MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Longitudinal MRI Based Radiomics: Classifying Rectal Cancer Response to Radiation Therapy

A Chen1*, Y Gao2, Y Yang2, P Hu2, (1) Jericho High School,(2) University of California, Los Angeles, Los Angeles, CA

Presentations

(Tuesday, 7/14/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: Our study was designed to: 1. evaluate the feasibility of using longitudinal MRI-based radiomics to predict rectal cancer response to neoadjuvant radiation therapy (RT); and 2. identify the most predictive features extracted from both diffusion weighted MRI (DWI) and 3D anatomical MRI acquired every 6-10 treatment fractions, and the optimal imaging time during treatment.

Methods: Ten rectal cancer patients treated with standard fractionated RT were recruited. DWI and high-resolution 3D TrueFISP MRI were acquired five times throughout treatment using a 0.35T MRI-guided radiotherapy system: fractions (fx)1-10; fx9-15; fx14-21; fx19-29; fx25-39. A 4-point scale pathological treatment response score was obtained from the post-radiotherapy resection as a clinical endpoint. Patients were divided into 2 groups based on the score (group1: 0-1; group2: 2-4).

Mean tumor apparent diffusion coefficient (ADC) and 2,760 delta and static radiomics features were extracted from the longitudinal MRIs (ADC maps and 3D TrueFISP MRI) within GTV contour transferred from planning CT. The area under the receiver operating characteristic curve (AUC) was calculated to rank the features. Logistic regression and support vector machine models were built to evaluate each feature’s prediction performance through the leave-one-out approach. High-accuracy features were identified by calculating frequencies of various feature categories, times, and delta features.

Results: TrueFISP-based features yielded optimal AUCs (AUC=1) and testing and training accuracies, indicating superior estimating abilities compared to mean ADC (max-AUC=0.63) and ADC-based features (max-AUC=0.88). Delta, first-order and gray level co-occurrence matrix features acquired at times 1 (between fx1-10) and 3 (between fx14-21) were most prevalent among high-accuracy features, suggesting that these features and acquisition times provide the best treatment response estimation.

Conclusions: Based on our small cohort of rectal cancer patients, radiomics features from longitudinal anatomical MRI outperformed DWI. Two optimal MRI acquisition time points were identified as the beginning and middle of the treatment course.

Funding Support, Disclosures, and Conflict of Interest: Dr. Hu and Dr. Yang have served as short term consultant at Viewray Inc.

Keywords

Low-field MRI, Radiation Therapy

Taxonomy

TH- Response Assessment: Imaging-based: MRI

Contact Email