Preserving Tumor Volumes for Unsupervised Medical Image Registration

TL;DR

This work tackles the challenge that learning-based medical image registration often distorts tumor volumes, limiting its clinical utility for tracking tumor growth. It introduces a two-stage, unsupervised framework: Stage 1 estimates a soft tumor mask by analyzing volume changes from a similarity-based registration, and Stage 2 trains a volume-preserving registration with an adaptive loss weighted by the mask to preserve tumor sizes while maintaining image alignment. A new metric, Square Tumor Size Ratio (STSR), quantifies tumor-volume preservation, and experiments across LiTS17 and BraTS20 show comparable registration performance to state-of-the-art methods with substantially improved tumor-volume preservation, demonstrating robustness across architectures. The approach is network-agnostic, scales to different imaging modalities and lesions, and provides a practical pathway for reliable tumor growth assessment through registration.

Abstract

Medical image registration is a critical task that estimates the spatial correspondence between pairs of images. However, current traditional and deep-learning-based methods rely on similarity measures to generate a deforming field, which often results in disproportionate volume changes in dissimilar regions, especially in tumor regions. These changes can significantly alter the tumor size and underlying anatomy, which limits the practical use of image registration in clinical diagnosis. To address this issue, we have formulated image registration with tumors as a constraint problem that preserves tumor volumes while maximizing image similarity in other normal regions. Our proposed strategy involves a two-stage process. In the first stage, we use similarity-based registration to identify potential tumor regions by their volume change, generating a soft tumor mask accordingly. In the second stage, we propose a volume-preserving registration with a novel adaptive volume-preserving loss that penalizes the change in size adaptively based on the masks calculated from the previous stage. Our approach balances image similarity and volume preservation in different regions, i.e., normal and tumor regions, by using soft tumor masks to adjust the imposition of volume-preserving loss on each one. This ensures that the tumor volume is preserved during the registration process. We have evaluated our strategy on various datasets and network architectures, demonstrating that our method successfully preserves the tumor volume while achieving comparable registration results with state-of-the-art methods. Our codes is available at: \url{https://dddraxxx.github.io/Volume-Preserving-Registration/}.

Figures (5)

• Figure 1: Red boxes represent the location of tumors in the moving image before warped. In image registration with tumors, similarity-based registration typically leads to significant alterations in tumor size while our volume-preserving registration is capable of preserving tumor anatomy.
• Figure 2: Main Framework. Our framework consists of two stages. In the first stage, a soft mask indicating tumor regions is estimated by analyzing the Jacobian matrix of the deformation field obtained by a registration network pre-trained on similarity loss. In the second stage, the soft mask is used to guide the calculation of both volume-preserving and similarity losses for training the volume-preserving registration network. This deformation field estimated by the stage two ensures that both the image similarity and the preservation of tumor sizes.
• Figure 3: Visualization of the soft tumor mask. (a) is the estimated soft tumor mask without pre-registration and (b) is the one with pre-registration. The gray colors with intensity variations denote the soft mask, while the green and red colors represent the ground truth segmentation for the liver and tumors, respectively.
• Figure 4: Qualitative comparison of different volume-preserving (VP) methods trained on the Liver Tumor Segmentation (LiTS) dataset. The left side of the figure shows two sets of images: Fixed and Ground Truth (GT), and Moving and GT. The first row of the figure displays the warped moving image, while the second row illustrates the organ outlines in green and red for the moving and fixed images, respectively. The yellow overlay highlights the tumors. Our proposed volume-preserving (VP) method ensures the preservation of tumor volume while aligning the images, as demonstrated by reduced number of visible changes in tumor size. In the third row, the Jacobian matrix of the deformation field is visualized. The green and red lines represent the organ and tumor outlines, respectively. The white areas indicate a large Jacobian, which corresponds to a more significant change in volume. The method without volume-preserving loss demonstrates a larger white area in the tumor, indicating a greater volume change of tumor volume. The last row of the figure displays the deformation field. Due to space limitations, qualitative results for BraTS20 dataset are provided in the supplemental material.
• Figure 5: Robustness of our tumor mask estimation. We evaluated the robustness of our tumor mask estimation by running our framework using RCN on the LiTS datasets with different noisy ground truth (GT) masks as tumor masks, each with specific Dice scores with tumors. We plotted the results on a graph with the Dice score on the x-axis and the STSR in the test set on the y-axis (where smaller values are better).