Deep Learning-Based Prediction of Transarterial Chemoembolization Target Lesion Response in Hepatocellular Carcinoma Post-TACE Using DSA Data

Tuesday, April 1, 2025

3:18 PM – 3:27 PM CT

Location: Music City Center, 209 ABC

Presenting Author(s)

Tej I. Mehta, MD (he/him/his)

Interventional Radiology Resident
Johns Hopkins Hospital, United States

Author/Co-author(s)

David Gage, MD (he/him/his)

Interventional Radiology Resident
The Johns Hopkins Hospital, United States
CH

Caleb Heiberger, MD

Radiology Resident
Johns Hopkins, United States
Christopher Bailey, MD (he/him/his)

Assistant Professor
Interventional Radiology, Johns Hopkins University School of Medicine, United States
RL

Robert P. Liddell, MD, FSIR

Division Chief, Associate Professor Radiology and Radiological Sciences
The Johns Hopkins Hospital, Baltimore, Maryland, United States, United States
HB

Harrison Bai, MD, MS

Director, Radiology AI Lab
Johns Hopkins University, United States

Purpose: Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide. Transarterial chemoembolization (TACE) is a primary locoregional therapy for patients with intermediate-stage HCC. Early prediction of target lesion (TL) response to TACE, as defined by the modified Response Evaluation Criteria in Solid Tumors (mRECIST), is critical for optimizing patient management and improving outcomes. We sought to develop and validate a radiomics based machine learning (ML) model utilizing intra-TACE digital subtraction angiography (DSA) data for early prediction of TL response by mRECIST criteria in patients with HCC.

Materials and Methods: 240 patients with biopsy-confirmed HCC who underwent TACE, were naïve to locoregional liver therapy and had pre-TACE, 6 months post-TACE MRIs and first time TACE DSAs were included. Subselective DSA sequences demonstrating the region of tumor blush were spatially registered using Demons Elastic Registration, the median pixel values for each DSA image in the run were extracted and these pixel values were superimposed to create a single DSA image highlighting the region of tumor blush. Segment Anything, a deep learning-based semi-automated segmentation algorithm, was used for semi-automated segmentation of target HCC lesions on DSA. 2-D radiomic features of these target lesions were extracted using Pyradiomics. TL response criteria were taken from the pre and post TACE MRIs, which are reported at our institution per protocol. The ML model was generated using Autogluon as tabular predictors with TL response as the outcome with an 80/20 train/test split. Model performance was evaluated using accuracy (AC), sensitivity (SE), specificity (SP), and area under the receiver operating characteristic curve (AUC).

Results: The highest performing model was an ensemble model with a Neural Net Fast AI model in the first layer. The model achieved an AUC of 0.87 for TL prediction by mRECIST criteria at 6 months post-TACE. It demonstrated an AC of 84%, SE of 90% and SP of 78%.

Conclusion: The ML model effectively predicts TL response by mRECIST criteria using DSA data in patients with HCC. This approach could be utilized to help predict future TL response and guide HCC therapies.