Deep Learning Prediction of MRI Data for Evaluating Combined Transcatheter NK Cell Local Delivery and Sorafenib Therapy in Hepatocellular Carcinoma

Monday, March 31, 2025

3:27 PM – 3:36 PM CT

Location: Music City Center, 205 A

Presenting Author(s)

Guangbo Yu, MS

PhD Student
University of California, Irvine, United States

Author/Co-author(s)

Zigeng Zhang, MD

Postdoctoral Fellow
University of California, Irvine, United States
AE

Aydin Eresen, PhD (he/him/his)

Assistant Researcher
University of California, Irvine, United States
QH

Qiaoming Hou, n/a

Lab Manager
University of California, Irvine, United States
VY

Vahid Yaghmai, n/a

Chairman of Radiology
University of California, Irvine, United States
ZZ

Zhuoli Zhang, MD, PhD

Principle Investigator
University of California Irvine, Department of Radiological Sciences, United States

Purpose: Accurate early assessment of treatment responses in hepatocellular carcinoma (HCC) post-NK cell therapy is complicated by pseudoprogression, presenting a significant clinical challenge. This study aims to compare deep learning and traditional radiomics approaches to precise evaluation of HCC treatment outcomes following combined intraarterial transcatheter NK cell delivery and Sorafenib administration, demonstrating the superior ability of deep learning models in capturing nuanced texture features from MRI images that are invisible to the naked eye.

Materials and Methods: Twenty-four Sprague-Dawley Rats were implanted with N1S1 tumors, randomly divided into four groups: control, NK cell therapy, Sorafenib, and a combination of NK cell and Sorafenib therapies. Weekly MRI scans over four weeks provided multi-parametric data (T1w, T2w, T1w post-contrast), with tumor regions delineated. Radiomics features were extracted using four filters, followed by feature selection to eliminate highly correlated features and Recursive Feature Elimination (RFE) to further narrow down to five key features, minimizing overfitting. An XGBoost model was then developed for treatment outcome differentiation. Concurrently, an EfficientNet-based deep learning model was designed to autonomously extract and classify features from the ROI. The models' performance was evaluated through accuracy and the area under the curve (AUC) of the receiver operating characteristic (ROC) curve.

Results: The analysis identified five key MRI features for the predictive model. The XGBoost model achieved an accuracy of 86.6%, AUC of 0.94, sensitivity of 86.5%, and specificity of 95.8%. In contrast, the deep learning approach demonstrated an accuracy of 88.28%, AUC of 0.96, sensitivity of 87.4%, and specificity of 96.8%.

Conclusion:

Integrating T1-weighted, T2-weighted, and T1-weighted post-contrast multi-parametric MRI analyses via deep learning outperforms traditional radiomics in predicting treatment responses, marking a crucial shift towards utilizing deep learning for the dynamic and early assessment of combination therapies in HCC. This development offers a promising path to improve clinical outcomes through precise treatment evaluation.