Program Information
Pathological Response Prediction by Radiomic Data From Primary and Lymph Nodes in NSCLC
T Coroller*, V Agrawal , V Narayan , S Lee , R Mak , H Aerts , Dana-Farber Cancer Institute, Brigham Women's hospital, Harvard Medical School, Boston, MA
Presentations
TU-D-207B-6 (Tuesday, August 2, 2016) 11:00 AM - 12:15 PM Room: 207B
Purpose: In advanced non-small cell lung cancer (NSCLC) patient, metastasis can spread from the primary tumor to the lymph nodes and hence could have a distinct phenotype compared to unaffected nodes. In this study we investigated the complementary information of radiomics extracted from lymph nodes and the primary tumor in order to predict pathological response at time of surgery after chemoradiation in patients with stage II-III NSCLC.
Methods: 86 NSCLC patients with primary tumor and involved lymph nodes (LN) were included in this study. Twenty radiomic features were selected based on stability and variance. Predictive power was evaluated using AUC and false discovery rate (FDR) corrected p-values. Conventional imaging features (total tumor / LN volume and axial tumor diameter) and clinical characteristics were included for comparison. Classification power was investigated using random forest. Performances were assessed using cross validation (1000 iterations, 70% training / 30% validation).
Results: Three radiomic features were predictive for pathologic (GRD) gross residual disease (AUC range 0.69-0.75, p<0.05) and two pathologic (pCR) complete response (AUC range 0.65-0.68, p<0.05). No conventional imaging features were predictive of either outcome (range AUC = 0.51 to 0.61, p>0.05). Patients with pCR were likely to have more homogeneous LN (large area emphasis, AUC=0.74, p<0.05) when in the other hand patients with GRD present rounder primary tumor shape (spherical disproportionality, AUC= 0.68 p<0.05). Cross validation AUC values shown that the combined radiomics-clinical features set outperformed for each outcomes (median AUC = 0.68 and 0.74 respectively for pCR and GRD).
Conclusion: We demonstrate that LN phenotypic information, ascertained from radiomic features, is complementary to imaging features obtained from the primary tumor. These features are strongly associated with response to chemoradiation as determined by pathologic response and provide greater predictive performance than clinical characteristics and conventional assessment of tumor burden by diameter or volume.
Funding Support, Disclosures, and Conflict of Interest: R.M. have consulting interest with Amgen.
Contact Email: