97 - Multi-Dimensional Vision Transformer Modeling of Lipiodol Staining for Outcome Prediction Following TACE in Hepatocellular Carcinoma

Monday, April 13, 2026

4:03 PM - 4:12 PM ET

Presenting Author(s)

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

Interventional Radiology Resident
Johns Hopkins Hospital, United States

Author/Co-author(s)

Tony Tae Kyung Kim, MD

Resident Physician
Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins Hospital, United States
Alec Jost, MD

Resident
Johns Hopkins Hospital, United States
CL

Carolina Lugo-Fagundo, MD

Resident Physician
Johns Hopkins University, United States
KR

Kent Reichel, MD

Integrated IR/DR Resident
Johns Hopkins Hosptial, 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

Purpose: To develop a multi-dimensional 3D vision transformer (ViT) model that analyzes lipiodol staining patterns on post-TACE CT imaging to predict key clinical outcomes in hepatocellular carcinoma (HCC), including Milan criteria status, mRECIST response, progression-free survival (PFS), and overall survival (OS).

Materials and Methods: This retrospective study included 200 patients with biopsy-confirmed HCC treated with transarterial chemoembolization (TACE). Non-contrast CTs within 30 days post-TACE were used to evaluate lipiodol staining. Lesions were segmented using Segment Anything 2 (SAM2), generating 3D masks. A Multi-Dimensional ViT model was constructed with 3D patch embeddings (4×4×4 voxels), hierarchical transformer blocks, and shifted window attention across four stages (C=96→768). Spatial context was captured using relative positional encoding. Clinical variables (age, sex, MELD, Child-Pugh, systemic therapy) were embedded and fused with imaging features. Output heads included softmax classifiers (for 6-month Milan and mRECIST) and Cox proportional hazards layers (for PFS/OS). Training was performed on 160 patients; 40 patients were used for testing. Optimization included AdamW with Cox, cross-entropy, and MSE losses, plus dropout and data augmentation. Model performance was evaluated using AUC, sensitivity (SE), specificity (SP), PPV, NPV, accuracy, F1 score, and R².

Results: The Multi-D ViT model achieved strong predictive performance with results as below:

AUC (Milan) = 0.95; SE = 92.0%, SP = 87.5%, PPV = 88.5%, NPV = 93.0%, Accuracy = 90.0%, F1 = 90.1%

AUC (mRECIST) = 0.85

R² (PFS) = 0.89, R² (OS) = 0.84

Conclusion: Lipiodol staining patterns captured on post-TACE CT contain spatial features predictive of future transplant eligibility, radiologic response, and survival in patients with HCC. Our 3D vision transformer model effectively integrates volumetric imaging and clinical data, enabling automated, accurate prognostication following TACE. These results highlight the promise of transformer-based AI tools in guiding personalized management for liver cancer patients.