An important feature of the Q-DRI device is the acquisition of maps of the specimen margins, reflective of a variety of biochemical and morphological sources of contrast in the breast. Sources of absorption contrast include hemoglobin (both oxygenated and deoxygenated forms) and β-carotene, whereas sources of scattering contrast include the size and density of scatterers within the tissue (e.g., varying number densities of subcellular organelles such as mitochondria, collagen content, and variations in the size of cell nuclei). These extracted parameters may be considered alone, or may be combined (for instance, by computing ratios) to further exploit diagnostic information. Although the Q-DRI device provides extracted parameter maps of the margin with spatial information intact, traditional image analysis is difficult because there is no prior information about the spatial characteristics of positive and close margins to leverage upon. Thus, it is necessary to devise strategies for reducing the image data into substitute variables which can best predict margin status.
One approach to image reduction we have pursued is computation of descriptive statistical variables which capture important features of the margin images. These include, but are not limited to, the mean, maximum, minimum, or variance of the values in a particular extracted parameter image. Another successful approach is to analyze the distribution of values within a particular image, for which histograms are useful tools. For example, by plotting histograms for each margin parameter image, natural distinctions between positive and negative margins may be observed in the distribution of the image values. A number of parameters including, total hemoglobin concentration, β-carotene concentration, <μs′>, hemoglobin saturation, oxy-hemoglobin concentration, deoxy-hemoglobin concentration, and the ratios of various combinations of these parameters, were evaluated.
contains representative maps of two such parameters: the ratio of β-carotene:< μs
′> (left image) as well as the ratio of total hemoglobin:< μs
′> (right image), and the corresponding histograms which graphically represent the distribution of values within each image for a margin negative for residual disease. contains representative images of the same parameters in a margin positive for invasive ductal carcinoma (IDC). These figures are full screenshots taken from the custom software application. β-carotene is a dietary carotenoid known to be stored primarily in adipocytes, and is thus reflective of the amount of fat present in the sensing volume. The wavelength-averaged reduced scattering coefficient, <μs
′>, is a measure of the amount of light elastically scattered in the tissue, with higher scattering coefficients associated with more connective tissue and denser arrangements of cells and their subcellular scatterers such as organelles and membranes (scatter density) as well as with changes in the distribution of sizes of these scatterers (scatter size) [47
Fig. 5 Representative results from the clinical study, showing results from a pathologically-confirmed negative margin. These screenshots depict the β-carotene:< μs′> (left) and total hemoglobin:< μs′> (more ...)
Fig. 6 Representative results from the clinical study, showing results from a pathologically-confirmed positive margin (IDC). The black circles indicate pixels which were confirmed to contain residual invasive ductal carcinoma, whereas white circles indicate (more ...)
Additionally, the extracted total hemoglobin concentration is reflective of the vascular volume present within the sensing volume of the Q-DRI device. (Separate studies have shown that this parameter is stable with respect to time following tissue removal, as described later). It is well known that angiogenesis, or creation of new blood vessels, is a hallmark of cancer. Thus, increased levels of total hemoglobin could be reflective of angiogenic processes in the tissue and thus indicative of cancer.
As observed in , positive margins were hallmarked by decreased levels in both β-carotene:< μs′> as well as total hemoglobin:< μs′>. Therefore, in and , the colormaps are set such that lower values of β-carotene:< μs′> and total hemoglobin:< μs′> appear red, whereas higher values appear blue. As seen in the images, the negative margin images () are characterized by a higher proportion of blue pixels, whereas the positive margin images () are characterized by increased proportions of red pixels, for both parameters of interest. The black circles in the images of indicate the locations of path-confirmed cancerous pixels, and the white circles indicate the locations of normal pixels, which confirm these generalizations. Analysis of these path-confirmed pixels suggests that the most specific contrast appears to be contained in the β-carotene:< μs′> images.
Fig. 7 Boxplots of image-descriptive variables with highest diagnostic potential. A) Boxplot of percentage of β-carotene:< μs′> image pixels below 6 μM-cm, B) boxplot of percentage of total hemoglobin:< (more ...)
It must be noted, that breast tissue is highly heterogeneous, and it is difficult to isolate one or two unique parameters in our data which contain all of the diagnostic information. However, based on a separate analysis of path-confirmed pixels, we suspect that the primary value of dividing β-carotene and total hemoglobin by <μs′> is to normalize for differences in highly scattering collagen content (i.e. breast density) between patients. However, <μs′> is also associated with malignancy and is expected to provide additional diagnostic information in addition to β-carotene and total hemoglobin concentration. Thus, normalization by <μs′> could change the expected trends, since both numerator and denominator are affected by malignancy.
The percentage of pixels below a particular threshold for β-carotene:< μs′> and total hemoglobin:< βs′> exhibited the greatest differences between positive and negative margins in this dataset, as determined by Wilcoxon rank-sum testing. The following text describes how the optimal threshold was selected for each of the above two parameters in order to build predictors for margin-level assessment from the parameter maps. A threshold value for pixel intensity was determined and the percentage of pixels below that threshold was computed. A Wilcoxon Rank Sum test was carried out to determine if the percentage of pixels below that threshold was statistically different between positive and negative margins. The optimal threshold was determined by repeating the Wilcoxon tests across the full range of threshold values, the results of which showed that 6 μM-cm for β-carotene:< μs′> showed the greatest degree of association with pathology (p < 0.002). A similar process was applied to total hemoglobin:< μs′>, such that the percentage of pixels below a threshold value of 8 μM-cm resulted in the statistically most significant differences between positive and negative margins for that parameter (p < 0.01). contains boxplots that graphically demonstrate these differences.
Using the feature extraction algorithm described by Wilke et al. [38
] and summarized below, the program then determines a margin-level diagnosis for the imaged tissue. A tree-based approach was taken to build the two-parameter model, such that a margin was classified as positive if the percentage of image pixels for the β-carotene:< μs
′> OR total hemoglobin:< μs
′> parameters were above their respective thresholds; otherwise it was classified as negative. These percentages were each varied across the complete set of different threshold values from 0–100%, and the sensitivity and specificity was then calculated against margin-level histopathology. The optimal pair of threshold values was determined by a receiver operator characteristic analysis and the Youden index, in order to maximize the sensitivity and specificity in an additive manner. The average optimal pair of threshold values in the final cross-validated model for β-carotene:< μs
′> and total hemoglobin:< μs
′> were 98 ± 0.19 % and 72 ± 1.0 %, respectively. The percentage of pixels below 6 μM-cm for β-carotene:< μs
′> and below 8 μM-cm for total hemoglobin:< μs
′> is displayed within each respective histogram in and , and the overall margin diagnosis is displayed using this algorithm. With a leave-one-out cross validation technique (an alternative to prospective testing in the absence of a large sample size), an unbiased estimate of 80% sensitivity and 67% specificity was achieved. Of the 55 margins, 41 were correctly identified by the algorithm.
contains a summary of the overall classification performance of the device, as well as the performance of the device in detecting disease as a function of depth from the surface. For this analysis, the positive margins were subdivided into four categories: 1) truly positive margins, with cancer extending to the surface, 2) close margins, with cancer within 1 mm of the surface, 3) close margins, with cancer between 1–2 mm of the margin surface, and 4) margins which were diagnosed as containing cancer within 2 mm of the surface, but in which the exact distance was not specified and is therefore unknown (could be in either of categories 1–3).
Table 3 Performance of decision tree predictive model on all margins, as well as on positive or close margins only, stratified by depth of disease from the margin surface. A category of ‘Unknown’ is given to close margins in which disease depth (more ...)
The results indicate that the device was more accurate in detecting cancer at the surface than in detecting cancer within 1 mm of the margin, which is not unexpected given the average sensing depth noted previously. Interestingly, however, the accuracy of detecting close margins between 1–2 mm of the margin surface exceeded that of close margins within 1 mm of the surface. It should be noted that the majority of the positive margins in this study were truly positive (17), and there were only 7 margins close within 1 mm, and 6 margins close between 1–2 mm. Because the disease variant was not considered in this depth-wise analysis, it is possible that the results were biased by an uneven distribution of disease variants across the depth categories. Specifically, of the 7 margins close within 1 mm, the 2 misdiagnosed margins contained DCIS and mixed IDC/DCIS, respectively. Conversely, in the close between 1–2 mm category, of the 6 margins, only 1 margin (containing IDC) was misdiagnosed. However, the significance of the cancer type on the difference in depth-wise accuracy is not clear because of the small sample sizes and the fact that the two categories differ by only one missed margin. presents the performance of the device in classifying positive or close margins as a function of disease type found at the margin. Interestingly, the device correctly identified 8 of 9 margins positive for ductal carcinoma in situ (DCIS) which corresponds to a sensitivity of 89%. Detection of DCIS intraoperatively is particularly important, since DCIS presents a challenge for surgeons due to its low mammographic density (making it difficult or impossible to see in specimen radiographs), as well as its indistinct gross characteristics.
Performance of decision tree predictive model on positive or close margins, stratified by disease variant found at the margin
One consideration in any study concerned with examination of excised tissue is potential degradation (or temporal change) of the tissue or its associated optical properties. To address this question, degradation studies have been performed on freshly excised tissues to assess the effect of time after excision on the predictive features used to identify positive tumor margins with this technology. It was found that within a 30-minute window (the time frame over which intra-operative margin assessment would occur), time after excision has comparable effects on the optical contrast between positive and negative tumor margins, as compared to the error in the measurement established by taking multiple measurements of the same tissue by removing and replacing the probe. Specifically, the coefficient of variation due to temporal effects in 13 tumor tissues was 0.07 ± 0.03 and 0.21 ± 0.21 for β-carotene:< μs′> and total hemoglobin:< μs′>, respectively, whereas in 15 benign tissues the coefficients of variation of the same variables were 0.23 ± 0.48 and 0.17 ± 0.21. Conversely, in the reproducibility study, the coefficient of variation in 32 measurements from a 4 benign samples was 0.08 ± 0.05 and 0.11 ± 0.11 for β-carotene:< μs′> and total hemoglobin:< μs′>, respectively.