Program Information
Predicting Risk in NSCLC Patients Using Learned Tumor Sub-Region Appearance From Quantitative Features in CT Images
R Neph*, K Sheng , UCLA School of Medicine, Los Angeles, CA
Presentations
TH-AB-201-11 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 201
Purpose: To learn tumor sub-regions in NSCLC patients and predict risk by classification on sub-region quantitative imaging (QI) features.
Methods: Radiologist segmented pre-treatment CT volumes for 313 NSCLC patients were included. For each sequence, 57 patch-based quantitative imaging features were calculated for each voxel in the Gross Tumor Volume (GTV). Per-subject clustering produced 20 noise-robust supervoxels. Inter-subject hierarchical clustering follows, identifying 7 distinct appearance sub-regions across the cohort. Geometric and textural features were averaged in each appearance group for each subject and used as classification inputs. The risk ground truth was labeled as high or low by median survival time thresholding of each subject. A non-linear classification model was trained using a 5-fold cross validation procedure. For each fold, the classifier was independently trained on 80% (n=251) of the subjects and validation performed on the remaining 20% (n=62). Predictions for the full cohort were obtained during the cross validation. Classifier accuracy and log-rank test p-value were assessed to evaluate the performance of the method.
Results: Classification accuracy for the trained predictor and median thresholding on full GTV volume was 61.7% and 54.1% respectively, indicating the predictive value of QI features in addition to the dominant tumor volume influence. The log-rank test for the classifier showed statistical significance in survival group partitioning (p-value: 8.34e-5). The number of subjects predicted as high-risk by the learned classifier and naive classifier were 126 and 158 respectively while the true high-risk count was 121.
Conclusion: Classification of risk in CT images using sub-region discovery and QI features improved accuracy and prediction of the correct number of subjects in each risk population than naive thresholding on the unclustered GTV volume. The prognostic performance for this CT-only dataset is an indicator that a localizing radiomic signature may be learned from quantitative feature analysis of CT images.
Funding Support, Disclosures, and Conflict of Interest: The authors disclose the following as sources of funding for the work herein: DOE DE-SC0017057, NIH R44CA183390, NIH R01CA188300, NIH R43CA183390, NIH U19AI067769.
Contact Email: