Source-linked AI summary
Shape-Aware Organ Segmentation by Predicting Signed Distance Maps
Yuan Xue, Hui Tang, Zhi Qiao, Guanzhong Gong, Yong Yin, Zhen Qian, Chao Huang, Wei Fan, Xiaolei Huang
TL;DR
The paper targets organ-segmentation systems that may fail to preserve overall anatomy and smoothness. It predicts signed distance maps jointly with segmentation maps through an approximated Heaviside function, and reports strong segmentation with smoother, more continuous shapes across hippocampus and multi-organ datasets.
Problem
Current deep learning organ-segmentation systems often lack awareness of feasible organ shape and produce nonsmooth results, particularly with inconsistent 2D-slice annotations.
Method
The method adapts a 3D segmentation network to jointly predict signed distance maps and segmentation maps using a differentiable approximation to the Heaviside function.
Results
Across hippocampus and multi-organ segmentation, SDM learning retains strong segmentation performance while producing smoother shapes and improving reported distance-based measures; the backbone also achieves state-of-the-art results.
Takeaways & Limitations
SDM learning can be incorporated into existing 3D segmentation networks with nearly no additional overhead and may support organ-segmentation applications involving shape analysis and radiotherapy planning.
Takeaways & Limitations
In multi-organ segmentation, SDMs are not perfectly learned for small organs, and the shared network architecture is constrained by GPU memory despite requiring greater capacity.
Abstract
from arXiv · showhide
In this work, we propose to resolve the issue existing in current deep learning based organ segmentation systems that they often produce results that do not capture the overall shape of the target organ and often lack smoothness. Since there is a rigorous mapping between the Signed Distance Map (SDM) calculated from object boundary contours and the binary segmentation map, we exploit the feasibility of learning the SDM directly from medical scans. By converting the segmentation task into predicting an SDM, we show that our proposed method retains superior segmentation performance and has better smoothness and continuity in shape. To leverage the complementary information in traditional segmentation training, we introduce an approximated Heaviside function to train the model by predicting SDMs and segmentation maps simultaneously. We validate our proposed models by conducting extensive experiments on a hippocampus segmentation dataset and the public MICCAI 2015 Head and Neck Auto Segmentation Challenge dataset with multiple organs. While our carefully designed backbone 3D segmentation network improves the Dice coefficient by more than 5% compared to current state-of-the-arts, the proposed model with SDM learning produces smoother segmentation results with smaller Hausdorff distance and average surface distance, thus proving the effectiveness of our method.
Introduction
The paper addresses organ-segmentation outputs that may lack anatomical shape awareness and smoothness, especially when annotations are drawn on 2D slices. It proposes learning signed distance maps alongside segmentation maps to encode global shape information while retaining segmentation performance.
- Organ segmentation supports disease diagnosis and surgical planning, but manual annotation is labor-intensive and error-prone.
- Current deep learning organ-segmentation systems may lack feasible-shape awareness and produce nonsmooth 3D results because annotations are inconsistent across 2D slices.Figure 1 contrasts these limitations in hippocampus segmentation.
- Signed distance maps encode distance, inside/outside position, and global structural information, so predicting them encourages continuity and smoothness.Unlike binary maps, local shape changes can alter SDM values across multiple points.
- The method jointly predicts SDMs and segmentation maps through a differentiable Heaviside approximation, leveraging their rigorous mapping while retaining high-quality segmentation outputs.The SDM branch supplies shape priors and the segmentation branch contributes conventional segmentation information.
- The proposed 3D U-Net backbone uses deeper layers and large receptive fields to segment both large and small organs, achieving state-of-the-art results on the MICCAI 2015 dataset.
Related Works
Prior organ-segmentation work spans statistical, atlas-based, contour, level-set, deep-learning, boundary, surface, and distance-function approaches. These methods improve accuracy or shape handling in different ways, but the cited literature motivates end-to-end methods that integrate shape information into segmentation.
- Statistical, atlas-based, active-contour, and level-set methods require registration, label fusion, or iterative optimization, whereas deep learning offers faster and more accurate organ segmentation.
- Deep learning outputs can contain extraneous regions and fail to preserve organ anatomy, motivating post-processing such as fully connected CRFs for refinement.
- End-to-end alternatives incorporate level-set smoothness terms or adversarial learning during training to reduce reliance on post-processing.
- Boundary and surface-based methods represent shape through organ boundaries, including boundary losses and elastic-shell surface prediction.
- Prior SDM or SDF work includes volumetric-surface reconstruction, saliency refinement, continuous shape learning, and unsigned distance maps for organ prediction.
Methodology
The method predicts signed distance maps alongside segmentation maps using a 3D UNet, linking shape-aware representation with conventional segmentation supervision. It combines differentiable Heaviside conversion, specialized SDM losses, and a backbone evaluated across organ-segmentation settings.
- SDM learning: The model jointly predicts SDMs and segmentation maps by converting SDMs through a differentiable approximated Heaviside function during training.The final objective combines segmentation loss with SDM loss.
- Deep 3D UNet: 3D inputs and a modified 3D UNet provide global and multi-scale features for capturing organ shape across large and small organs.The architecture uses multiple downsampling operations while retaining mixed receptive-field sizes through skip connections.
- SDM representation: An SDM encodes signed distance to the organ surface, so shape changes affect values globally rather than only at local points.Negative values indicate interior points, positive values exterior points, and the zero level set marks the organ surface.
- SDM learning: Groundtruth SDMs are computed from segmentation maps, normalized to [-1, 1], and predicted with a tanh output layer.The normalization accounts for varying fields of view and organ volumes.
- SDM learning: The SDM objective combines a product-based regression loss with L1 loss to penalize incorrect signs and emphasize values near the boundary.The combined loss is described as providing smoother gradients and more stable SDM training than L1 alone.
- Evaluation: On the hippocampus dataset, joint SDM and segmentation training achieves the best scores across all reported evaluation metrics using the same backbone.The comparison includes Dice coefficient, Hausdorff Distance, 95% Hausdorff Distance, and Average Symmetric Surface Distance.
Experiments
Experiments evaluate the proposed models on hippocampus and multi-organ head-and-neck segmentation, using qualitative comparisons and standard segmentation metrics. Joint SDM and segmentation training generally improves boundary quality, smoothness, and false-positive control.
- Hippocampus: The joint SDM-and-segmentation model achieves the best hippocampus agreement with ground truth across the qualitative comparison.Segmentation-only outputs contain isolated or inaccurate false-positive regions, while joint training produces the closest contours.
- Hippocampus: Joint training produces more accurate hippocampus boundaries, while SDM-only training yields the smoothest contours and both preserve organ shape.These results support direct SDM prediction as feasible for capturing hippocampus shape information.
- Multi-organ segmentation: The MICCAI 2015 dataset evaluates methods on 38 training CT images and 10 testing images, with comparisons against state-of-the-art multi-organ systems.The study targets the more challenging multi-organ segmentation setting.
- Multi-organ segmentation: The proposed methods improve Dice and HD95 over prior state-of-the-art methods, especially for small organs, while joint training slightly lowers Dice versus Dice-only training.Joint training nevertheless outperforms other state-of-the-art methods by a large margin and performs better on the other reported metrics.
- Ablation comparison: The proposed SDM loss achieves the best HD and ASD scores across the ablation comparison and preserves continuous shape on all organs.SDM-only training and joint training with L1 loss fail to produce meaningful results for some small organs or classes.
Discussion
The discussion finds that SDM learning improves hippocampus segmentation and preserves shape, but multi-organ SDMs remain imperfect for small organs. Capacity and cross-organ modeling limitations constrain the current multi-organ design.
- Findings: SDM learning gives superior hippocampus performance versus segmentation-only or SDM-only training while producing smooth fixed-distance contours that retain hippocampus shape.For multi-organ segmentation, learned SDMs preserve shape in larger and medium organs but not reliably in the smallest organ, Chiasm.
- Benchmark results: The MICCAI testing tables report that the proposed backbone outperforms current methods in Dice, while joint training improves HD95 by 0.64 on average.The HD95 comparison is reported on the MICCAI 2015 testing set.
- Limitations: A key limitation is that GPU memory forced the same network architecture for single-organ and multi-organ datasets, despite the latter requiring greater capacity.The authors state that multi-organ segmentation requires more information propagation through the network.
- Limitations: Each organ’s SDM is predicted independently in the final layer, leaving no connection between organs and limiting feature sharing in the multi-organ model.The authors propose exploring relationships among organ SDMs in future work.
Conclusions
The work identifies SDM learning as a promising approach for organ segmentation and reports improvements over previous methods. The mechanism also adapts existing 3D segmentation networks with nearly no additional overhead.
- SDM learning combined with traditional segmentation-map training improves state-of-the-art organ-segmentation results by a large quantitative and qualitative margin.
- Existing 3D segmentation networks can incorporate the SDM prediction model with nearly no additional overhead.
- The mechanism has potential applications in radiotherapy planning, shape analysis, and general segmentation tasks.