Multi-Atlas Image Soft Segmentation via Computation of the Expected Label Value

Iman Aganj, Bruce FischlIEEE TMI 2021 · OpenReview

TL;DR

This paper introduces a novel method for medical image segmentation that avoids computationally expensive deformable registration. Instead, it computes the "expected label value" (ELV) for each voxel by considering all possible transformations from atlases, effectively creating a soft segmentation without relying on a single "optimal" registration. This approach offers a faster and more robust alternative to traditional multi-atlas segmentation.

What Problem Does This Paper Address?

Accurate segmentation of anatomical structures in medical images is crucial for diagnosis, treatment planning, and research. Traditional multi-atlas segmentation methods rely on deforming multiple reference images (atlases) to match a new target image. This process, known as deformable registration, is computationally intensive and prone to errors, such as getting stuck in local optima during optimization. Furthermore, even if a "correct" registration is found, it might discard valuable information from other plausible transformations. This paper aims to overcome these limitations by developing a more efficient and robust segmentation technique.

Methods

The core of the proposed method is the computation of the "Expected Label Value" (ELV). Instead of finding a single optimal transformation, the authors propose to integrate over all possible transformations, weighted by their probability. This is achieved without explicitly performing deformable registration.

  1. Key Image Creation: A "key" image is created from the training atlases. This key image effectively encodes the spatial information of the target structures from multiple atlases.
  2. ELV Computation: For a new image, the ELV map is computed through a simple convolution operation between the (phase-only) image and the pre-computed key. This convolution efficiently approximates the expected label value by considering the probability distribution of transformations.
  3. Intensity Prior (ELV+IP): To further improve segmentation, an intensity prior is incorporated. This step uses the image intensity information to modulate the ELV map, making it more sensitive to organ-specific intensity characteristics.
  4. Post-processing: The soft ELV map is then converted into a hard segmentation mask using thresholding and connected component analysis.

Datasets

The authors evaluated their method on several publicly available datasets:

  • Brain Segmentation:
    • OASIS-3 dataset (T1-weighted MR images): Used for training and generating FreeSurfer-like labels.
  • Internet Brain Segmentation Repository (IBSR) dataset.
  • A dataset from [24] (1 mm³ isotropic voxel images).
  • Liver Segmentation:
    • Liver Tumor Segmentation (LiTS) Challenge dataset (abdominal CT images).
  • Pancreas Segmentation:
    • Two CT databases from The Cancer Imaging Archive (TCIA) (NIH Clinical Center and Memorial Sloan Kettering Cancer Center).

Key Insights and Ideas

  • Shifting from "Optimal" to "Expected": The fundamental insight is that instead of relying on a single, potentially suboptimal, "best" registration, considering the weighted average of all plausible transformations provides a more robust and informative segmentation.
  • Convolution as a Registration-Free Alternative: The clever use of convolution to approximate the expected label value bypasses the need for computationally expensive deformable registration, significantly speeding up the process.
  • Phase Information for Robustness: Using only the phase information of images for similarity measurement makes the method less sensitive to intensity variations and more robust to different acquisition protocols.
  • Key Image Encapsulates Atlas Information: The "key" image acts as a powerful summary of the anatomical information across multiple atlases, simplifying the subsequent segmentation step.

Key Results

  • Brain Segmentation: The ELV+IP method achieved a median Dice score of 0.782 and a mean of 0.766 ± 0.001 on reproducing FreeSurfer labels from the OASIS-3 dataset. External evaluation on IBSR and another dataset yielded median Dice scores of 0.69 and 0.677, respectively.
  • Liver Segmentation: The ELV+IP method achieved a median Dice score of 0.92 (mean: 0.91 ± 0.01) on the LiTS dataset, outperforming traditional multi-atlas methods (ANTs) and showing comparable results to a U-Net CNN.
  • Pancreas Segmentation: Median Dice scores of 0.59 and 0.50 were achieved on the two CT datasets, respectively. The authors note that pancreas segmentation is inherently more challenging due to its anatomical variability.
  • Speed: The ELV method is significantly faster than traditional multi-atlas methods. For example, the ELV+IP method takes approximately 32.8 minutes for the entire experiment (including key creation), while ANTs methods take hours. The core ELV computation and hard segmentation take less than a minute.

Limitations and Future Work

  • Pancreas Segmentation Accuracy: The accuracy for pancreas segmentation was lower compared to brain and liver, attributed to the organ's anatomical flexibility and variability.
  • Intensity Prior Sensitivity: The performance of the intensity prior (ELV+IP) was found to be sensitive to inter-dataset variations in image acquisition protocols and intensity harmonization.
  • Future Work:
    • Augmenting convolutional neural networks (CNNs) with the ELV map as input to potentially improve segmentation accuracy.
  • Incorporating large rotations and scaling directly into the ELV computation.
  • Evaluating segmentation using metrics beyond Dice, such as region volume and surface distances.
  • Combining the ELV approach with deep learning architectures into a single network.

Summary updated

BibTeX
@article{aganj2021multi,
  author = {Iman Aganj and Bruce Fischl},
  title = {{Multi-Atlas Image Soft Segmentation via Computation of the Expected Label Value}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2021},
  url = {https://openreview.net/forum?id=ROuiAaTu7kv},
}
Figure