MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Rectal Cancer Survival Outcome Prediction Based On Radiomics Features for Different Imaging Modalities

M Huang*, H Zhong , P Boimel , J Janopaul-Naylor , K Men , H Geng , C Cheng , P Gabriel , Y Fan , L Ungar , M Rosen , E Ben-Josef , Y Xiao , University of Pennsylvania, Philadelphia, PA

Presentations

(Monday, 7/30/2018) 4:30 PM - 6:00 PM

Room: Davidson Ballroom B

Purpose: To implement Machine learning Random Survival Forest(RSF) prediction method, for overall survival (OS) prediction based on radiomics features from CT, MRI, T1_post, and T2 imaging modalities, for patients with locally advanced rectal cancer treated with neoadjuvant chemo-radiotherapy followed by surgery. Results were compared to penalized COX regression.

Methods: 225 patients were included in the study. The patient characteristics and 170 radiomics features were collected from 93 CT, 82 MRI T1, and 84 MRI T2 images, respectively. Extracted radiomics features included 3 geometry features, 12 histogram features, 100 GLCM texture features, and 55 GLRLM texture features for each imaging modality. These features were clustered to 20 features for RSF and COXPH prediction modeling. The matching radiomics prediction model for OS were built for each imaging modality. RSF, an ensemble tree method for analysis of right-censored survival data, combines both machine learning and providing statistical event-specific inference. RSF analysis was performed and benchmarked to KM curve over 5-year span. All prediction models were five-fold cross-validated.

Results: Performance indices (AUC) for RSF were 0.56 for CT radiomics features, 0.62 for T1 radiomics measures, 0.66 for T2 images, and 0.68 with combined MRI. Using COXPH with L1 regulation, the AUC was 0.58 for CT, 0.53 for T1 and 0.57 for T2. The RSF model also calculates survival prediction over time, 97.4%, 90.7%, 86.4%, 85.2%, 77.1% at 1-, 2-, 3-, 4-,5-year, aligned with KM estimates for both T1 and T2 images.

Conclusion: It is feasible to apply machine learning RSF for radiomics outcome predictive modeling, which has the potential to out perform conventional COXPH. Performance with different image modalities varied. In MR T1,T2 images, the radiomics features give better predictive power over CT. More samples, strict tumor contour guideline and better image quality will further improve the prediction power.

Funding Support, Disclosures, and Conflict of Interest: This project was supported by U24CA180803 (IROC) from the National Cancer Institute (NCI).

Keywords

Not Applicable / None Entered.

Taxonomy

IM- MRI : General (Most aspects)

Contact Email