The purpose of this project was to develop a fully automated deep learning algorithm to predict the response of hepatocellular carcinoma to transcatheter arterial chemoembolization therapy using pre-therapeutic quantitative CT images and clinical data.

Sponsor

 

Team Members

Ares Vega | Jena Everett | Liam Raehsler | Zeyuan Wang | Ziyu Wang | Zhimin Zhang | | | | | |

 

Project Poster

Click on any image to enlarge.


Project Video

video player icon

Project Summary PDF

pdf icon


Project Summary

Overview

Hepatocellular carcinoma (HCC) is the most common primary hepatic malignancy. As patients have variable responses to therapy for HCC, individualized treatment is paramount for improving survival and quality of life. Patients who diagnosed this disease are usually in the intermediate stage which is not available for curative therapies anymore. For unresectable HCC, patients require other treatments and will have different responses to them.

Objectives

Develop a deep learning prediction model based on the computerized tomography (CT) images to predict the HCC response to one of the treatments: transcatheter arterial chemoembolization (TACE) before the therapy starts.

Approach

– Review literature works on the related topic.

– Meet with sponsor contact and faculty advisor biweekly to make sure the project is on track.

– Generate database using patients’ CT images and clinical data provided by the sponsor.

– Develop and train Neural Network (3D CNN) models for liver and tumor segmentation.

– Identify the clinical gold standards for TACE prediction and use it to format the training labels for classifier.

– Build and train a Random Forest classification model with the liver segmentations and the derived labels.

– Evaluate accuracy of the models using different approaches.

Outcomes

– Use DICE loss and over sampling method to solve imbalanced data problem for the CNN models.

– Improve CNN model prediction accuracy against the manual segmentation through implementing ResNet Models and use DICE score to evaluate the segmentation accuracy (Liver DICE score: 94.0%, Tumor DICE score: 64.1%).

– Build and fine-turn random classifier to improve the prediction accuracy to 86.56