Utilizing machine learning dimensionality reduction for risk stratification of chest pain patients in the emergency department

BACKGROUND

Chest pain is among the most common presenting complaints in the emergency department (ED). Swift and accurate risk stratification of chest pain patients in the ED may improve patient outcomes and reduce unnecessary costs. Traditional logistic regression with stepwise variable selection has been used to build risk prediction models for ED chest pain patients. In this study, we aimed to investigate if machine learning dimensionality reduction methods can improve performance in deriving risk stratification models.

METHODS

A retrospective analysis was conducted on the data of patients > 20 years old who presented to the ED of Singapore General Hospital with chest pain between September 2010 and July 2015. Variables used included demographics, medical history, laboratory findings, heart rate variability (HRV), and heart rate n-variability (HRnV) parameters calculated from five to six-minute electrocardiograms (ECGs). The primary outcome was 30-day major adverse cardiac events (MACE), which included death, acute myocardial infarction, and revascularization within 30 days of ED presentation. We used eight machine learning dimensionality reduction methods and logistic regression to create different prediction models. We further excluded cardiac troponin from candidate variables and derived a separate set of models to evaluate the performance of models without using laboratory tests. Receiver operating characteristic (ROC) and calibration analysis was used to compare model performance.

RESULTS

Seven hundred ninety-five patients were included in the analysis, of which 247 (31%) met the primary outcome of 30-day MACE. Patients with MACE were older and more likely to be male. All eight dimensionality reduction methods achieved comparable performance with the traditional stepwise variable selection; The multidimensional scaling algorithm performed the best with an area under the curve of 0.901. All prediction models generated in this study outperformed several existing clinical scores in ROC analysis.

CONCLUSIONS

Dimensionality reduction models showed marginal value in improving the prediction of 30-day MACE for ED chest pain patients. Moreover, they are black box models, making them difficult to explain and interpret in clinical practice.

Adaptive denoising for chemical exchange saturation transfer MR imaging

High image signal-to-noise ratio (SNR) is required to reliably detect the inherently small chemical exchange saturation transfer (CEST) effects in vivo. In this study, it was demonstrated that identifying spectral redundancies of CEST data by principal component analysis (PCA) in combination with an appropriate data-driven extraction of relevant information can be used for an effective and robust denoising of CEST spectra. The relationship between the number of relevant principal components and SNR was studied on fitted in vivo Z-spectra with artificially introduced noise. Three different data-driven criteria to automatically determine the optimal number of necessary components were investigated. In addition, these criteria facilitate straightforward assessment of data quality that could provide guidance for CEST MR protocols in terms of SNR. Insights were applied to achieve a robust denoising of highly sampled low power Z-spectra of the human brain at 3 and 7 T. The median criterion provided the best estimation for the optimal number of components consistently for all three investigated artificial noise levels. Application of the denoising technique to in vivo data revealed a considerable increase in image quality for the amide and rNOE contrast with a considerable SNR gain. At 7 T the denoising capability was quantified to be comparable or even superior to an averaging of six measurements. The proposed denoising algorithm enables an efficient and robust denoising of CEST data by combining PCA with appropriate data-driven truncation criteria. With this generally applicable technique at hand, small CEST effects can be reliably detected without the need for repeated measurements.

Principal component analysis fosr fast and model-free denoising of multi b-value diffusion-weighted MR images

Despite the utility of tumour characterisation using quantitative parameter maps from multi-b-value diffusion-weighted MRI (DWI), clinicians often prefer the use of the image with highest diffusion-weighting (b-value), for instance for defining regions of interest (ROIs). However, these images are typically degraded by noise, as they do not utilize the information from the full acquisition. We present a principal component analysis (PCA) approach for model-free denoising of DWI data. PCA-denoising was compared to synthetic MRI, where a diffusion model is fitted for each voxel and a denoised image at a given b-value is generated from the model fit. A quantitative comparison of systematic and random errors was performed on data simulated using several diffusion models (mono-exponential, bi-exponential, stretched-exponential and kurtosis). A qualitative visual comparison was also performed for in vivo images in six healthy volunteers and three pancreatic cancer patients. In simulations, the reduction in random errors from PCA-denoising was substantial (up to 55%) and similar to synthetic MRI (up to 53%). Model-based synthetic MRI denoising resulted in substantial (up to 29% of signal) systematic errors, whereas PCA-denoising was able to denoise without introducing systematic errors (less than 2%). In vivo, the signal-to-noise ratio (SNR) and sharpness of PCA-denoised images were superior to synthetic MRI, resulting in clearer tumour boundaries. In the presence of motion, PCA-denoising did not cause image blurring, unlike image averaging or synthetic MRI. Multi-b-value MRI can be denoised model-free with our PCA-denoising strategy that reduces noise to a level similar to synthetic MRI, but without introducing systematic errors associated with the synthetic MRI method.

Improved accuracy of quantitative parameter estimates in dynamic contrast-enhanced CT study with low temporal resolution

PURPOSE

A previously proposed method to reduce radiation dose to patient in dynamic contrast-enhanced (DCE) CT is enhanced by principal component analysis (PCA) filtering which improves the signal-to-noise ratio (SNR) of time-concentration curves in the DCE-CT study. The efficacy of the combined method to maintain the accuracy of kinetic parameter estimates at low temporal resolution is investigated with pixel-by-pixel kinetic analysis of DCE-CT data.

METHODS

The method is based on DCE-CT scanning performed with low temporal resolution to reduce the radiation dose to the patient. The arterial input function (AIF) with high temporal resolution can be generated with a coarsely sampled AIF through a previously published method of AIF estimation. To increase the SNR of time-concentration curves (tissue curves), first, a region-of-interest is segmented into squares composed of 3 × 3 pixels in size. Subsequently, the PCA filtering combined with a fraction of residual information criterion is applied to all the segmented squares for further improvement of their SNRs. The proposed method was applied to each DCE-CT data set of a cohort of 14 patients at varying levels of down-sampling. The kinetic analyses using the modified Tofts' model and singular value decomposition method, then, were carried out for each of the down-sampling schemes between the intervals from 2 to 15 s. The results were compared with analyses done with the measured data in high temporal resolution (i.e., original scanning frequency) as the reference.

RESULTS

The patients' AIFs were estimated to high accuracy based on the 11 orthonormal bases of arterial impulse responses established in the previous paper. In addition, noise in the images was effectively reduced by using five principal components of the tissue curves for filtering. Kinetic analyses using the proposed method showed superior results compared to those with down-sampling alone; they were able to maintain the accuracy in the quantitative histogram parameters of volume transfer constant [standard deviation (SD), 98th percentile, and range], rate constant (SD), blood volume fraction (mean, SD, 98th percentile, and range), and blood flow (mean, SD, median, 98th percentile, and range) for sampling intervals between 10 and 15 s.

CONCLUSIONS

The proposed method of PCA filtering combined with the AIF estimation technique allows low frequency scanning for DCE-CT study to reduce patient radiation dose. The results indicate that the method is useful in pixel-by-pixel kinetic analysis of DCE-CT data for patients with cervical cancer.

Sensitivity enhancement of (Hyper-)CEST image series by exploiting redundancies in the spectral domain

CEST has proven to be a valuable technique for the detection of hyperpolarized xenon-based functionalized contrast agents. Additional information can be encoded in the spectral dimension, allowing the simultaneous detection of multiple different biosensors. However, owing to the low concentration of dissolved xenon in biological tissue, the signal-to-noise ratio (SNR) of Hyper-CEST data is still a critical issue. In this work, we present two techniques aiming to increase SNR by exploiting the typically high redundancy in spectral CEST image series: PCA-based post-processing and sub-sampled acquisition with low-rank reconstruction. Each of them yields a significant SNR enhancement, demonstrating the feasibility of the two approaches. While the first method is directly applicable to proton CEST experiments as well, the second one is particularly beneficial when dealing with hyperpolarized nuclei, since it distributes the non-renewable initial polarization more efficiently over the sampling points. The results obtained are a further step towards the detection of xenon biosensors with spectral Hyper-CEST imaging in vivo.

Improving quantitative CT perfusion parameter measurements using principal component analysis

RATIONALE AND OBJECTIVES

To evaluate the improvements in measurements of blood flow (BF), blood volume (BV), and permeability-surface area product (PS) after principal component analysis (PCA) filtering of computed tomography (CT) perfusion images. To evaluate the improvement in CT perfusion image quality with poor contrast-to-noise ratio (CNR) in vivo.

MATERIALS AND METHODS

A digital phantom with CT perfusion images reflecting known values of BF, BV, and PS was created and was filtered using PCA. Intraclass correlation coefficients and Bland-Altman analysis were used to assess reliability of measurements and reduction in measurement errors, respectively. Rats with C6 gliomas were imaged using CT perfusion, and the raw CT perfusion images were filtered using PCA. Differences in CNR, BF, BV, and PS before and after PCA filtering were assessed using repeated measures analysis of variance.

RESULTS

From simulation, mean errors decreased from 12.8 (95% confidence interval [CI] = -19.5 to 45.0) to 1.4 mL/min/100 g (CI = -27.6 to 30.4), 0.2 (CI = -1.1 to 1.4) to -0.1 mL/100 g (CI = -1.1 to 0.8), and 2.9 (CI = -2.4 to 8.1) to 0.2 mL/min/100 g (CI = -3.5 to 3.9) for BF, BV, and PS, respectively. Map noise in BF, BV, and PS were decreased from 51.0 (CI = -3.5 to 105.5) to 11.6 mL/min/100 g (CI = -7.9 to 31.2), 2.0 (CI = 0.7 to 3.3) to 0.5 mL/100 g (CI = 0.1 to 1.0), and 8.3 (CI = -0.8 to 17.5) to 1.4 mL/min/100 g (CI = -0.4 to 3.1), respectively. For experiments, CNR significantly improved with PCA filtering in normal brain (P < .05) and tumor (P < .05). Tumor and brain BFs were significantly different from each other after PCA filtering with four principal components (P < .05).

CONCLUSIONS

PCA improved image CNR in vivo and reduced the measurement errors of BF, BV, and PS from simulation. A minimum of four principal components is recommended.

Standardization of radiological evaluation of dynamic contrast enhanced MRI: application in breast cancer diagnosis

Dynamic contrast enhanced MRI is applied as an adjuvant tool for breast cancer detection, diagnosis, and follow-up of therapy. Despite improvements through the years in achieving higher spatial and temporal resolution, it still suffers from lack of scanning and processing standardization, and consequently, high variability in the radiological evaluation, particularly differentiating malignant from benign lesions. We describe here a hybrid method for achieving standardization of the radiological evaluation of breast dynamic contrast enhanced (DCE)-magnetic resonance imaging (MRI) protocols, based on integrating the model based three time point (3TP) method with principal component analysis (PCA). The scanning and image processing procedures consisted of three main steps: 1. 3TP standardization of the MRI acquisition parameters according to a kinetic model, 2. Applying PCA to test cases and constructing an eigenvectors' base related to the contrast-enhancement kinetics and 3. Projecting all new cases on the eigenvectors' base and evaluating the clinical outcome. Datasets of overall 96 malignant and 26 benign breast lesions were recorded on 1.5T and 3T scanners, using three different MRI acquisition parameters optimized by the 3TP method. The final radiological evaluation showed similar detection and diagnostic ability for the three different MRI acquisition parameters. The area under the curve of receiver operating characteristic analysis yielded a value of 0.88 ± 0.034 for differentiating malignant from benign lesions. This 3TP + PCA hybrid method is fast and can be readily applied as a computer aided diagnostic tool of breast cancer. The underlying principles of this method can be extended to standardize the evaluation of malignancies in other organs.

Measurement of myocardial extracellular volume fraction by using equilibrium contrast-enhanced CT: validation against histologic findings

PURPOSE

To develop and validate equilibrium contrast material-enhanced computed tomography (CT) to measure myocardial extracellular volume (ECV) fraction by using a histologic reference standard and to compare equilibrium CT with equilibrium contrast-enhanced magnetic resonance (MR) imaging.

MATERIALS AND METHODS

A local ethics committee approved the study, and all subjects gave fully informed written consent. An equilibrium CT protocol was developed using iohexol at 300 mg of iodine per milliliter (bolus of 1 mg per kilogram of body weight administered at a rate of 3 mL/sec, followed immediately by an infusion of 1.88 mL/kg per hour with CT imaging before and at 25 minutes after injection of bolus of contrast agent) and ECV within the myocardial septum measured using both equilibrium CT and equilibrium MR imaging in patients with severe aortic stenosis. Biopsy samples of the myocardial septum collected during valve replacement surgery were used for histologic quantification of extracellular fibrosis with picrosirius red staining. Equilibrium CT- and equilibrium MR imaging-derived ECV measurements were compared with histologically quantified fibrosis by using Pearson correlation. Agreement between equilibrium CT and equilibrium MR imaging was assessed by using Bland-Altman comparison.

RESULTS

Twenty-three patients (16 male, seven female; mean age, 70.8 years; standard deviation, 8.3) were recruited. The mean percentage of histologic fibrosis was 18% (intersubject range, 5%-40%). There was a significant correlation between both equilibrium CT- and equilibrium MR imaging-derived ECV and percentage of histologic fibrosis (r = 0.71 [P < .001] and r = 0.84 [P < .0001], respectively). Equilibrium CT-derived ECV was significantly correlated to equilibrium MR imaging-derived ECV (r = 0.73).

CONCLUSION

ECV measured by using equilibrium CT in patients with aortic stenosis correlates with histologic quantification of myocardial fibrosis and with ECV derived by using equilibrium MR imaging.

The 2D Hotelling filter - a quantitative noise-reducing principal-component filter for dynamic PET data, with applications in patient dose reduction

Background

In this paper we apply the principal-component analysis filter (Hotelling filter) to reduce noise from dynamic positron-emission tomography (PET) patient data, for a number of different radio-tracer molecules. We furthermore show how preprocessing images with this filter improves parametric images created from such dynamic sequence.

We use zero-mean unit variance normalization, prior to performing a Hotelling filter on the slices of a dynamic time-series. The Scree-plot technique was used to determine which principal components to be rejected in the filter process. This filter was applied to [¹¹C]-acetate on heart and head-neck tumors, [¹⁸F]-FDG on liver tumors and brain, and [¹¹C]-Raclopride on brain. Simulations of blood and tissue regions with noise properties matched to real PET data, was used to analyze how quantitation and resolution is affected by the Hotelling filter. Summing varying parts of a 90-frame [¹⁸F]-FDG brain scan, we created 9-frame dynamic scans with image statistics comparable to 20 MBq, 60 MBq and 200 MBq injected activity. Hotelling filter performed on slices (2D) and on volumes (3D) were compared.

Results

The 2D Hotelling filter reduces noise in the tissue uptake drastically, so that it becomes simple to manually pick out regions-of-interest from noisy data. 2D Hotelling filter introduces less bias than 3D Hotelling filter in focal Raclopride uptake. Simulations show that the Hotelling filter is sensitive to typical blood peak in PET prior to tissue uptake have commenced, introducing a negative bias in early tissue uptake. Quantitation on real dynamic data is reliable. Two examples clearly show that pre-filtering the dynamic sequence with the Hotelling filter prior to Patlak-slope calculations gives clearly improved parametric image quality. We also show that a dramatic dose reduction can be achieved for Patlak slope images without changing image quality or quantitation.

Conclusions

The 2D Hotelling-filtering of dynamic PET data is a computer-efficient method that gives visually improved differentiation of different tissues, which we have observed improve manual or automated region-of-interest delineation of dynamic data. Parametric Patlak images on Hotelling-filtered data display improved clarity, compared to non-filtered Patlak slope images without measurable loss of quantitation, and allow a dramatic decrease in patient injected dose.

Computed tomography perfusion imaging for therapeutic assessment: has it come of age as a biomarker in oncology?

With the emergence of novel targeted therapies, imaging techniques that assess tumor vascular support have gained credence for response assessment alongside standard response criteria. Computed tomography (CT) perfusion techniques that quantify regional tumor blood flow, blood volume, flow-extraction product, and permeability-surface area product through standard kinetic models are attractive, but the level of evidence for CT perfusion to be a truly mature biomarker remains insufficient. Studies to date have not been powered to assess this. Future studies that include good quality prospective validation correlating perfusion CT to outcome end points in the trial setting are needed to take CT perfusion forward as a biomarker in oncology.