Multi-view Cardiac Image Segmentation via Trans-Dimensional Priors

TL;DR

This work tackles the challenge of accurate multi-view cardiac segmentation by exploiting trans-dimensional priors that transform segmentation maps between short-axis and long-axis MR images. It introduces a sequential 3D-to-2D-to-3D pipeline augmented with trans-dimensional segmentation priors (TDSP) and a Heart Localization and Cropping (HLC) module to focus on the heart ROI. The approach yields state-of-the-art performance on the M&Ms-2 dataset, improving robustness across centers and pathologies while reducing computation through ROI cropping. By leveraging cross-view information and anatomically guided priors, the method delivers plausible segmentations of LV, RV, and MYO across SA and LA views, with publicly available code and pretrained models.

Abstract

We propose a novel multi-stage trans-dimensional architecture for multi-view cardiac image segmentation. Our method exploits the relationship between long-axis (2D) and short-axis (3D) magnetic resonance (MR) images to perform a sequential 3D-to-2D-to-3D segmentation, segmenting the long-axis and short-axis images. In the first stage, 3D segmentation is performed using the short-axis image, and the prediction is transformed to the long-axis view and used as a segmentation prior in the next stage. In the second step, the heart region is localized and cropped around the segmentation prior using a Heart Localization and Cropping (HLC) module, focusing the subsequent model on the heart region of the image, where a 2D segmentation is performed. Similarly, we transform the long-axis prediction to the short-axis view, localize and crop the heart region and again perform a 3D segmentation to refine the initial short-axis segmentation. We evaluate our proposed method on the Multi-Disease, Multi-View & Multi-Center Right Ventricular Segmentation in Cardiac MRI (M&Ms-2) dataset, where our method outperforms state-of-the-art methods in segmenting cardiac regions of interest in both short-axis and long-axis images. The pre-trained models, source code, and implementation details will be publicly available.

Figures (6)

• Figure 1: Overview of other related methods and our proposed pipeline. The proposed framework (d) segments $I_{SA}$ using TriggerNet to produce $S_{SA1}$. It then generates $S_{LA}$ using LA-SegNet and refines $S_{SA2}$ using SA-SegNet, along with relevant segmentation prior and image. Compared to (a), (b), and (c), our method (d) performs a two-way transformation (LA2SA and SA2LA) along with the utilisation of transformed maps as guidance for the HLC module.
• Figure 2: Comparison of segmentation network complexities regarding the number of parameters and Multiply-Accumulate (MAC) operations with and without using the HLC module. (a) SA-SegNet without HLC Module (TriggerNet), and (Depth, Height, Width = 64,192,192), (b) SA-SegNet With HLC Module (SA-SegNet) and (Depth, Height, Width = 112,128,112), (c) LA-SegNet without HLC Module and (Height, Width = 384,384), and (d) LA-SegNet with HLC Module and (Height, Width = 128,128).
• Figure 3: Proposed Heart Localization and Cropping (HLC) module. The heart region is localized and cropped in both intensity and label images.
• Figure 4: Comparison of visual results under different settings for LA segmentation. From the left, the first column is the baseline results, and the second and third columns utilize HLC module and segmentation priors, respectively. The fourth column is the best results using both. The fifth column represents the SA2LAmap.
• Figure 5: Comparison of visual results under different settings for SA segmentation. From the left, the first column is the baseline results, and the second utilizes the HLC module and $S_{SA1}$ as segmentation priors. The third column shows the best results using both segmentation priors. The fourth column represents the LA2SAmap.