Xiaomei Song

Characterization of hemoglobin, water, and NIR scattering in breast tissue: analysis of intersubject variability and menstrual cycle changes.

Near-infrared imaging was used to quantify typical values of hemoglobin concentration, oxygen saturation, water fraction, scattering power, and scattering amplitude within the breast tissue of volunteer subjects. A systematic study of the menstrual variations in these parameters was carried out by measuring a group of seven premenopausal normal women (aged 41 to 47 years) in the follicular (days 7 to 14 of the cycle) and secretory phases (days 21 to 28) of the cycle, for two complete menstrual cycles. An average increase in hemoglobin concentration of 2.6 microM or 13% of the background breast values was observed in the secretory phase relative to the follicular phase (p<0.0001), but no other average near-infrared parameter changes were significant. While repeatable and systematic changes were observed in all parameters for individual subjects, large intersubject variations were present in all parameters. In a survey of thirty-nine normal subjects, the total hemoglobin varied from 9 to 45 microM, with a systematic correlation observed between total hemoglobin concentration and breast radiographic density. Scattering power and scattering amplitude were also correlated with radiographic density, but oxygen saturation and water fraction were not. Images of breast lesions indicate that total hemoglobin-based contrast can be up to 200% relative to the background in the same breast. Yet, since the background hemoglobin values vary considerably among breasts, the maximum hemoglobin concentrations observed in cancer tumors may vary considerably as well. In light of these observations, it may be important to use hemoglobin contrast values relative to the background for a given breast, rather than absolute hemoglobin contrast when trying to compare the features of breast lesions among subjects.

Image analysis methods for diffuse optical tomography.

Three major analytical tools in imaging science are summarized and demonstrated relative to optical imaging in vivo. Standard resolution testing is optimal when infinite contrast is used and hardware evaluation is the goal. However, deep tissue imaging of absorption or fluorescent contrast agents in vivo often presents a different problem, which requires contrast-detail analysis. This analysis shows that the minimum detectable sizes are in the range of 1/10 the outer diameter, whereas minimum detectable contrast values are in the range of 10 to 20% relative to the continuous background values. This is estimated for objects being in the center of the domain being imaged, and as the heterogeneous region becomes closer to the surface, the lower limit on size and contrast can become arbitrarily low and more dictated by hardware specifications. Finally, if human observer detection of abnormalities in the images is the goal, as is standard in most radiological practice, receiver operating characteristic (ROC) curve and location receiver operating characteristic curve (LROC) are used. Each of these three major areas of image interpretation and analysis are reviewed in the context of medical imaging as well as how they are used to quantify the performance of diffuse optical imaging of tissue.

Approximation of Mie scattering parameters in near-infrared tomography of normal breast tissue in vivo

A method for estimating Mie theory scattering parameters from diffuse light tomography measurements in breast tissue is discussed. The approach provides an estimate of the mean particle size and number density given assumptions about the index of refraction change expected in lipid-membrane-bound scatterers. When using a sparse number of wavelengths in the reduced scattering spectra, the parameter extraction technique is limited to representing a continuous distribution of scatterer sizes that appears to be dominated by an exponential decrease with increasing particle size. The fitting method is tested on simulated data and then on Intralipid-based tissue-phantom data, giving a mean particle size of 93+/-17 nm, which is in excellent agreement with expectations. The approach is also applied retrospectively to breast tissue spectra acquired from normal healthy volunteers, where the average particle size and number density were found to be in the range of 20 to 1400 nm. Grouping of the data based on radiographic breast density, as a surrogate measure of tissue composition yielded values of 20 to 65, 25 to 200, 140 to 1200, and 150 to 1400 nm, respectively, for the four BI-RADS (American College of Radiology Breast Imaging Reporting and Data System) density classifications of extremely dense, heterogeneously dense, scattered, and fatty. These results are consistent with the microscopic characteristics of each breast type given the expected progression from predominantly collagenous connective tissue (extremely dense category) to increasing proportions of glandular epithelium and fat (intermediate density categories) to predominantly fat (fatty category).

Improved quantification of small objects in near-infrared diffuse optical tomography

Diffuse optical tomography allows quantification of hemoglobin, oxygen saturation, and water in tissue, and the fidelity in this quantification is dependent on the accuracy of optical properties determined during image reconstruction. In this study, a three-step algorithm is proposed and validated that uses the standard Newton minimization with Levenberg-Marquardt regularization as the first step. The second step is a modification to the existing algorithm using a two-parameter regularization to allow lower damping in a region of interest as compared to background. This second stage allows the recovery of the actual size of an inclusion. A region-based reconstruction is the final third step, which uses the estimated size and position information from step 2 to yield quantitatively accurate average values for the optical parameters. The algorithm is tested on simulated and experimental data and is found to be insensitive to object contrast and position. The percentage error between the true and the average recovered value for the absorption coefficient in test images is reduced from 47 to 27% for a 10-mm inclusion, from 38 to 13% for a 15-mm anomaly, and from 28 to 5.5% for a 20-mm heterogeneity. Simulated data with absorbing and scattering heterogeneities of 15 mm diam located in different positions show recovery with less than 15% error in absorption and 6% error in reduced scattering coefficients. The algorithm is successfully applied to clinical data from a subject with a breast abnormality to yield quantitatively increased absorption coefficients, which enhances the contrast to 3.8 compared to 1.23 previously.

Image reconstruction of effective Mie scattering parameters of breast tissue in vivo with near-infrared tomography

A method for image reconstruction of the effective size and number density of scattering particles is discussed within the context of interpreting near-infrared (NIR) tomography images of breast tissue. An approach to use Mie theory to estimate the effective scattering parameters is examined and applied, given some assumptions about the index of refraction change expected in lipid membrane-bound scatterers. When using a limited number of NIR wavelengths in the reduced scattering spectra, the parameter extraction technique is limited to representing a continuous distribution of scatterer sizes, which is modeled as a simple exponentially decreasing distribution function. In this paper, image formation of effective scatterer size and number density is presented based on the estimation method. The method was evaluated with Intralipid phantom studies to demonstrate particle size estimation to within 9% of the expected value. Then the method was used in NIR patient images, and it indicates that for a cancer tumor, the effective scatterer size is smaller than the background breast values and the effective number density is higher. In contrast, for benign tumor patients, there is not a significant difference in effective scatterer size or number density between tumor and normal tissues. The method was used to interpret magnetic resonance imaging-coupled NIR images of adipose and fibroglandular tissues, and it indicated that the fibroglandular tissue has smaller effective scatterer size and larger effective number density than the adipose tissue does.

Statistical analysis of nonlinearly reconstructed near-infrared tomographic images. I. Theory and simulations

Near-infrared (NIR) diffuse tomography is an emerging method for imaging the interior of tissues to quantify concentrations of hemoglobin and exogenous chromophores noninvasively in vivo. It often exploits an optical diffusion model-based image reconstruction algorithm to estimate spatial property values from measurements of the light flux at the surface of the tissue. In this study, mean-squared error (MSE) over the image is used to evaluate methods for regularizing the ill-posed inverse image reconstruction problem in NIR tomography. Estimates of image bias and image standard deviation were calculated based upon 100 repeated reconstructions of a test image with randomly distributed noise added to the light flux measurements. It was observed that the bias error dominates at high regularization parameter values while variance dominates as the algorithm is allowed to approach the optimal solution. This optimum does not necessarily correspond to the minimum projection error solution, but typically requires further iteration with a decreasing regularization parameter to reach the lowest image error. Increasing measurement noise causes a need to constrain the minimum regularization parameter to higher values in order to achieve a minimum in the overall image MSE.

Statistical analysis of nonlinearly reconstructed near-infrared tomographic images. II. Experimental interpretation

For pt. I see ibid., vol. 21, no. 7, p. 755-63 (2002). Image error analysis of a diffuse near-infrared tomography (NIR) system has been carried out on simulated data using a statistical approach described in pt. I of this paper (Pogue et al., 2002). The methodology is used here with experimental data acquired on phantoms with a prototype imaging system intended for characterizing breast tissue. Results show that imaging performance is not limited by random measurement error, but rather by calibration issues. The image error over the entire field of view is generally not minimized when an accurate homogeneous estimate of the phantom properties is available; however, local image error over a target region of interest (ROI) is reduced. The image reconstruction process which includes a Levenberg-Marquardt style regularization provides good minimization of the objective function, yet its reduction is not always correlated with an overall image error decrease. Minimization of the bias in an ROI which contains localized changes in the optical properties can be achieved through five to nine iterations of the algorithm. Precalibration of the algorithm through statistical evaluation of phantom studies may provide a better measure of the image accuracy than that implied by minimization of the standard objective function.

Receiver operating characteristic and location analysis of simulated near-infrared tomography images

Receiver operating characteristic (ROC) analysis was performed on simulated near-infrared tomography images, using both human observer and contrast-to-noise ratio (CNR) computational assessment, for application in breast cancer imaging. In the analysis, a nonparametric approach was applied for estimating the ROC curves. Human observer detection of objects had superior capability to localize the presence of heterogeneities when the objects were small with high contrast, with a minimum detectable threshold of CNR near 3.0 to 3.3 in the images. Human observers were able to detect heterogeneities in the images below a size limit of 4 mm, yet could not accurately find the location of these objects when they were below 10 mm diameter. For large objects, the lower limit of a detectable contrast limit was near 10% increase relative to the background. The results also indicate that iterations of the nonlinear reconstruction algorithm beyond 4 did not significantly improve the human detection ability, and degraded the overall localization ability for the objects in the image, predominantly by increasing the noise in the background. Interobserver variance performance in detecting objects in these images was low, suggesting that because of the low spatial resolution, detection tasks with NIR tomography is likely consistent between human observers.

Effect of image reconstruction bias upon spectroscopy-based quantification of chromophores in near-infrared tomography

Accuracy of chromophore concentrations in near-infrared tomography is affected by negative bias in the recovery of absorption coefficients and availability of limited wavelengths. The effect of these factors has been studied with suitable error models.

Spectral-encoding for parallel source implementation in NIR tomography

While near-infrared tomography has advanced considerably over the past decade, key technological designs still limit what can be achieved, especially in terms of imaging acquisition speed. One of these fundamental limitations is the requirement that the source light be delivered sequentially or through frequency encoding of the time signal. Sequential delivery inherently limits the speed at which images can be acquired. Modulation frequency-dependent encoding of the sources solves the problem by allowing sources near the same location to be turned on simultaneously, thereby improving the speed for acquisition, but suffers from dynamic range problems. In this study, we demonstrate an alternative parallel source implementation approach which uses spectral wavelength encoding of the source. This new technique allows many sources to be input into the tissue at the same time, as long as the spectrally encoded signals can be decoded at the output. To test the implementation of this approach, 8 single-mode laser diodes of wavelengths distributed within a narrow range of 10 nm are used, and the lights are all input into tissue phantom simultaneously. On the detection side, a high-resolution spectrometer is used to spatially spread out the signals to facilitate parallel detection of the signal from each spectrally-encoded source. This robust approach allows rapid parallel sampling of all sources at all detection locations. The implementation of this technique in a NIR tomography application is examined, and the preliminary results of video-rate imaging at 30 Hz is presented.