MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Depthwise Separable Convolution Neural Network for Survival Prediction of Head & Neck Cancer

R Li1*, A Das2, N Bice1, P Rad2, A Roy2, N Kirby1, N Papanikolaou1, (1) University of Texas HSC SA, San Antonio, Texas, (2) The University of Texas at San Antonio

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
To evaluate the performance of a novel deep learning radiomics model in predicting the short-term and long-term survival probability for patients with Head & Neck cancer.

Methods:
A cohort of 466 patients with Stage I to IV Head & Neck squamous cell carcinoma (HNSCC) from three independent datasets were selected. For each patient in the training set, combinations of CT slice and pre-segmented gross tumor volume (GTV) slice (input size: 2×100×100 voxels) were fed into the CNN model. A deep learning model with depthwise separable convolutions (DSC) was trained to predict survivability from the voxel information. Prediction endpoints included 2-year survival, 5-year survival and locoregional recurrence (LR). For comparison, a random forest (RF) classifier was built using the top 13 features selected from 1189 handcrafted features by the LASSO logistic regression method. To assess the performance, the area under receiver operator characteristic curve (AUC) for each endpoint was calculated on the test set. Kaplan-Meier curves were calculated for high and low mortality risk groups with an optimal score stratification.

Results:
The proposed CNN model achieved AUC of 0.762 [0.699-0.854, 95% CI], 0.659 [0.593-0.741, 95% CI] and 0.645 [0.560-0.751, 95% CI] for 2-year survival, 5 -year survival, LR prediction in the test set, while the RF classifier based on handcrafted radiomics features achieved AUC of 0.802[0.703-0.887, 95% CI], 0.727 [0.654-0.813, 95% CI], 0.599 [0.499-0.692, 95% CI]. Both models were able to stratify patients into a high and low modality risk group in the test set, with a log-rank test p value<0.01.

Conclusion:
These results suggest that a training-from-scratch DSC network shows potential in improving the performance of survival prediction using a relatively small training dataset. The prediction power was comparable to the existing handcrafted radiomics model, as well as other complicated CNN models.

Keywords

Statistical Analysis, Feature Extraction

Taxonomy

TH- Response Assessment: Modeling: Machine Learning

Contact Email