Click here to


Are you sure ?

Yes, do it No, cancel

Identifying the Prognostic Value of PET and Clinical Information for Triple-Negative Breast Cancer Patients Treated with Radiotherapy Using Deep Learning Approach

Y Yue*, S Bose , X Cui , M Burnison , H Sandler , B Fraass , Cedars-Sinai Medical Center, Los Angeles, CA


(Sunday, 7/29/2018) 5:05 PM - 6:00 PM

Room: Karl Dean Ballroom C

Purpose: Recent studies show 18-F FDG PET for triple-negative breast cancer (TNBC) patients has significant tumor uptake compared to other breast cancers. This study uses a deep learning approach to identify the prognostic value of the combination of image and clinicopathological data for improving clinical outcomes of TNBC patients treated with radiotherapy.

Methods: With IRB approval, 43 consecutive TNBC breast cancer patients treated with radiotherapy were identified. Tissue samples were examined by immunohistochemistry, including P53, Ki-67, CK5/6 and EGFR, and clinical variables were collected, including age, tumor size, grade, pathologic tumor status. All patients had pre-treatment PET/CT scans, and FDG uptake was measured by standard uptake value (SUV). Deep learning convolutional neural networks (DL-CNNs) were used to extract imaging features with 2 convolution layers, and then were connected with a separate regression model to include clinicopathological information. The final layer was constructed to account for image-clinical features for prediction of survival risk of patients. Further, the prognostic values of individual variables were examined by disease-free survival (DFS), and the results of neural networks were compared with multivariate Cox analysis.

Results: DFS was significantly associated with tumor size(p=0.03), lymph node status(p=0.04), T stage(p=0.02), EGFR(p=0.002), and mean uptake(SUVmean, p=0.016). The survival regression was validated using 5-fold cross-validation. The proposed imaging-clinical learning approach achieved better performance for prediction of DFS(C-Index=0.745) than multivariate Cox analysis(C-index=0.692). As a result, patients were stratified into two groups based on the risk score with log-rank p=0.016: a low risk group(n=24) with a longer mean DFS(38.1 months), and a high risk group(n=19) with a shorter mean DFS(22.2 months).

Conclusion: The integration of learning-based deep features and clinicopathological information provides a feasible approach for predicting clinical outcomes for TNBC patients. The proposed method can potentially be used to stratify patient risk and select appropriate clinical strategies before treatment.


PET, Quantitative Imaging, Breast


TH- response assessment : PET imaging-based

Contact Email