OASIS: One-pass aligned Atlas Set for Image Segmentation

Qikui Zhu, Bo Du, Pingkun YanarXiv 2019 · arXiv

Introduction

Deep Convolutional Neural Networks (CNNs) have set new benchmarks in medical image segmentation but typically require extensive labeled datasets to achieve high performance. In clinical settings, obtaining expert annotations is costly and time-consuming. Traditional atlas-based segmentation methods offer a data-efficient alternative by warping a few "gold standard" labeled images (atlases) to a target scan. However, these methods are often computationally expensive due to iterative registration and can be prone to errors when background noise interferes with organ alignment.

OASIS (One-pass aligned Atlas Set for Image Segmentation) addresses these challenges by combining the representational power of deep learning with the data efficiency of atlas-based frameworks. It introduces a label-constrained registration process and a novel weighting strategy to achieve state-of-the-art results on small datasets.

Methodology

OASIS streamlines the atlas-based pipeline through two primary innovations:

1. One-Pass Label-Constrained Registration

Instead of slow, iterative alignment, OASIS utilizes a Spatial Transform Model (STM) based on the VoxelMorph architecture. This allows for rapid, "one-pass" registration. A critical feature of this model is the inclusion of a label constraint during training. By incorporating a Dice loss ($\\mathcal{L}_{dice}$) between the warped atlas label and the target label, the model learns to prioritize the alignment of the specific organ of interest over irrelevant background structures.

The OASIS Spatial Transform Model (STM) architecture, highlighting the integration of image similarity, field smoothness, and label-based Dice losses.

2. Weighted Label Fusion via LOA

Standard label fusion techniques often weight atlas contributions based on global image similarity, which can be misleading in noisy medical scans. OASIS introduces Label Overlapping Accuracy (LOA), a strategy that calculates weights based on the similarity of labels within the region of interest (ROI). This ensures that atlases with the most accurate local alignment have the greatest influence on the final segmentation.

Experimental Evaluation

The framework was evaluated using the PROMISE12 dataset, which contains 50 transversal T2-weighted prostate MR images. This dataset is particularly challenging due to variations in image resolution and acquisition protocols across different vendors. OASIS was compared against several baselines, including 2D and 3D CNNs (ConvNet2D/3D) and standard atlas-based variations.

Results and Discussion

OASIS demonstrated superior performance across multiple metrics, particularly in the challenging "apex" and "base" regions of the prostate.

Quantitative comparison of OASIS against baseline methods across different prostate regions (Apex, Base, Whole).

Key findings include:

  • High Accuracy: OASIS achieved a Dice Similarity Coefficient (DSC) of 92.81% for the whole prostate, significantly outperforming the 3D ConvNet (89.43%) and the 2D ConvNet (71.03%).
  • Regional Robustness: The label-constrained approach provided a 4.1% performance boost in the difficult-to-segment apex and base regions compared to the improvement seen in the whole organ.
  • Efficiency: By leveraging the STM for registration, the framework avoids the computational bottlenecks of classical iterative methods, making it suitable for real-time clinical applications.

Conclusion

OASIS proves that hybridizing "old school" atlas concepts with modern deep learning can overcome the data limitations of pure CNN approaches. By focusing registration on the region of interest and employing a label-aware fusion strategy, OASIS provides a robust and efficient solution for medical image segmentation in small-data regimes. Future work will explore extending this framework to other organs and imaging modalities such as CT and Ultrasound.

Summary updated

Abstract

Medical image segmentation is a fundamental task in medical image analysis. Despite that deep convolutional neural networks have gained stellar performance in this challenging task, they typically rely on large labeled datasets, which have limited their extension to customized applications. By revisiting the superiority of atlas based segmentation methods, we present a new framework of One-pass aligned Atlas Set for Images Segmentation (OASIS). To address the problem of time-consuming iterative image registration used for atlas warping, the proposed method takes advantage of the power of deep learning to achieve one-pass image registration. In addition, by applying label constraint, OASIS also makes the registration process to be focused on the regions to be segmented for improving the performance of segmentation. Furthermore, instead of using image based similarity for label fusion, which can be distracted by the large background areas, we propose a novel strategy to compute the label similarity based weights for label fusion. Our experimental results on the challenging task of prostate MR image segmentation demonstrate that OASIS is able to significantly increase the segmentation performance compared to other state-of-the-art methods.
BibTeX
@misc{zhu2019oasis,
  author = {Qikui Zhu and Bo Du and Pingkun Yan},
  title = {{OASIS: One-pass aligned Atlas Set for Image Segmentation}},
  year = {2019},
  eprint = {1912.02417},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.02417},
}
Figure