Rahul Mehta, Presenter: Nothing to Disclose

Nishant Kumar MD, Abstract Co-Author: Nothing to Disclose

Hui Lu, Abstract Co-Author: Nothing to Disclose

Aladin Mariano MD, Abstract Co-Author: Nothing to Disclose

Grace Knuttinen, Abstract Co-Author: Nothing to Disclose

Thomas M. Anderson MD, Abstract Co-Author: Nothing to Disclose

Yang Lu MD, PhD, Abstract Co-Author: Nothing to Disclose

To develop a prediction algorithm capable of determining the effectiveness of Y90-SIRT treatment in patients with primary and metastatic liver cancers through the use of imaging biomarkers extracted from PET/CT scans.

We designed a strategy of associating changes in imaging features of tumors after treatment through the use of pattern recognition and machine learning. We modified a fuzzy clustering algorithm to automatically detect and segment liver tumors to calculate individual tumor features such as SUV, morphology, texture, and gray-level statistics. Next, we built a support vector machine (SVM) and a Bayesian model to identify critical imaging markers relevant to improvement in Y90-SIRT therapy. Finally, we evaluated the prognostic significance of the model on patients to determine whether Y90-SIRT is an effective treatment in the current state of cancer. The strategy was applied on a set of 15 pretherapy FDG PET/CT scans in patients with Cholangiocarcinoma (n=6), or liver metastases from colon cancer (n=8) and ovarian cancer (n=1). Each patient had at least a 6 month follow-up with PET/CT. Additionally, some had contrast CT or MRI studies. Y90-SIRT therapy responses were analyzed with PET/CT based PERCIST criteria.

The model was able to predict the effectiveness of treatment with an accuracy of 85%-95% in determining if a patient would improve based on PET/CT scan. The sensitivity was found to be 90%, while the specificity was 100%. We found the Bayesian model to have a higher accuracy rate, most likely because our cohort of data is relatively small. Furthermore, we found tumor volume, number of curves of a tumor, and edge shape had greatest prognostic significance.

The model is self-learning. As further data is accumulated, the prediction accuracy will improve. Furthermore, we can add additional imaging biomarkers to increase the sensitivity rate. The ability to predict the outcome of a treatment based on imaging biomarkers may reduce or prevent unnecessary, expensive, and invasive procedures, along with the potential to provide personalized treatments.

The computer aided pre therapy PET/CT based prediction algorithm can predict responsiveness of liver directed Y90-SIRT therapy, thus avoiding ineffective treatment and unnecessary costly procedures.

Mehta, R, Kumar, N, Lu, H, Mariano, A, Knuttinen, G, Anderson, T, Lu, Y, Computer Aided Response Prediction Based on Pre-therapy FDG PET/CT Imaging Biomarkers of Y90-SIRT Therapy in Patients with Primary and Metastatic Liver Cancers.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14008235.html