Machine Learning Strategies to Analyze Quantitative Ultrasound Multi-Parametric Images for Prediction of Therapy Response in Breast Cancer Patients

In this thesis project, two novel machine learning strategies were investigated to predict tumor response to neoadjuvant chemotherapy (NAC) at pre-treatment using quantitative ultrasound (QUS) multi-parametric images. The ultrasound data for analytical development and evaluation of the methodologies investigated in this project were acquired from 181 patients diagnosed with locally advanced breast cancer (LABC) and planned for NAC followed by surgery. The QUS multi-parametric images were generated using spectral analyses on the raw ultrasound radiofrequency (RF) data acquired before starting the NAC. In the first machine learning approach investigated in this project, distinct intra-tumor regions were identified within the parametric maps using a hidden Markov random field (HMRF) and its expectation-maximization (EM) algorithm. Several hand-crafted features characterizing the tumor, intra-tumor regions, and the tumor margin were extracted from different parametric images. A multi-step feature selection procedure was applied to construct a QUS biomarker for response prediction. Evaluation results on an independent test set indicated that the developed biomarker using the characteristics of intra-tumor regions and tumor margin in conjunction with a decision tree model with adaptive boosting (AdaBoost) as the classifier could predict the treatment response of patients at pre-treatment with an accuracy of 85.4% and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.89. In the second machine learning approach investigated in this project, two deep convolutional neural network (DCNN) architectures including the residual network (ResNet) and residual attention network (RAN) were explored for extracting optimal feature maps from the parametric images, with a fully connected network for response prediction. Results demonstrated that the developed model with the RAN architecture to extract feature maps from the expanded parametric images of the tumor core and margin had a superior performance with an accuracy of 0.88 and an AUC of 0.86 on the independent test set. Also, survival analysis demonstrated a statistically significant difference between survival curves of the two response cohorts identified at pre-treatment based on both the conventional machine learning method and the deep learning model. Obtained results in this study demonstrated a great promise of QUS multi-parametric imaging integrated with both unsupervised learning methods in identifying distinct breast cancer intra-tumor regions and traditional classification techniques, as well as deep convolutional neural networks in predicting tumor response to NAC prior to start of treatment.