ICON: Learning Regular Maps Through Inverse Consistency.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Learning maps between data samples is crucial for applications like representation learning and spatial deformation estimation.
• Spatial transformations require regularity for well-posedness, traditionally achieved through explicit regularizers in optimization models.
• Recent deep learning methods explore population-based learning to avoid explicit regularizers.

Purpose of the Study:

• To investigate if spatial regularity can be achieved using solely an inverse consistency loss in deep learning models.
• To understand the factors contributing to map regularity in such contexts.
• To evaluate the performance of this simplified approach in image registration.

Main Methods:

• Utilized deep neural networks combined with an inverse consistency loss.
• Incorporated randomized off-grid interpolation to enhance spatial transformation properties.
• Tested the approach on both synthetic and real-world image datasets.

Main Results:

• Deep networks with inverse consistency loss and randomized interpolation produced well-behaved, approximately diffeomorphic spatial transformations.
• The proposed method achieved competitive registration performance.
• Demonstrated that regular maps can be obtained without meticulously tuned explicit regularizers.

Conclusions:

• Inverse consistency loss, when combined with deep networks and randomized interpolation, is sufficient for inducing spatial regularity.
• This simplified approach offers a viable alternative to traditional regularization methods in image registration.
• The findings support the potential of data-driven methods for learning regular spatial transformations efficiently.