From December 2007 to September 2008, 54 patients were enrolled in this study. Optical spectral imaging data from 6 patients were not utilized in this analysis as the pathological outcomes could not be accurately co-registered with the specified margins. From the remaining 48 patients, 55 margins were evaluated with the optical spectral imaging device. contains a summary of the patient and margin characteristics for the participants in this study. Twenty one patients had negative margins, and 15 had a positive margin on the main specimen which was analyzed with the optical probe with additional margin specimens taken by the surgeons at the original surgery to obtain negative margins. Of the 48 patients, 12 had to have a re-excision lumpectomy.
Characteristics of the study population.
The majority of patients (73%) enrolled had invasive and/or in-situ ductal carcinomas in the lumpectomy specimen. The remaining patients had a variety of histologic subtypes including lobular (invasive and in-situ), tubular, and papillary carcinoma. Three of the patients in this study had a complete pathologic response to their pre-operative chemotherapy. For each patient enrolled in this study, only 1–2 margins were assessed with the optical spectral imaging device. As such, the re-excision rate and device outputs could not be compared to determine what outcome could have occurred had the device results been revealed to the surgeons.
contains representative parameter maps of β-carotene concentration to wavelength-averaged reduced scattering coefficient (β-carotene:scattering) (A,C,E) and corresponding histograms which graphically represent the distribution of values within each image (B,D,F) for a margin negative for residual disease (A,B), a margin positive for ductal carcinoma in-situ (DCIS) (C,D) and a margin positive for invasive ductal carcinoma (IDC) (E,F). 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 is a measure of the amount of light elastically scattered in the tissue, with higher scattering coefficients associated with more dense 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).22
Since malignant tissues are expected to have less fat (due to displacement of adipocytes by carcinoma cells) and higher scattering (due to increased cell density and changes in nuclear morphology), β-carotene:scattering is expected to be decreased in cancer tissue relative to normal breast tissue. Color maps in the images of are set such that lower values of β-carotene:scattering appear red, whereas higher values appear blue. As seen in the images, the negative margin () is characterized by a higher proportion of blue pixels, whereas the positive margins () are characterized by increased proportions of red pixels. Site-level histology of these margins indicated that generally, the blue areas were indicative of cancer-free regions of the margin, whereas the orange-red areas indicated regions of the margin containing residual disease.
Figure 2 Maps of β-carotene:scattering coefficient for A) negative margin, C) margin positive for DCIS, E) margin positive for IDC. Site-level pathology for the margins indicated that the blue areas generally corresponded to cancer-free areas, whereas (more ...)
In order to build predictors for margin-level assessment from the parameter maps, a threshold value for pixel intensity was determined (e.g., 6 μM-cm for β-carotene:scattering) and the percentage of pixels below that threshold was computed (as illustrated in ). 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:scattering showed the greatest degree of association with pathology (p < 0.002, ). A similar process was applied to total hemoglobin:scattering, 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 ).
Figure 3 Boxplots of A) percentage of pixels < 6 μM-cm β-carotene:scattering, and B) percentage of pixels < 8 μM-cm total hemoglobin:scattering, stratified by margin status (negative: blue, positive/close: red). P-values (more ...)
Next a multivariate predictive model was developed for classifying a margin as positive or negative based on the predictors shown in . A tree-based approach was taken to build the two-parameter model, such that a margin was classified as positive if the percentage for the β-carotene:scattering OR total hemoglobin:scattering parameters were above their respective thresholds; otherwise it was classified as negative. The percentages shown on the y-axes in were each varied across the complete set of different threshold values (for example, 98% in and 72% in ), and the sensitivity and specificity was then calculated against margin assessment by pathology. The optimal pair of threshold values was determined by a receiver operator characteristic analysis (ROC, ) and the Youden index, in order to maximize the sensitivity and specificity in an additive manner. Then, a leave-one-out cross validation scheme was used to obtain an unbiased estimate of the operating characteristics of the predictive model using the same guiding principles as above, and resulted in a sensitivity and specificity of 79% and 67%, for the two parameter decision tree. 23
The percentage pixel thresholds in the final model for β-carotene: scattering and total hemoglobin: scattering was 98 ± 0.19 % and 72 ± 1.0 %, respectively.
ROC of the two predictors from .
contains the prediction accuracy resulting from the decision-making model described above. Of the 34 path-confirmed positive/close margins in the dataset, the predictive model correctly identified 27 of them as positive, yielding a sensitivity of 79.4%. In addition, of the 21 path-confirmed negative margins, the predictive model correctly identified 14 of them as negative, yielding a specificity of 66.7%. The performance of the model in predicting path-positive/close margins is also shown as a function of type of cancer found at the margin. For margins that were positive for IDC only, the predictive model correctly identified 11 out of 14, or 78.6% of positive margins. For margins that were positive for DCIS only, the predictive model correctly identified 8 out of 9, or 88.9% of positive margins. In the current dataset, in 6 of the positive margins the type of cancer tissue present at the margin was not specified, whereas a further 5 margins were positive for less common malignancies (lobular, mixed DCIS/IDC, and tubular). Of these “other” positive margins, the predictive model correctly identified 8 of 11, or 72.7% of positive margins. The margins were equally weighted with respect to the number of close margins (n = 17) and the number of positive margins (n = 17). The performance of the predictive model was also not biased significantly toward either positive or close margins, with positive margins being correctly predicted only slightly more frequently (14/17 or 82.4% sensitivity) than close margins (13/17 or 76.5% sensitivity).
Prediction accuracy of cross-validated algorithm for margin classification on 55 margins from 48 patients. Sensitivity is also given for path positive margins separated by cancer subtype, as well as by positive versus close margin status.
Although only 7 of 34 positive margins (from 6 patients) were misclassified as falsely negative, it is important to understand why these margins may have been misclassified. Within this patient group were 2 patients who received neoadjuvant therapy; one with endocrine therapy and the other with chemotherapy. is imaged from the margin (posterior margin pathologically positive for IDC) of the patient who underwent neoadjuvant endocrine therapy. This patient had a significant decrease in proliferation rate between her pre- and post-therapy biopsies which may have decreased the scattering coefficient, which in turn could have resulted in a more flat histogram, which is typical of negative margins. In the second patient who had received neoadjuvant therapy, the device incorrectly called the posterior margin diagnosed with IDC at 2 mm as negative. The device likely called this margin negative because the residual disease was at a depth that was just at the 2 mm sensing depth of the device.
Figure 5 Maps of β-carotene:scattering coefficient (A, C) for margins positive for IDC but falsely classified as negative by the predictive model from 2 different patients. Corresponding frequency histograms are given to the right of each map (B and D, (more ...)
Of the remaining 5 patients, is imaged from a patient with IDC just beneath her nipple where there is greater ductal tissue and less fat. The histogram in this particular case looks like that of a positive margin but was just above the percentage pixel cutoff for margin positivity. This is an example of a near-miss. Another of the false negative cases had DCIS which was misclassified as negative by the optical spectral imaging device. The final pathology showed close margins anterior-inferiorly and anterior-laterally with additional shaved anterior, lateral and inferior margins negative for DCIS. This tumor specimen was large with a volume measuring greater than 1200 cm3. In this case, although the majority of the margin is reflective of normal tissue, there are some relatively small areas of very low β-carotene:scattering values, which are suggestive of the presence of residual disease. However, in the current method of automated image analysis, the contribution of these pixels to the overall pixel distribution was small, due to the very large size of these margins. This resulted in misclassification of the margin, since the percentage of pixels below the threshold was small. This is a weakness of the current paradigm for automated image analysis, in that error may be introduced when the area of suspicious pixels is small compared to the overall margin size.