Program Information

Feature Selection and Prediction of Radiation-Induced Lung Damage in Radiotherapy of Lung Cancer with Deep Multi-Layer Neural Network-Based Methods

S Cui

s cui*, Y Luo , R Ten Haken , I El Naqa , University of Michigan, Ann Arbor, MI

Presentations

SU-J-CAMPUS-JT-2 (Sunday, July 30, 2017) 4:00 PM - 5:00 PM Room: Joint Imaging-Therapy Theater

Purpose: A major challenge in application of deep learning methods to predict radiotherapy outcomes is the ratio of large number of variables to limited sample size. Therefore, to enable prediction of acute lung radiation pneumonitis (RP) using deep neural network, different selection methods for identifying optimal features were evaluated on a large-scale heterogeneous dataset.

Methods: We applied deep learning feature selection onto a pool of 216 variables including clinical factors (e.g. dose) and biomarkers (e.g. SNPs, cytokines µ-RNAs) in a population of 106 patients with 22 grade 2 or higher RP. Extremely Randomized (ER) Trees method was applied to reduce the size of the feature set from 216 to 40. Multi-layer perceptron (MLP) was then employed to make further selection. Specifically, 6 MLP-based methods were investigated, including weight pruning (WP), saliency based pruning (SBP), percentile absolute derivatives(PAD), mean absolute derivatives (MAD), feature quality index (FQI) and feature-based sensitivity of posterior probability (FSPP). To account for overfitting, we did 5-fold cross-validation randomized 20 times. Synthetic minority oversampling was applied to correct for class imbalance. The Top15% and Kemeny rule were combined to aggregate feature rankings obtained from different folds and cross-validations. Finally, the selected features by each method were fed into the deep MLP to predict RP2.

Results: MLP based on WP and FSPP features yielded the best predictions. Moreover, WP showed preferable higher convergence rate. Subsets of top 9/10 features selected by forward selection and FSPP/ WP achieved AUC of 0.843/0.850 on cross-validation. All MLP methods outperformed random forest (RF) in the range from 6 to 10 features.

Conclusion: This work demonstrates the potential of deep learning in modeling radiotherapy toxicities, which is made possible by application of advanced MLP-based optimal feature selection methods. Particularly, WP is a promising feature selection method for deep learning applications in radiotherapy outcome models.


Contact Email: