Automated image-based registration of ultrasound and magnetic resonance imaging for endoscopic endonasal surgery

Project summary:

Pituitary tumours are common, affecting 15–20% of the population. Surgery is typically performed through the nose using RGB cameras and minimally invasive instruments. Despite advances in surgical techniques, some pituitary adenomas are challenging to cure. The use of intraoperative ultrasound during pituitary surgery could significantly improve surgeons’ ability to visualise tumour tissue and other critical surrounding structures however its routine clinical adoption is limited by the fact that ultrasound images are acquired in unfamiliar planes which can make recognizing landmarks challenging. We aim to develop robust intraoperative ultrasound-MR image-based registration methods for pituitary imaging. Image-based registration must be able to account for different in appearance and field of view between the two modalities in addition to the fact that tissues may be removed or deform during the course of the procedure. Registration  will facilitate the use of intraoperative pituitary ultrasound, improving safe surgical resection and patient outcomes. 

 

Project description:

Background 

Pituitary adenomas affect 15–20% of the population. In the UK over 1200 pituitary operations are performed every year with the transsphenoidal approach (via the nasal cavity) being the choice for resecting most pituitary adenomas. However, despite advances in surgical technique, some pituitary tumours remain difficult to cure and a third of patients undergoing transsphenoidal surgery have an incomplete resection. 

 

Neuronavigation may be used to help guide pituitary surgery but intra-operative tissue deformation can cause brain shift, moving surgical targets and other vital structures, thus invalidating the pre-surgical plan. Intraoperative ultrasound (iUS) has become an increasingly popular tool in neurosurgery to provide a relatively inexpensive and simple method of real-time feedback. However due to a limited field-of-view and unfamiliar imaging planes recognizing anatomical landmarks is challenging. Image registration techniques that can update pre-surgical MRI based on iUS are crucial to ensuring the surgical plan can be achieved, but accurate image registration remains an open challenge.  

​

Previous work including the MICCAI Challenge 2018 for Correction of Brain shift with Intra-Operative UltraSound (CuRIOUS2018) and Discrete Unsupervised 3D Registration Methods for the Learn2Reg Challenge using the REtroSpective Evaluation of Cerebral Tumors (RESECT) database. Recent image registration methods have reported promising results including using a multi-modal similarity metric denoted linear correlation of linear combinations (LC2), the use of a probabilistic dense displacement net (PDD-net), or hybrid approaches that combine image intensity and segmentation. Further improvements are required before US-MRI registration becomes routine in clinical practice. The generalisability of previously-developed methods are yet to evaluated on pituitary ultrasound data.  

 

This project will leverage a dataset to be collected as part of a clinical trial evaluating the use of a novel endonasal pituitary ultrasound probe. This will form the largest dataset of pituitary ultrasound images and will be made publicly-available at the conclusion of this project. 

 

Aim 

This project will explore novel methods of ultrasound-MR image-based registration developed on a large dataset of pituitary ultrasound and MRI images. 

The project will pursue the following aims: 

1. Develop an image-based metric to determine correspondence between MRI and iUS.  

1. Fine-tune algorithm for pituitary imaging and evaluate registration on alignment of clinically important structures.  

1. Create a high-fidelity phantom model for acquisition of additional training data 

1. Disseminate the study’s outcomes and the device’s potential to patients and the public 

​

Methodology 

Y1: Early algorithmic development will be performed on publicly-available US-MRI datasets such as the RESECT database. The student will investigate deep learning approaches that can determine good and poor correspondence between MRI and iUS images of the brain with varying appearances (A1). A key focus will be on data augmentation to synthetically create images with different appearances and artifacts (e.g. shatter, shadows). 

 

Y2 and Y3: Techniques developed in Y1 will be adapted to an endoscopic transsphenoidal dataset (A2) and data acquired from a phantom model (A3). Evaluation will incorporate alignment of clinically important structures to pituitary surgery. 

 

Research results will be disseminated through publications in high-impact journals, presentations at leading conferences and public engagement events (A4). 

​