A Cooperative Autoencoder for Population-Based Regularization of CNN Image Registration

Riddhish Bhalodia, Shireen Y. Elhabian, Ladislav Kavan, Ross T. WhitakerMICCAI 2019 · OpenReview

TL;DR

This paper introduces a novel "cooperative autoencoder" (CAE) architecture for unsupervised medical image registration. By learning population-level statistics of transformations, the CAE acts as a regularizer for a primary CNN registration network, leading to more anatomically meaningful and statistically compact deformation fields compared to existing methods, while maintaining computational efficiency.

What Problem Does This Paper Address?

Aligning medical images (registration) is crucial for tasks like comparing anatomical structures across a population or tracking changes over time. Traditional methods often rely on generic smoothness penalties for deformation fields, which can overlook anatomical variability within a population. While deep learning (CNNs) has accelerated registration, these networks still typically use similar smoothness regularizers and don't inherently capture population-specific transformation patterns. This paper aims to bridge this gap by developing a CNN-based registration method that explicitly learns and leverages population statistics to enforce anatomically plausible transformations.

Methods

The core innovation is a "cooperative network" architecture comprising two subnetworks:

  • Primary Network: A U-Net-based CNN that learns to predict a displacement field to align a source image to a target image. This is a standard approach for CNN-based registration.
  • Secondary Network (Cooperative Autoencoder - CAE): This is a bottleneck autoencoder that takes the displacement fields generated by the primary network and attempts to reconstruct them. The bottleneck layer learns a low-dimensional representation (manifold) of the population of transformations.

The CAE acts as a regularizer by adding its reconstruction loss to the primary network's registration loss. This encourages the primary network to produce displacement fields that can be accurately represented by the CAE's learned low-dimensional manifold, effectively enforcing population-level consistency. The training involves an initial phase with only smoothness regularization, followed by a cooperative phase where both networks are trained together.

Datasets

The paper evaluates their method on:

  • Synthetic Datasets:
    • 2D "linear box-bump" images: Representing a single linear mode of shape variation.
  • 2D "rotating box-bump" images: Representing a single rotational mode of shape variation.
  • Medical Datasets:
    • Corpus Callosum (CC): 324 mid-sagittal 2D slices from the OASIS Brains dataset.
  • Left Atrium Appendage (LAA): 3D dataset represented as signed distance transforms.

Key Insights and Ideas

  • Population Statistics as a Regularizer: The paper demonstrates that explicitly learning and enforcing population-level statistics of transformations is more effective for anatomical correspondence than generic smoothness penalties.
  • CAE for Manifold Learning: The bottleneck autoencoder effectively learns a low-dimensional manifold that captures the essential variations within the population of transformations, acting as a powerful, data-driven regularizer.
  • Cooperative Learning: The interaction between the primary registration network and the secondary CAE network allows for a more nuanced and anatomically relevant regularization.
  • Unsupervised Advantage: The method is unsupervised, meaning it doesn't require pre-aligned image pairs for training, which is a significant advantage in medical imaging where labeled data is scarce.

Key Results

  • Comparable Accuracy, Better Correspondence: The proposed method achieves registration accuracy (measured by Dice coefficients) comparable to unsupervised deformable registration (UnDR) methods, but significantly improves anatomical correspondence, as evidenced by lower landmark errors. For instance, on synthetic datasets, landmark errors were substantially lower with the CAE approach.
  • Efficient Registration: The CNN-based approach maintains computational efficiency, with test-time runtimes similar to other CNN-based methods and much faster than conventional optimization-based methods.
  • Discovering Shape Variation: For synthetic datasets, the CAE successfully captures the dominant modes of shape variation (linear or rotational), whereas UnDR tends to compress/expand shapes without discovering these underlying population characteristics.
  • Medical Data Performance: The method shows promising results on medical datasets like Corpus Callosum and Left Atrium Appendage, demonstrating its applicability beyond synthetic examples.

Limitations and Future Work

  • Initialization Sensitivity: While the authors state that the initialization phase (with smoothness only) doesn't significantly affect results, the exact impact and optimal duration might warrant further investigation.
  • Bottleneck Dimension Selection: For medical datasets where intrinsic dimensionality is unknown (e.g., Corpus Callosum), selecting the optimal bottleneck dimension for the CAE requires experimentation. The paper shows that increasing the bottleneck can improve Dice coefficients but may lead to worse landmark errors, suggesting a trade-off and the potential for overfitting.
  • Future Work: The authors suggest that this cooperative network framework is applicable to other training strategies and representations (e.g., momentum fields). Further exploration of different network architectures and regularization techniques within this cooperative framework could yield even better results.

Summary updated

Abstract

Spatial transformations are enablers in a variety of medical image analysis applications that entail aligning images to a common coordinate systems. Population analysis of such transformations is expected to capture the underlying image and shape variations, and hence these transformations are required to produce anatomically feasible correspondences. This is usually enforced through some smoothness-based generic regularization on deformation field. Alternatively, population-based regularization has been shown to produce anatomically accurate correspondences in cases where anatomically unaware (i.e., data independent) fail. Recently, deep networks have been for unsupervised image registration, these methods are computationally faster and maintains the accuracy of state of the art methods. However, these networks use smoothness penalty on deformation fields and ignores population-level statistics of the transformations. We propose a novel neural network architecture that simultaneously learns and uses the population-level statistics of the spatial transformations to regularize the neural networks for unsupervised image registration. This regularization is in the form of a bottleneck autoencoder, which encodes the population level information of the deformation fields in a low-dimensional manifold. The proposed architecture produces deformation fields that describe the population-level features and associated correspondences in an anatomically relevant manner and are statistically compact relative to the state-of-the-art approaches while maintaining computational efficiency. We demonstrate the efficacy of the proposed architecture on synthetic data sets, as well as 2D and 3D medical data.
BibTeX
@inproceedings{bhalodia2019cooperative,
  author = {Riddhish Bhalodia and Shireen Y. Elhabian and Ladislav Kavan and Ross T. Whitaker},
  title = {{A Cooperative Autoencoder for Population-Based Regularization of CNN Image Registration}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  url = {https://openreview.net/forum?id=IYnXfpYMyMH},
}
Figure