Measuring Chemotherapy Response in Breast Cancer Using Optical and Ultrasound Spectroscopy

Purpose: This study comprises two subprojects. In subproject one, the study purpose was to evaluate response to neoadjuvant chemotherapy (NAC) using quantitative ultrasound (QUS) and diffuse optical spectroscopy imaging (DOS) in locally advanced breast cancer (LABC) during chemotherapy. In subproject two, DOS-based functional maps were analysed with texture-based image features to predict breast cancer response before the start of NAC. Patients and Measurements: The institution’s ethics review board approved this study. For subproject one, subjects (n=22) gave written consent before participating in the study. Participants underwent non-invasive, DOS and QUS imaging. Data were acquired at weeks 0 (i.e. baseline), 1, 4, 8 and before surgical removal of the tumour (mastectomy and/or lumpectomy); corresponding to chemotherapy schedules. QUS parameters including the midband fit (MBF), 0-MHz intercept (SI), and the spectral slope (SS) were determined from tumour ultrasound data using spectral analysis. In the same patients, DOS was used to measure parameters relating to tumour haemoglobin and tissue composition such as %Water and %Lipids. Discriminant analysis and receiver-operating characteristic (ROC) analyses were used to correlate the measured imaging parameters to Miller-Payne pathological response during treatment. Additionally, multivariate analysis was carried out for pairwise DOS and QUS parameter combinations to determine if an increase in the classification accuracy could be obtained using combination DOS and QUS parametric models. For subproject two, 15 additional patients we recruited after first giving their written informed consent. A pooled analysis was completed for all DOS baseline data (subproject 1 and subproject 2; n=37 patients). LABC patients planned for NAC had functional DOS maps and associated textural features generated. A grey-level co-occurrence matrix (texture) analysis was completed for parameters associated with haemoglobin, tissue composition, and optical properties (deoxy-haemoglobin [Hb], oxy-haemoglobin [HbO2], total haemoglobin [HbT]), %Lipids, %Water, and scattering power [SP], scattering amplitude [SA]) prior to treatment. Textural features included contrast (con), vi correlation (cor), energy (ene), and homogeneity (hom). Patients were classified as ‘responders’ or ‘non-responders’ using Miller-Payne pathological response criteria after treatment completion. In order to test if baseline univariate texture features could predict treatment response, a receiver operating characteristic (ROC) analysis was performed, and the optimal sensitivity, specificity and area under the curve (AUC) was calculated using Youden’s index (Q-point) from the ROC. Multivariate analysis was conducted to test 40 DOS-texture features and all possible bivariate combinations using a naïve Bayes model, and k-nearest neighbour (k-NN) model classifiers were included in the analysis. Using these machine-learning algorithms, the pretreatment DOS-texture parameters underwent dataset training, testing, and validation and ROC analysis were performed to find the maximum sensitivity and specificity of bivariate DOS-texture features. Results: For subproject one, individual DOS and QUS parameters, including the spectral intercept (SI), oxy-haemoglobin (HbO2), and total haemoglobin (HbT) were significant markers for response outcome after one week of treatment (p<0.01). Multivariate (pairwise) combinations increased the sensitivity, specificity and AUC at this time; the SI+HbO2 showed a sensitivity/specificity of 100%, and an AUC of 1.0 after one week of treatment. For subproject two, the results indicated that textural characteristics of pre-treatment DOS parametric maps can differentiate treatment response outcomes. The HbO2-homogeneity resulted in the highest accuracy amongst univariate parameters in predicting response to chemotherapy: sensitivity (%Sn) and specificity (%Sp) = 86.5 and 89.0%, respectively and an accuracy of 87.8%. The highest predictors using multivariate (binary) combination features were the Hb-Contrast + HbO2-Homogeneity which resulted in a %Sn = 78.0, a %Sp = 81.0% and an accuracy of 79.5% using the naïve Bayes model. Conclusion: DOS and QUS demonstrated potential as coincident markers for treatment response and may potentially facilitate response-guided therapies. Also, the results of this study demonstrated that DOS-texture analysis can be used to predict breast cancer response groups prior to starting NAC using baseline DOS measurements.