Registration of Longitudinal Liver Examinations for Tumor Progress Assessment

TL;DR

This work addresses the difficulty of assessing liver tumor progression across longitudinal CT scans, where non-rigid liver deformations and temporal changes can distort direct image comparisons. It introduces a segmentation-guided cyclic diffeomorphic registration method that operates on liver segmentation masks, using a forward field $\phi_{AB'}$ and a cyclic refinement to ensure smooth, invertible deformations while preserving tumor burden. The approach combines segmentation-alignment loss $L_{sim}$ with regularization terms for smoothness, anti-folding, and inverse consistency, extended through an Incremental Cyclic framework to handle large deformations. Experimental results on a multi-patient dataset show superior field regularity and comparable content-based similarity to baselines, with tumor-specific metrics indicating that smoother deformations reduce bias in tumor measurements and support more reliable progression assessment. This method has practical implications for reducing radiologists' workload and improving RECIST-based tumor monitoring in longitudinal liver CT studies.

Abstract

Assessing cancer progression in liver CT scans is a clinical challenge, requiring a comparison of scans at different times for the same patient. Practitioners must identify existing tumors, compare them with prior exams, identify new tumors, and evaluate overall disease evolution. This process is particularly complex in liver examinations due to misalignment between exams caused by several factors. Indeed, longitudinal liver examinations can undergo different non-pathological and pathological changes due to non-rigid deformations, the appearance or disappearance of pathologies, and other variations. In such cases, existing registration approaches, mainly based on intrinsic features may distort tumor regions, biasing the tumor progress evaluation step and the corresponding diagnosis. This work proposes a registration method based only on geometrical and anatomical information from liver segmentation, aimed at aligning longitudinal liver images for aided diagnosis. The proposed method is trained and tested on longitudinal liver CT scans, with 317 patients for training and 53 for testing. Our experimental results support our claims by showing that our method is better than other registration techniques by providing a smoother deformation while preserving the tumor burden (total volume of tissues considered as tumor) within the volume. Qualitative results emphasize the importance of smooth deformations in preserving tumor appearance.

Figures (7)

• Figure 1: Liver longitudinal exams: The images show a growing lesion in red and reveal changes in the liver appearance after two months, e.g., effusion around the liver (orange arrows). The liver segmentation mask is presented in green, existing tumors in blue, and new tumors in red.
• Figure 2: Cyclic Diffeomorphic Registration:$g_{\theta_1}$ takes as input the moving and fixed segmentation masks $S_A$ and $S_B$ and generates a displacement field $\phi_{AB'}$. This field is applied to $S_A$ to obtain $S_{B'}$, the aligned segmentation mask. $S_{B'}$ goes through $g_{\theta_2}$ (with $S_A$) to obtain the cyclically transformed segmentation $S_{A'}$.
• Figure 3: Incremental Cyclic Diffeomorphic Registration Framework: The model takes as input $S_A$ and $S_B$ and generates two displacement fields, $\phi_{AB'_{temp}}$ and $\phi_{B'_{temp}B'}$, to produce the aligned segmentation $S_{B'}$. $S_{B'}$ then goes through the backward path (with $S_A$) to obtain the cyclically transformed segmentation $S_{A'}$.
• Figure 4: Left to right: Moving image $A$, displacement field $\phi$ (in red), transformed image $B'$, and fixed image $B$. 3D liver masks are presented in blue for A, red for B', and green for B. Red arrows highlight unrealistically stretched regions.
• Figure 5: Left to right: Moving image A (tumor in blue), transformed image B' for NiftyReg and our proposed framework DiffeoCyc_inc-2 (tumor in red), and fixed image B (tumor in orange). The transformed liver masks are represented in green, and the fixed image mask B is represented in red.