A Robust Pipeline For Personalised Localisation of Brain Regions Using Image Quality Transfer

1400 

Abstract Submission 

Ying-Qiu Zheng¹, Harith Akram², Stephen Smith¹, Saad Jbabdi³

¹University of Oxford, Oxford, Oxfordshire, ²UCL Queen Square Institute of Neurology, London, UK, ³Oxford University, Oxford, United Kingdom

Ying-Qiu Zheng  
University of Oxford
Oxford, Oxfordshire

Harith Akram, MD  
UCL Queen Square Institute of Neurology
London, UK

Stephen Smith  
University of Oxford
Oxford, Oxfordshire

Saad Jbabdi  
Oxford University
Oxford, United Kingdom

Accurate identification and localisation of brain regions in individual subjects is increasingly important in both neuroscience research and clinical practice. In functional neurosurgery, e.g. deep brain stimulation, targeting precision is essential for symptom relief in various neurological and psychiatric disorders. However, this task is hampered by the lack of distinct contrast for many target brain structures on conventional medical images and significant variability in individual brain anatomies. A common practice is to rely on standardised atlases, but this often fails to account for inter-individual variations [1-3].

Advancements in neuroimaging have enabled brain region localisation in individual brains. For instance, a popular approach is the use of connectivity-based functional localisation with tractography or resting-state connectivity, alongside prior knowledge of a region's connectional fingerprint. This however requires high-quality diffusion/functional MRI data, limiting application in clinical settings with scan-time constraints.

To address this challenge, we developed LOCALISE, a tool that enables individualised delineation of brain regions. The main purpose of this tool is to enable connectivity-based localisation in low-quality datasets. Using Image Quality Transfer (IQT) techniques [9-10], LOCALISE transfers anatomical information from large-scale high-quality MRI (e.g. HCP data) enabling functional localisation on clinical-quality data. Here, we focus on the ventral intermediate nucleus of the thalamus (Vim), a key DBS target for the treatment of tremor in Parkinson's patients, to demonstrate LOCALISE's capabilities, with ongoing work to incorporate more brain structures.

Localising Vim illustrates the idea behind LOCALISE. In theory, Vim can be found using literature-based connectional information, namely the connections to M1 and the contralateral Cerebellum. However, these connections cannot be robustly extracted. Instead, LOCALISE creates a large number of more "robust" connectional features, and learns to segment Vim using these features and the high-quality prior-based segmentation as a target "ground truth". This mapping is then used on low-quality data where the direct prior-based method fails. To train LOCALISE for Vim, we used high-quality diffusion MRI data from the Human Connectome Project (HCP) to generate Vim locations and trained an IQT model [10] to locate Vim in surrogate low-quality datasets based on their connectivity profiles.

The segmentation model employs a Conditional Random Field (CRF) [11] to map high-quality Vim labels to low-quality features. This CRF model is optimised to maximise the likelihood of correct HQ-Vim label assignments based on voxel-wise connectivity profiles. The trained model was then applied to left-out low-quality subjects and evaluated against their high-quality counterparts.

Using Vim localisation as an example, LOCALISE's accuracy was evaluated using the Dice coefficient and centroid displacement, the latter a measure of distance between the centroids of predicted and actual clusters. LOCALISE outperformed two established methods, the (original, simple) connectivity-driven and atlas-based approach, on HCP surrogate low-quality datasets, showing higher overlap with high-quality "ground truth" and greater reliability across varying data quality and scanning sessions. Notably, when applied to UK Biobank low-quality datasets, LOCALISE, trained on HCP data, generalised effectively, surpassing alternative methods in Vim localisation.

LOCALISE offers robust localisation of brain regions even when using clinical-standard data, and has the potential to enhance targeting accuracy in neurosurgical procedures. Its adaptability to diverse data conditions makes it a valuable tool in settings where high-quality MRI is unavailable or inappropriate for certain cohorts. The tool can be found in https://git.fmrib.ox.ac.uk/yqzheng1/python-localise.

Deep Brain Stimulation

Invasive Stimulation Methods Other

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural)

Segmentation and Parcellation ²

Design and Analysis

Machine Learning

MRI

Open-Source Software

Segmentation

Tractography

Other - Surgical Targeting