πŸ“ Paper: https://arxiv.org/pdf/1809.05231.pdf

🎯 What?

Pairwise medical image registration. Image registration, also known as image fusion or image matching, is the process of aligning two or more images based on image appearances. Medical Image Registration seeks to find an optimal spatial transformation that best aligns the underlying anatomical structures. in other words, is the process of finding is the process of finding optimal transformation that puts different images optimal transformation that puts different images into spatial correspondence.

❓ Why?

Medical Image Registration is used in many clinical applications such as image guidance, motion tracking, segmentation, dose accumulation, image reconstruction and so on.

☝🏻 Pior work and SoTa

Modelled the matching as an optimisation problem with an objective function.

(-) cons : inference take time (few minutes by using GPU)

πŸ₯Š Proposed solution

Use CNN to learn to generate the deformable field. Then applied it the the moving images.

(+) pros : Fast inference (under second using GPU), accuracy close to sota

βš™οΈ Architecture

The deformation field is generate by U-net architecture

model

The loss can be decomposed in two component : A image level similarly measure and the deformation field smoothness measure.

There two possible similarity measure : MSE and Cross-correlation

🦾 Results

Metrics : The evaluation metric is the Dice score on anatomical segmentation. This measure in a certain way measure the volume overlap.

Experimental setup : They did an atlas based registration. This means that we take the atlas model of the brain as reference and we try to transform our input to that reference.

They use an atlas computed using an external dataset . Each input volume pair consists of the atlas (image f) and a volume from the dataset (image m). image pairs using the same fixed atlas for all examples.