Room: Track 2
Purpose: we showed that CT delta-radiomic features (DRF) could predict treatment response in mid-treatment during preoperative chemoradiation therapy (CRT) of pancreatic cancer. This study aims to investigate possibility of earlier prediction by normalizing DRF (NDRF) with respect to those from healthy pancreas.
Methods: analyses were performed with dataset consisting of 1) non-contrast CTs acquired in 28 fractions during routine CT-guided preoperative CRT of 40 patients with pancreatic tumors along with their pathological response data, and 2) daily non-contrast CTs acquired during radiotherapy of 20 patients with other abdominal tumors but with healthy pancreas. DRFs were extracted from pancreatic heads and T-test and coefficient of variance (COV) were used to assess the differences in DRFs between normal-(NP) and diseased-(DP) pancreas. NDRFs for DRFs with significant t-test p-value and COV>5% were calculated. T-test and linear-mixed-effect model were used to assess NDRF correlation to response. Bayesian classifier with leave-one-out cross-validation was used to identify NDRFs with improved response predictions. Performance was judged using the AUC of the ROC curve.
Results: expected, DRFs of NP had much narrow variations compared to DRFs for DP. A total of 23 DRFs were found to be significantly different between their NP and DP values, of which 48% had COV> 5% and were selected to calculate NDRFs. Eight NDRFs passed t-test and linear mixed effect (p<0.05) and showed significant changes as early as at fractions 5-21 during CRT between good and bad response groups. This is an improvement from the fractions 14-18 when using the DRFs. Bayesian classifier AUC increased from 0.93 with using 3 DRFs (kurtosis, contrast, coarseness) combination to 0.97 for using the corresponding NDRFs.
Conclusion: delta radiomics to the corresponding healthy structure can reduce variability and improve performance of DRF. NDRFs can predict treatment response two weeks earlier than DRFs during CRT for pancreatic cancer.