Abstract
Standard medical image segmentation architectures, such as UNets and Transformers, often struggle to delineate sharp anatomical boundaries due to the low-pass filtering nature of grid-based operations. We present MORSE (iMplicit anatOmical Rendering with Stochastic Experts), a framework that treats segmentation as a continuous rendering problem. By combining a stochastic mixture-of-experts (SMoE) module with implicit neural representations, MORSE adaptively refines high-frequency boundary regions. Evaluated on the Synapse multi-organ and MP-MRI liver datasets, MORSE consistently improves boundary accuracy and overall Dice scores across both 2D and 3D backbones, outperforming existing point-based refinement methods.
Introduction
In clinical settings, the precise definition of organ boundaries is critical for diagnostic accuracy and surgical planning. However, traditional deep learning models for medical image segmentation typically operate on discrete pixel or voxel grids. The fundamental operations of these models—convolutions and self-attention—tend to act as low-pass filters, which inherently blur high-frequency details. This leads to a loss of precision at the anatomical interfaces where organs meet.
Existing attempts to solve this, such as point-based refinement, often suffer from "representation collapse" or require prohibitively expensive pixel-level supervision. MORSE addresses these challenges by reformulating segmentation as a continuous coordinate-based rendering task, allowing the model to query and refine specific coordinates with high-frequency spatial information.
Methodology
The MORSE framework introduces three primary innovations to the standard encoder-decoder segmentation pipeline:
1. Stochastic Mixture-of-Experts (SMoE)
Rather than relying on a single feature map from the final decoder layer, MORSE utilizes multi-scale representations from various decoder blocks, treating them as individual "experts." To ensure robust feature specialization and prevent the model from over-relying on a single expert, the authors implement Stochastic Routing. Similar to dropout, this mechanism randomly activates different experts during training, while a gating network learns to optimally weight their contributions during inference.
2. Implicit Anatomical Rendering (IAR)
The IAR module serves as the continuous rendering engine. Instead of processing the entire image at high resolution, it employs an uncertainty-based point selection strategy. The model identifies coordinates where the initial coarse prediction is most uncertain—typically at the anatomical boundaries. For these points, the model concatenates coarse features with Fourier features (sinusoidal encodings of coordinates) to capture high-frequency spatial details. A compact three-layer MLP then "renders" the final, fine-grained label for those specific coordinates.
3. Adaptive Weight Adjustment (AWA)
Training a refinement head simultaneously with a base model can be unstable. MORSE utilizes an Adaptive Weight Adjustment (AWA) scheduling mechanism. This ensures the model first learns a stable global "coarse" structure before gradually increasing the importance of the rendering loss ($\mathcal{L}_{rend}$) to polish the fine details.
Experimental Results
The framework was validated on two major benchmarks: the Synapse Multi-organ Segmentation dataset (CT) and a Multi-phasic MRI (MP-MRI) Liver dataset.
Quantitative Performance
MORSE demonstrated significant performance gains across multiple backbones, including UNet, TransUNet, and UNETR:
- Synapse Dataset: Integrating MORSE into a standard UNet improved the Dice Similarity Coefficient (DSC) from 70.11% to 76.59%. When paired with TransUNet, it achieved a state-of-the-art 80.85% DSC.
- Boundary Precision: The model significantly reduced the 95% Hausdorff Distance (95HD), confirming that the rendered boundaries are geometrically closer to the ground truth than those produced by grid-based methods.
- Comparison with PointRend: MORSE consistently outperformed PointRend by margins of 1.5% to 5%, suggesting that implicit representations are better suited for the complexities of medical imaging than standard point-sampling techniques.
Discussion and Insights
The success of MORSE highlights several key principles for the future of medical imaging:
- Continuous vs. Discrete: Viewing segmentation as a rendering task allows models to bypass the resolution limits of the underlying grid, enabling sub-pixel precision.
- High-Frequency Injection: Explicitly using sinusoidal positional encodings forces the network to overcome its natural bias toward low-frequency functions, which is essential for capturing sharp organ edges.
- Stochasticity in Experts: The use of stochastic routing is vital for maintaining diversity in multi-scale features, preventing the "collapse" where the model ignores useful information from shallower or deeper layers.
Limitations and Future Work
While MORSE provides superior accuracy, the addition of the gating network and multiple MLPs increases the parameter count and computational complexity compared to vanilla architectures. Furthermore, the IAR module is dependent on the quality of the initial coarse prediction; if the base model fails to detect a small organ entirely, the point selection mechanism may not have the opportunity to refine it. Future research may explore applying this coordinate-based approach to semi-supervised learning to improve generalization in low-data regimes.
Conclusion
MORSE represents a significant step forward in high-precision medical image segmentation. By shifting from discrete grid labeling to continuous implicit rendering, it provides a robust solution for the long-standing problem of boundary blurring in deep learning models.
Summary updated