour cell lines BT-20, MCF-7 and OCI-AML3) as well as by an extensive simulation that allows precise determination of the actual performance of the models in question.

Characterization of atherosclerotic plaque depositions by Raman and FTIR imaging

Spectroscopy-based imaging techniques can provide useful biochemical information about tissue samples. Here, we employ Raman and Fourier transform infrared (IR) imaging to characterize composition and constitution of atherosclerotic plaques of rabbits, fed with a high cholesterol diet. The results were compared with conventional light microscopy after staining with hematoxylin eosin, and elastica van Gieson. The spectral unmixing algorithm vertex component analysis was applied for data analysis and image reconstruction. IR microscopy allowed for differentiation between lipids and proteins in plaques of full aortic cross sections. Raman microscopy further discriminated cholesterol esters, cholesterol and triglycerides. FTIR and Raman images were recorded at a resolution near 20 micrometer per pixel for a large field of view. High resolution Raman images at 1 micrometer per pixel revealed structural details at selected regions of interest. The intima-media and the lipid-protein ratio were determined in five specimens for quantitation. These results correlate well with histopathology. The described method is a promising tool for easy and fast molecular imaging of atherosclerosis.

Poly-l-lysine-Coated Silver Nanoparticles as Positively Charged Substrates for Surface-Enhanced Raman Scattering

Positively charged nanoparticles to be used as substrates for surface-enhanced Raman scattering (SERS) were prepared by coating citrate-reduced silver nanoparticles with the cationic polymer poly-l-lysine. The average diameter of the coated nanoparticles is 75 nm, and their zeta potential is +62.3 ± 1.7 mV. UV-vis spectrophotometry and dynamic light scattering measurements show that no aggregation occurs during the coating process. As an example of their application, the so-obtained positively charged coated particles were employed to detect nanomolar concentrations of the anionic chromophore bilirubin using SERS. Because of their opposite charge, bilirubin molecules interact with the coated nanoparticles, allowing SERS detection. The SERS intensity increases linearly with concentration in a range from 10 to 200 nM, allowing quantitative analysis of bilirubin aqueous solutions.

Combining multiset resolution and segmentation for hyperspectral image analysis of biological tissues.

Hyperspectral images can provide useful biochemical information about tissue samples. Often, Fourier transform infrared (FTIR) images have been used to distinguish different tissue elements and changes caused by pathological causes. The spectral variation between tissue types and pathological states is very small and multivariate analysis methods are required to describe adequately these subtle changes. In this work, a strategy combining multivariate curve resolution-alternating least squares (MCR-ALS), a resolution (unmixing) method, which recovers distribution maps and pure spectra of image constituents, and K-means clustering, a segmentation method, which identifies groups of similar pixels in an image, is used to provide efficient information on tissue samples. First, multiset MCR-ALS analysis is performed on the set of images related to a particular pathology status to provide basic spectral signatures and distribution maps of the biological contributions needed to describe the tissues. Later on, multiset segmentation analysis is applied to the obtained MCR scores (concentration profiles), used as compressed initial information for segmentation purposes. The multiset idea is transferred to perform image segmentation of different tissue samples. Doing so, a difference can be made between clusters associated with relevant biological parts common to all images, linked to general trends of the type of samples analyzed, and sample-specific clusters, that reflect the natural biological sample-to-sample variability. The last step consists of performing separate multiset MCR-ALS analyses on the pixels of each of the relevant segmentation clusters for the pathology studied to obtain a finer description of the related tissue parts. The potential of the strategy combining multiset resolution on complete images, multiset segmentation and multiset local resolution analysis will be shown on a study focused on FTIR images of tissue sections recorded on inflamed and non-inflamed palatine tonsils.

Validation of soft classification models using partial class memberships: An extended concept of sensitivity & co. applied to grading of astrocytoma tissues

Abstract We use partial class memberships in soft classification to model uncertain labeling and mixtures of classes. Partial class memberships are not restricted to predictions, but may also occur in reference labels (ground truth, gold standard diagnosis) for training and validation data. Classifier performance is usually expressed as fractions of the confusion matrix, like sensitivity, specificity, negative and positive predictive values. We extend this concept to soft classification and discuss the bias and variance properties of the extended performance measures. Ambiguity in reference labels translates to differences between best-case, expected and worst-case performance. We show a second set of measures comparing expected and ideal performance which is closely related to regression performance, namely the root mean squared error RMSE and the mean absolute error MAE. All calculations apply to classical crisp as well as to soft classification (partial class memberships as well as one-class classifiers). The proposed performance measures allow to test classifiers with actual borderline cases. In addition, hardening of e.g. posterior probabilities into class labels is not necessary, avoiding the corresponding information loss and increase in variance. We implemented the proposed performance measures in R package “softclassval” which is available from CRAN and at softclassval.r-forge.r-project.org . Our reasoning as well as the importance of partial memberships for chemometric classification is illustrated by a real-word application: astrocytoma brain tumor tissue grading (80 patients, 37,000 spectra) for finding surgical excision borders. As borderline cases are the actual target of the analytical technique, samples which are diagnosed to be borderline cases must be included in the validation.

Recognition of tumor cells by immuno-SERS-markers in a microfluidic chip at continuous flow

SERS active nanoparticles were labeled with a reporter molecule and conjugated with anti-EpCAM antibodies. These immuno SERS markers were mixed with leukocytes, MCF-7 breast cancer cells and polystyrene beads, and the mixture was injected into a microfluidic quartz chip. Raman spectra were acquired at 785 nm excitation with 25 milliseconds exposure time in a continuous flow regime. Spectral unmixing by N-FINDR identified spectral signatures of SERS-labelled cells and polystyrene beads. This approach demonstrated rapid and reproducible SERS-assisted cell detection. Strategies are discussed to further increase the throughput for cell sorting.

Application of R-mode analysis to Raman maps: a different way of looking at vibrational hyperspectral data

Hierarchical cluster analysis (HCA) is extensively used for the analysis of hyperspectral data. In this work, hyperspectral data sets obtained from Raman maps were analyzed using an alternative mode of cluster analysis, clustering “images” instead of spectra, under the assumption that images showing similar spatial distributions are related to the same chemical species. Such an approach was tested with two Raman maps: one simple “test map” of micro-crystals of four different compounds for a proof of principle and a map of a biological tissue (i.e., cartilage) as an example of chemically complex sample. In both cases, the “image-clustering” approach gave similar results as the traditional HCA, but at lower computational effort. The alternative approach proved to be particularly helpful in cases, as for the cartilage tissue, where concentration gradients of chemical composition are present. Moreover, with this approach, yielded information about correlation between bands in the average spectrum makes band assignment and spectral interpretation easier.

Raman and Infrared Spectroscopy Distinguishing Replicative Senescent from Proliferating Primary Human Fibroblast Cells by Detecting Spectral Differences Mainly Due to Biomolecular Alterations

Cellular senescence is a terminal cell cycle arrested state, assumed to be involved in tumor suppression. We studied four human fibroblast cell strains (BJ, MRC-5, IMR-90, and WI-38) from proliferation into senescence. Cells were investigated by label-free vibrational Raman and infrared spectroscopy, following their transition into replicative senescence. During the transition into senescence, we observed rather similar biomolecular abundances in all four cell strains and between proliferating and senescent cells; however, in the four aging cell strains, we found common molecular differences dominated by protein and lipid modifications. Hence, aging induces a change in the appearance of biomolecules (including degradation and storage of waste) rather than in their amount present in the cells. For all fibroblast strains combined, the partial least squares-linear discriminant analysis (PLS-LDA) model resulted in 75% and 81% accuracy for the Raman and infrared (IR) data, respectively. Within this validation, senescent cells were recognized with 93% sensitivity and 90% specificity for the Raman and 84% sensitivity and 97% specificity for the IR data. Thus, Raman and infrared spectroscopy can identify replicative senescence on the single cell level, suggesting that vibrational spectroscopy may be suitable for identifying and distinguishing different cellular states in vivo, e.g., in skin.

Raman-based identification of circulating tumor cells for cancer diagnosis

Circulating tumor cells (CTCs) that can be extracted from body fluids offer new prospects in cancer diagnostics. An overview about our recent achievements is presented to use Raman-based methodologies to distinguish cancer cells from normal blood cells. In a first approach, a microfluidic chip was developed to collect Raman spectra from optically trapped cells. Whereas sensitivities and specificities were promising, the throughput was not compatible with the expected low number of CTCs per million white blood cells. A second strategy immobilized up to 200,000 cells onto a microhole array made of silicon nitride. Rapid microscopic screening can be applied to pre-select a subset of cells from which Raman spectra are collected for specific CTC identification. As this approach is compatible with living cells and Raman spectroscopy with 785 nm excitation is non-destructive, a robotic arm can select positively identified CTCs for in-depth biochemical assessment. Finally, an in vivo approach directly collects CTCs from the blood stream. This way reduces the cell number to a manageable size that is subjected to Raman spectroscopy for cell typing and enumeration. An integrated acquisition mode was introduced to further increase the throughput and robustness of single cell classification.

Cell identification using Raman spectroscopy in combination with optical trapping and microfluidics

Cell identification by Raman spectroscopy has evolved to be an attractive complement to established optical techniques. Raman activated cell sorting (RACS) offers prospects to complement the widely applied fluorescence activated cell sorting. RACS can be realized by combination with optical traps and microfluidic devices. The progress of RACS is reported for a cellular model system that can be found in peripheral blood of tumor patients. Lymphocytes and erythrocytes were extracted from blood samples. Breast carcinoma derived tumor cells (MCF-7, BT-20) and acute myeloid leukemia cells (OCI-AML3) were grown in cell cultures. First, Raman images were collected from dried cells on calcium fluoride slides. Support vector machines (SVM) classified 99.7% of the spectra to the correct cell type. Second, a 785 nm laser was used for optical trapping of single cells in aqueous buffer and for excitation of the Raman spectrum. SVM distinguished 1210 spectra of tumor and normal cells with a sensitivity of >99.7% and a specificity of >99.5%. Third, a microfluidic glass chip was designed to inject single cells, modify the flow speed, accommodate fibers of an optical trap and sort single cells after Raman based identification with 514 nm for excitation. Forth, the microfluidic chip was fabricated by quartz which improved cell identification results with 785 nm excitation. Here, partial least squares discriminant analysis gave classification rates of 98%. Finally, a Raman-on-chip approach was developed that integrates fibers for trapping, Raman excitation and signal detection in a single compact unit.