MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Multitask-Based Supervised Deep Learning Using Contrast-Enhanced CT (CECT) Images for Hepatocellular Carcinoma (HCC) Intrahepatic Progression Risk Analysis

L Wei1*, D Owen2 , M Mendiratta-Lala3 , B Rosen2 , K Cuneo2 , T Lawrence2 , R Ten Haken2 , I El Naqa2 , (1) Applied Physics Program, University of Michigan, Ann Arbor, MI, (2) Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, (3) Department of Radiology, University of Michigan, Ann Arbor, MI

Presentations

(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 2

Purpose: To develop a deep learning-based classification and survival network using pre-treatment contrast-enhanced CT (CECT) images for the challenging problem of intrahepatic progression risk estimation of hepatocellular carcinoma (HCC) patients treated with stereotactic body radiation therapy (SBRT).

Methods: 130 HCC patients (77 progressions and 93 deaths) were retrospectively analyzed. 161 radiomic features with varying sampling and quantization parameters and 36 clinical factors were used as input to the network. A supervised variational autoencoder (VAE) was applied to imaging features to learn a compressed latent representation. The network consisting of VAE and multi-layer perceptron (MLP) structures was trained on two related tasks: progression classification and progression-free survival prediction by optimizing an objective function comprising cross-entropy loss (classification), log-sigmoid lower bound of c-index (survival), and VAE loss. A novel cross-stitch unit to learn optimal combination of tasks was introduced into the network architecture. Model performance was assessed by the area under curve (AUC) from receiver operating curves (ROCs), Harrel’s c-index and Kaplan-Meier plots in a 10-fold cross-validation framework.

Results: For the classification task, single task training, multitask training without cross-stitch units, multitask training with cross-stitch using imaging only, and multitask training with cross-stitch units yielded AUC of 0.585 (CI: 0.500-0.696), 0.586 (CI: 0.500-0.676), 0.674 (CI: 0.576-0.762), and 0.727 (CI: 0.633-0.827). For the progression-free survival task, single task training, multitask training without cross-stitch units, multitask training with cross-stitch using imaging only, and multitask training with cross-stitch units yielded c-index of 0.655 (CI: 0.589-0.721), 0.656 (CI: 0.593-0.718), 0.678 (CI: 0.614-0.741), and 0.700 (CI: 0.642-0.758).

Conclusion: Our proposed architecture generalizes across two tasks and gives improved performance for both classification and survival analyses. These results warrant further investigation of neural network structure design to explore underlying biological mechanisms and to leverage the enriched imaging data to improve HCC patient prognosis.

Keywords

Image Analysis, Tumor Control, Modeling

Taxonomy

IM- Dataset analysis/biomathematics: Machine learning

Contact Email