SuperCUT, an unsupervised multimodal image registration with deep learning for biomedical microscopy

Area of Science:

• * Biomedical imaging
• * Computational biology
• * Machine learning for microscopy

Background:

• * Correlative analysis of biological samples requires accurate registration of multimodal images.
• * Current registration methods are often supervised, demanding limited expert-annotated data.
• * Unsupervised approaches are needed to overcome data limitations in multimodal image registration.

Purpose of the Study:

• * To propose a general unsupervised deep learning pipeline for multimodal image registration.
• * To evaluate the pipeline's performance against state-of-the-art methods.
• * To demonstrate the efficacy of unsupervised style transfer for registration.

Main Methods:

• * Developed a deep learning pipeline for unsupervised multimodal image registration.
• * Employed style transfer techniques to align different imaging modalities.
• * Conducted comprehensive evaluations on four biological datasets using diverse microscopy techniques.

Main Results:

• * The proposed unsupervised pipeline achieved image registration accuracy comparable to supervised methods.
• * Style transfer combined with unsupervised training proved effective for multimodal registration.
• * The method successfully registered images from various microscopy modalities without human intervention.

Conclusions:

• * Unsupervised deep learning, particularly with style transfer, offers a viable solution for multimodal image registration.
• * This approach overcomes the limitations of expert-annotated data in supervised methods.
• * The pipeline provides a generalizable and efficient tool for correlative biological imaging analysis.