Generalised 3D Medical Image Registration with Learned Shape Encodings

Due to the high variability in medical images, task-specific solutions have often prevailed in the field of Deep Learning Image Registration (DLIR). Contrary to classical approaches, these are hardly transferable to other tasks or datasets. To overcome these limitations we propose to exploit easy-to-acquire anatomical segmentations, which provide strong semantic features as a common representation across different datasets and image modalities. We, therefore, present a new module for medical image registration based on generalised label encodings that can be applied to arbitrary label maps. We demonstrate the generalisability of our approach by employing the module trained on skeletal and pulmonary vascular structures to the registration of abdominal segmentations, yielding an average increase in Dice scores of 24% and compare their encoding qualities to classical Euclidean Distance Maps. Furthermore, we leverage the modular design to assess the generalisability of convolutional and vision-transformer based registration architectures. Source code and trained models are released on GitHub.