Weakly supervised alignment and registration of MR-CT for cervical cancer radiotherapy

TL;DR

The paper tackles the challenge of accurately registering multimodal MR-CT images to delineate paracervical tissue invasion for cervical cancer radiotherapy. It introduces a two-stage framework: an adaptive spatial position alignment that normalizes inter-layer depth using skeletal structures, and TransFlow, a pyramid-feature, cost-volume-based registration network that estimates dense optical flow for multimodal alignment, augmented by a connected-domain similarity measure and label denoising for weak supervision. Empirically, the approach outperforms volume-rendering-based alignment and several deep-registration baselines across multiple metrics (e.g., DSC, JC, HD, and JD), while maintaining smooth deformation fields. The work advances practical MR-CT fusion for radiotherapy planning by enabling more accurate, automated multimodal alignment with reduced manual intervention and improved anatomical plausibility.

Abstract

Cervical cancer is one of the leading causes of death in women, and brachytherapy is currently the primary treatment method. However, it is important to precisely define the extent of paracervical tissue invasion to improve cancer diagnosis and treatment options. The fusion of the information characteristics of both computed tomography (CT) and magnetic resonance imaging(MRI) modalities may be useful in achieving a precise outline of the extent of paracervical tissue invasion. Registration is the initial step in information fusion. However, when aligning multimodal images with varying depths, manual alignment is prone to large errors and is time-consuming. Furthermore, the variations in the size of the Region of Interest (ROI) and the shape of multimodal images pose a significant challenge for achieving accurate registration.In this paper, we propose a preliminary spatial alignment algorithm and a weakly supervised multimodal registration network. The spatial position alignment algorithm efficiently utilizes the limited annotation information in the two modal images provided by the doctor to automatically align multimodal images with varying depths. By utilizing aligned multimodal images for weakly supervised registration and incorporating pyramidal features and cost volume to estimate the optical flow, the results indicate that the proposed method outperforms traditional volume rendering alignment methods and registration networks in various evaluation metrics. This demonstrates the effectiveness of our model in multimodal image registration.

Figures (11)

• Figure 1: Spatial Position Alignment Steps. Volume Rendering: 3D Slicer using volume rendering for alignment; Spatial Alignment: Spatial position alignment algorithm using the original image and labeled image for automatic alignment.
• Figure 2: The $\mathcal{SIM}$ function denoising process is shown from left to right. (a): The top image is the CT labeled image before denoising, and the bottom image is the MR labeled image. (b) The range of attention of the $\mathcal{SIM}$ function. (c) progresses from left to right, including the local CT image, the local MR image, and the local MR image after resizing. (d) The top image shows the fused local labeled image, while the bottom image shows the denoised CT and MR labeled image.
• Figure 3: The overall process of multimodal registration is shown in the schematic diagram. Firstly, the multimodal images are spatially aligned for interlayer registration, then put into the network for training, the binary ROI contour masks and bone labels of the original image are transformed using the deformation field and the loss is calculated.$L_{l c}$ and $L_{g l}$
• Figure 4: Overview of the registration network
• Figure 5: Average similarity ($\mathcal{SIM}$), connected domain retention ($\mathbf{cd-num}$) (for individual images), and empty image retention ($\mathbf{em-num}$) (for the dataset) trend with the $\gamma$ value. $\gamma=2$ exhibits less noise and rounder edges than $\gamma=5$.