MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Combining Delta-Radiomics and Clinical Biomarkers Based On KNN-PCA Classification to Improve Treatment Outcome Prediction for Pancreatic Cancer

H Nasief1*, W Hall2, C Zheng3, S Tsai4, B Erickson5, X Li6, (1) Medical College of Wisconsin, Milwaukee, WI, (2) Medical College of Wisconsin, Milwaukee, WI, (3) University of Wisconsin Milwaukee, Milwaukee, WI, (4) Medical College of Wisconsin, Milwaukee, WI, (5) Medical College of Wisconsin, Milwaukee, WI, (6) Medical College of Wisconsin, Milwaukee, WI

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: investigate if treatment outcome prediction for pancreatic adenocarcinoma (PDAC) can be improved using K-nearest neighbor and principal component analysis (KNN-PCA) classifier to identify delta radiomic features (DRFs) to be combined with clinical biomarkers, CA19-9 and pancreatic intraepithelial neoplasia (PanIN) grade

Methods: non-contrast CTs acquired during routine CT-guided pre-operative chemoradiation therapy (CRT) for 26 PDAC patients, along with their CA19-9, PanIN biopsy-grade, pathological results, and follow up data, were analyzed. Radiomic features were extracted from pancreatic head on each daily CT and were used to calculate DRF between different days. Patients were divided into two groups based on their pathological responses. KNN-PCA based classifier was built to identify DRFs that can be combined with CA19-9 and PanIN grade to result in improved treatment outcome prediction. Specifically, PCA was incorporated to identify the DRFs with highest explained variance to reduce model complexity and avoid overfitting. A Cox analysis was performed to correlate PanIN, CA19-9 and DRFs with survival

Results: highest explained variance of DRFs with PCA was 0.96, an increase from 0.82 without using PCA. Incorporating CA19-9 and PanIN with the obtained DRFs increased AUC of the KNN-PCA classifier from 0.57 using single biomarker to 0.98 with 0.9 accuracy, indicating improved predicting power. The pathology response can be predicted by the 2nd week during CRT using combined DRF-CA199-PanIN biomarkers, compared to 4th week if using CA19-9 alone. Such an earlier prediction would allow enough time to adapt treatment if necessary. Cox multivariate-analysis showed that treatment related decrease in CA19-9 levels (p=0.031), low PanIN grade (p=0.03) and DRFs (p=0.001) were independent predictors of survival. The hazard ratio was reduced from 0.73, p=0.032 using CA199 alone to 0.43, p=0.04 using DRF-CA199-PanIN combination

Conclusion: based classifier can identify appropriate DRF-PanIN-CA199 combinations to improve the predictions of pathology response and survival for CRT of PDAC

Keywords

Feature Extraction, Image-guided Therapy, Classifier Design

Taxonomy

IM- CT: Biomarkers

Contact Email