A common issue in (multi-modal) image registration is its initialization: The 3D ultrasound scan is arbitrarily oriented in space without any reference to a typical global reference system. Since intensity-based registration algorithms have a limited capture range, a close initialization is important. We have developed two independent approaches to achieve this global initialization.
On the one hand, we have trained a network to regress the absolute orientation of ultrasound as well as a confidence for each prediction. By selecting most confident frames, and a segmentation-based center-of-gravity alignment of a given organ, the registration can be initialized.
Registration result of liver MRI (greyscale) and 3D ultrasound (red) after regression of the best orientation.
On the other end, we could show that by learning dense keypoint descriptors it becomes possible to directly estimate the registration using matching pairs and RANSAC.
Overview of the proposed method. The images are fed to two separate feature extraction networks outputting dense feature maps, which are combined in a matching module, all end-to-end trainable. The matches are extracted via a confidence threshold and processed with RANSAC to retrieve the pose.
Related Publications
2022
Orientation Estimation of Abdominal Ultrasound Images with Multi-Hypotheses Networks
Timo Horstmann, Oliver Zettinig, Wolfgang Wein, and Raphael Prevost
In Proceedings of The 5th International Conference on Medical Imaging with Deep Learning, Jul 2022
Ultrasound imaging can provide valuable information to clinicians during interventions, in particular when fused with other modalities. Multi-modal image registration algorithms however require a somewhat accurate initialization, which is particularly difficult to estimate for ultrasound images as their orientation is arbitrary and their content ambiguous (limited field of view, artifacts, etc.). In this work, we not only train neural networks to predict the absolute orientation of ultrasound frames, but also to produce a confidence for each prediction. This allows us to select only the most confident frames in the clip. Our networks are trained to produce multiple hypotheses using a simple yet overlooked meta-loss that is specifically designed to capture the ambiguity of the input data. We show on several abdominal ultrasound datasets that multi-hypotheses networks provide better uncertainty estimates than Monte-Carlo dropout while being more efficient than network ensembling. Generic, easy to implement and able to quantify both data ambiguity and out-of-distribution samples, they represent a preferable alternative to traditional baselines for uncertainty estimation. On a clinical test our method produces estimates within 20^∘ of the true orientation, which we can use to improve the accuracy of a subsequent registration algorithm down to less than 10^∘.
Global Multi-modal 2D/3D Registration via Local Descriptors Learning
Viktoria Markova, Matteo Ronchetti, Wolfgang Wein, Oliver Zettinig, and Raphael Prevost
In Medical Image Computing and Computer Assisted Intervention – MICCAI 2022, Sep 2022