Introduction
Standard 3D medical imaging models typically treat all spatial dimensions equally. However, real-world scans such as CT or MRI are inherently anisotropic; they often possess significantly different resolutions and information densities along the "depth" (z-axis) compared to the cross-sectional (xy-plane) slices. This "one-size-fits-all" approach in standard 3D Convolutional Neural Networks (CNNs) ignores axial disparities, leading to spatial biases and blurry boundaries in final segmentations.
To address this, we introduce the Post-Axial Refiner (PaR), a plug-and-play module designed to "disentangle" these axial features. By acknowledging that the z-axis represents different morphological characteristics than the horizontal plane, PaR significantly improves segmentation accuracy and boundary precision across a wide array of state-of-the-art architectures.
Methods: The Post-Axial Refiner (PaR)
The PaR module is designed to be appended to the end of any existing 3D segmentation backbone, such as 3D U-Net, Swin UNETR, or nnFormer. It refines the "coarse" mask produced by the initial backbone through two parallel pipelines:
- Permutative Dimensional Intensification (PDI): This pipeline reshuffles the feature map dimensions to view the data from multiple perspectives (transverse, sagittal, and coronal). It utilizes multi-head self-attention to capture global dependencies across these orientations, providing the model with a holistic understanding of the anatomical structure.
- Sequential Axial Intensification (SAI): This pipeline focuses specifically on a chosen axis (typically the z-axis). It employs dual attention mechanisms:
- Channel-wise attention: Identifies what features are critical along that axis.
- Spatial-wise attention: Identifies where those features are located.
- Spatial Redistribution: Ensures refined features are correctly integrated back into the original 3D volume.
Experimental Evaluation
The PaR module was validated across three diverse datasets representing different imaging modalities and anatomical structures:
- FLARE2021: Abdomen CT scans for multi-organ segmentation.
- OIMHS: Retinal Optical Coherence Tomography (OCT) for macular hole segmentation.
- SegTHOR: Thoracic CT images for segmenting organs at risk (heart, trachea, esophagus, and aorta).
Key Results
The integration of PaR consistently improved performance across 11 different baseline models. On the FLARE2021 dataset, adding PaR to a standard 3D U-Net improved the Mean Intersection over Union (mIoU) by 1.28% and the Dice score by 0.81%.
More importantly, the module dramatically improved boundary precision. For the 3D U-Net on FLARE2021, the Hausdorff Distance (HD95)—a measure of boundary error—dropped from 16.31 to 2.49. This leap in edge clarity is visually evident when comparing baseline outputs to those refined by PaR.
Discussion and Conclusion
The primary insight of this work is that axial disentanglement is crucial for high-precision 3D medical segmentation. By treating the z-axis as fundamentally different from the xy-plane, the PaR module helps models "fill in the gaps" in sparse data more intelligently than standard convolutions.
While the module adds a marginal computational overhead, its plug-and-play nature makes it a versatile enhancement for various architectures. Future work will focus on automating axis selection and further optimizing the module's efficiency for real-time clinical applications.
Summary updated