Click here to

Session: Multi-Disciplinary: Data Science/Radiomics [Return to Session]

CT-Based Deep Learning Radiomics for Predicting Chemoradiation Treatment Response in Locally Advanced Rectal Cancer

J Fu1*, Z Wang1, K Singhrao1, J Lewis2, X Qi1, (1) Department of Radiation Oncology, UCLA, Los Angeles, CA, (2) Department of Radiation Oncology, Cedars-Sinai Medical Center, Los Angeles, CA


MO-IePD-TRACK 4-6 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: Many studies have used the conventional handcrafted radiomic features to predict neoadjuvant chemoradiation treatment (nCRT) response in patients with locally advanced rectal cancer (LARC). The goal of this study is to investigate deep learning (DL)-based features for nCRT response prediction based on the planning CT.

Methods: 86 LARC patients treated at a single institution were retrospectively enrolled. All patients received preoperative nCRT, with a prescription dose of 50.4Gy in 28 factions, before total mesorectal excision. Patients were split into a responder group (n=45) and a non-responder group (n=41) based on pathological tumor regression grades. Each patient had a planning CT along with the manually delineated gross tumor volume. Handcrafted features and DL-based features were extracted from the planning CT using PyRadiomics and a pre-trained convolutional neural network (VGG19), respectively. Two logistic regression models with LASSO regularization were constructed using handcrafted and DL-based features, respectively. The model performance was evaluated with the stratified 5-fold cross-validation using receiver operating characteristic (ROC) curves.

Results: The model constructed using handcrafted features achieved an area under the ROC curve (AUC) of 0.63±0.13, while the one constructed using DL-based features yielded an AUC of 0.76±0.04.

Conclusion: Deep learning-based features extracted from planning CT achieved more accurate nCRT response prediction for LARC patients, as compared to handcrafted features. Deep learning-based features selected by the LASSO regularization are potential biomarkers for outcome prediction and personalized therapy.



    CT, Texture Analysis, Radiation Therapy


    TH- Response Assessment: Radiomics/texture/feature-based response assessment

    Contact Email