On the assessment of local tumor response to neoadjuvant chemotherapy

If a malignant tumor is detected, the patient is often given neoadjuvant chemotherapy (NAC) before surgery. The effectiveness of NAC is estimated using tumor size changes. However, such a method is not always reliable, so other techniques are being investigated. The paper shows how to assess the response of different tumor areas to therapy based on the analysis of scattered ultrasound signals. Understanding the local response of the tumor can help to evaluate the effectiveness of therapy. The study used a set of raw ultrasound data from 48 tumors undergoing NAC. Ultrasound scanner was used to collect RF data before the start of NAC and after each drug administration. After therapy, tumors were resected and histopathologically evaluated. The percentage of residual malignant cells (RMC) in each lesion was estimated and used for assessing the NAC effectiveness. The set of tumors was divided into a training and a test sets. In the training set each tumor Region of Interest (ROI) was divided into small square pieces – patches, and labelled as responding or non-responding basing on RMC of the tumor. Then 357 statistical and texture features were estimated from each patch. The support vector machines binary classifier (SVM) was trained on data collected after the 3 ^rd drug administration. The efficiency of the classifier using a different number of features in the range from 2 to 100 has been tested. The classifier was then used to determine the probability of high RMC of tumor patches from the test set. In this way, parametric images of tumors were obtained, showing the spatial distribution of the probability of a given area being unresponsive to treatment. The ’patch approach’ allows the use of a very large set of predictors without the risk of overfitting – although the number of tumors in the set was not very large, the number of patches obtained from them was. Spatial distribution of non-responsiveness probability can be a base for detection of non-responding tumors. A simple predictor based on 70 ^th percentile of probability of non-responsiveness after the 3 ^rd dose resulted in a classification with an area under the ROC curve of 0.92, indicating potential for identifying non-responding tumors.