Click here to

Session: MRI in Radiation Therapy [Return to Session]

A Hybrid Model Derived From Multi-Parametric MRIs for Predicting Neoadjuvant Chemoradiation Response in Locally Advanced Rectal Cancer

Hao CHEN1*, Xing Li2, X. Sharon Qi3, (1) Xi'an University of Posts and telecommunications, Xi'an, shaanxi, CN, (2) Xi'an University of Posts and Telecommunications, Xi'an, Shaanxi, CN, (3) UCLA School of Medicine, Los Angeles, CA


MO-E115-IePD-F6-6 (Monday, 7/11/2022) 1:15 PM - 1:45 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Accurate prediction of treatment response to neoadjuvant chemoradiotherapy (nCRT) facilitates subsequent treatment management for patients with locally advanced rectal cancer (LARC). We designed a 3D-CNN to extract high-level imaging features, and combined with radiomic features to predict post-nCRT treatment response.

Methods: A 3D-CNN, consisting of two convolutional layers and a pooling layer parallelized by global average pooling (GAP), was designed to extract high-level imaging features from gross tumor volume (GTV). Pre-treatment ADC images and T2 images were acquired for 43 LARC patients. The patients were stratified into good responder (GR, n=22) and poor responder (non-GR, n=21) according to post-operative pathology after completion of nCRT (4-8 weeks). A total of 1160 features, inclcuding 200 radiomic features (PyRadomics) and 960 3D-CNN features, were extracted from the pre-treatment ADC/T2 images. A feature vector consisting of 128 hybird features were formed by the random forest (RF) recursive feature elimination (RFE) algorithm. The eXtreme Gradient Boosting (Xgboost) classifier was used to classify the GRs and non-GRs, the model performance was evaluated using the area under the receiver operator characteristic (ROC) curve (AUC).

Results: The classifiers were built upon radiomic features, 3D-CNN features and fused features for predicting GR and non-GR with repeated 20 times using 5-fold cross-validation. The radiomic classifier and DL-based classifier achieved the mean AUCs (standard deviation) of 0.692±0.07 and 0.81±0.02, respectively, while the hybrid model achieved the mean AUC of 0.854±0.01. The corrected paired delong-test of the hybrid model showed P-value < 0.05 compared to radiomic and DL-based model alone.

Conclusion: We trained a 3D-CNN to capture high-level imaging features using MRIs of real rectal cancer patients. Compared with low-level radiomic and high-level DL features, the hybrid radiomic prediction model significantly improved the accuracy of treatment response prediction towards personalized treatment.

Funding Support, Disclosures, and Conflict of Interest: Key Research and Development Program of Shaanxi (ProgramNo.2022GY-315)


MRI, Feature Selection, Quantitative Imaging


TH- Response Assessment: Imaging-based: MRI

Contact Email