A Baseline Analysis of Cross-Modal Liver Tumor Segmentation and the Role of Frozen Encoder

Abstract

Accurate liver tumor segmentation is critical for surgical planning, volumetry, and treatment monitoring across diverse liver pathologies and clinical applications. Both Computed tomography (CT) and Magnetic resonance imaging (MRI) serve as essential modalities in clinical practice, yet segmentation models trained on one modality typically fail when applied to the other. This limitation reflects a fundamental gap in the understanding of why complex architectural innovations are necessary for cross-modal robustness. Most literature proposes sophisticated solutions without first establishing what simpler methods achieve or where they encounter irreducible obstacles.

This thesis provides a systematic baseline analysis of cross-modal liver and tumor seg- mentation, deliberately adopting simple approaches to characterize their capabilities and limitations. The study employs a frozen ResNet18 encoder combined with a U-Net decoder within a two-stage pipeline that first segments the liver, then detects lesions within the hepatic region. This straightforward architecture is evaluated across five public datasets spanning both modalities: LiverHCCSeg and CHAOS for MRI, and LiTS, 3D-IRCADb-01, and SLiver07 for CT.