MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Radiomics Prediction of Neoadjuvant Pathological Complete Response Outcome Based On Multi-Institutional Multiparametric MRI for Rectal Cancer

Radiomics Prediction of Neoadjuvant Pathological Complete Response Outcome Based On Multi-Institutional Multiparametric MRI for Rectal Cancer

M Huang1*, P Yang2 , L Xu3 , L Shi4 , X Sun5 , X Zhong6 , N Li7 , F Yin8 , T Niu9 , X Qi10 , (1) Duke Medical School Radiation Oncology, Durham, NC, (2) Zhejiang University, Hangzhou, ,(3) Zhejiang University, Hangzhou, ,(4) Department of Radiation Oncology, Sir Run Run Shaw Hospital,Zhejiang Univ.(5) Department of Radiation Oncology, Sir Run Run Shaw Hospital, Zhejiang Univ., Hangzhou, ,(6) UCLA, Los Angeles, CA, (7) Cancer hospital Chinese Academy of medical Sciences, Beijing, ,(8) Duke University Medical Center, Durham, NC, (9) Zhejiang University, Hangzhou, ,(10) UCLA School of Medicine, Los Angeles, CA

Presentations

(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM

Room: 225BCD

Purpose: To evaluate the radiomics model for predicting pathological-complete-response (pCR) for locally advanced rectal cancer(LARC) cross multi-institutional multiparametric MRI datasets.

Methods: Two independent LARC cohorts from two different institutions were included. Cohort1 includes 44 patients (pCR: n=9) receiving neoadjuvant chemo-radiation therapy (CRT). A separate cohort2 includes 38 patients (pCR: n=8). All patients underwent T2 with 3.0 Telsa MRI (GE scanner), and Diffusion-Weighted-MRI (DWI) before the CRT. The DWI-MRI images acquired with the B-value of 800 (cohort1) and 1000 (cohort2). A radiomics prediction model for predicting pCR was then built based on the T2 images from corhort1. A total of 547 radiomics features (7 histogram features, 22 GLCM features, 13 GLRLM features, 13 GLSZM features, 5 NGTDM features, 480 wavelet-based radiomic features and 7 shape features) were extracted for each patient. The univariate Mann-Whitney test and the LASSO regression with L-1 penalty were used for the feature selection and a multi-variable logistic regression model was applied toward final selected features. A 4-fold validation technique was utilized in determining the best parameter in LASSO. Cohort2 were used for external validation. The predictive performance was calculated using the area under the ROC curve(AUC).

Results: When Cohort1 was used for training model, the developed radiomics model comprised 3 selected features and showed a good performance (AUC =0.908, 0.820-0.996: 95% confidence interval [CI]). The validation AUC was 0.650 (95% CI: 0.438-0.862) when cohort2 used as validation. When cohort2 was used for training model, the developed radiomics model comprised 2 selected features and showed an AUC of 0.792 (0.575-1, 95% CI) and external validation AUC of 0.635 (95% CI: 0.424-0.846) from cohort1.

Conclusion: Applying radiomics prediction model cross multi-institutions is feasible, giving similar image quality and patient cohort population. Bias exists between datasets from different institutions in radiomics study requiring more careful design and exam.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links