Publicaciones científicas incluidas en el estudio.
G. F María, S.S Carmen, E.F Maria Victoria, C Isabel.
Previous diffuse optical studies of response to breast cancer neoadjuvant chemotherapy have typi- cally used the contralateral breast for normalization, have not utilized z-score normalization, and/or have correlated temporal changes in measured physiological parameters with response to treatment [248; 70; 59; 154; 236; 61; 220; 109; 250; 55; 223; 8]. These alternative approaches were explored in the current study to determine the optimal normalization method for prediction.
First, z-score normalization was implemented to place all parameters on the same magnitude scale, which mitigates systemic physiological differences among the subject population and ac- counts for the systemic effects of chemotherapy. For comparison, models were created that used fully un-normalized data and Tumor-to-Normal ratio normalization,i.e., an average value of a tu-
mor parameter divided by the average of the same parameter in the normal tissue. As with the z-score normalized data, the logistic regression algorithm was run to produce probabilities of re- sponse. With Tumor-to-Normal ratio normalization, a one-parameter model with Early time-point StO2produced anAU C= 0.80, the BaselineStO2model resulted in anAU C= 0.69, and the two-
parameter model with Baseline time-point StO2 and H2O produced an AU C = 0.67. The AUC
values for the same models but with no normalization were even lower (AU C= 0.75,AU C= 0.64, andAU C= 0.67, respectively). The advantage of un-normalized or Tumor-to-Normal data is that it can be acquired with fewer spatial data points. However, these mean values do not account for the inherent heterogeneity of the normal breast tissue. For example, without some knowledge of the normal tissue heterogeneity, such as a standard deviation across many spatial points, there is little context to indicate whether a given Tumor-to-Normal parameter value represents a significant contrast between the tumor and normal tissue or simply a random variation that could be expected to be found in the normal tissue itself. This limitation, along with the statistical issues inherent in regressing over features with large magnitude variations and features with non-Gaussian distribu- tions, could explain the improved correlation generated by z-score normalized data [206; 136; 141]. Many diffuse optical chemotherapy monitoring studies also utilize temporal normalization, whereby the change in parameters from Baseline are correlated with response to treatment [248;
59; 154; 236; 61; 220; 109; 55]. Z-score regression models with this temporally normalized data were also tested. However, even the most predictive of the models derived in this analysis that used the change in any DOSI physiological parameter between the Baseline and Early time-points only produced anAU C= 0.63. The best temporal change model between the Baseline and Mid- point produced an AU C = 0.74, which, although more predictive than the models at the Early time-point, has the disadvantage of providing predictions approximately two months later in the chemotherapy cycle. The temporal change models ofStO2, in particular, could be limited by the
large inter-subject dispersion of the Baseline oxygen saturation; this large dispersion prevents the change in StO2 between the Baseline and Early time-points alone from accurately reflecting the
oxygenation state of the tumor relative to the normal region. By contrast, theeStO2 model pre-
sented here does not depend on the Baseline StO2 and, as such, is not affected by inter-subject
Baseline variation. In other words, the set of subjects whose tumors exhibit high Early time-point StO2z-scores encompasses both those subjects whose Baseline StO2 was high and subjects whose
BaselineStO2 experienced a large increase after the first dose of therapy. Models that rely on the
change inStO2between the Baseline and Early time-points may not accurately reflect the normoxic
or hyperoxic states of tumors that began with highStO2 and experienced a limited change.
Finally, z-score normalization to the contralateral breast tissue, rather than the normal tissue on the tumor-bearing breast, was tested. This is a common method of normalization in the diffuse optical community due to the fear of tumor tissue signal contamination in the tumor-bearing breast normal tissue caused by the partial volume effect. If the contralateral breast was used for z-score normalization with the primary models discussed here, the one-parameter EarlyStO2model
produced anAU C= 0.67, the one-parameter BaselineStO2model produced andAU C= 0.68, and
the two-parameter model with Baseline StO2 and H2O produced an AU C = 0.64. Thus, z-score
normalization to normal tissue on the tumor-bearing breast appears to enable better predictions of response to therapy than use of the contralateral breast data. This could be a result of systemic differences in oxygen saturation between the ipsilateral and contralateral breast. Thus, comparison of the tumor oxygen saturation to the contralateral oxygen saturation may not be an accurate representation of the relative level of tumor hypoxia. The comparatively better quality of the tumor breast z-score normalized models suggests that measurement of the contralateral breast is less important for early prediction of response to therapy than previously thought, which confirms other work done with the DOSI instrument [163]. If this is true, then a new imaging paradigm could eliminate the need for contralateral measurement and reduce imaging time by half. This advance could enable easier clinical adoption of the DOSI system.