Bridging brain coordinates and machine learning for surgical targeting and morphometric mapping

6 

Abstract Submission 

Alaa Taha¹, Greydon Gilmore², Mohamad Abbass², Violet Liu³, Chris Zajner², Brendan Santyr¹, Abrar Ahmed², Ali Hadi², Sandy Wong⁴, Ali Khan⁵, Jonathan Lau⁴

¹University of Western Ontario, London, Ontario, ²Department of Clinical Neurological Sciences, Division of Neurosurgery, London, Ontario, ³Imaging Research Laboratories, Robarts Research Institute, London, Ontario, ⁴Department of Clinical Neurological Sciences, Division of Neurosurgery, London, ON, ⁵Department of Medical Biophysics, Schulich School of Medicine and Dentistry, Western University, London, Ontario

Alaa Taha  
University of Western Ontario
London, Ontario

Greydon Gilmore  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Mohamad Abbass  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Violet Liu  
Imaging Research Laboratories, Robarts Research Institute
London, Ontario

Chris Zajner  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Brendan Santyr  
University of Western Ontario
London, Ontario

Abrar Ahmed  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Ali Hadi  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, Ontario

Sandy Wong  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, ON

Ali Khan  
Department of Medical Biophysics, Schulich School of Medicine and Dentistry, Western University
London, Ontario

Jonathan Lau  
Department of Clinical Neurological Sciences, Division of Neurosurgery
London, ON

A deviation of 2 millimeters (mm) in deep brain stimulation (DBS) electrode positioning can result in variability of upwards of 60% in therapeutic benefit [1]. Suboptimal targeting may require reimplantation, which can pose additional risks. Localizing DBS targets is not always possible because of their small size, lack of contrast, and patient motion. Tools which involve non-linear alignment of an atlas to patient images, considered 'gold standard' for automatic localization, yield errors on the order of 2-3 mm [2] and highly depend on image quality (Figure 1A). Gadolinium enhanced T1w MRI (MRI-gad) is employed during DBS planning, as it helps with avoiding blood vessels. However, MRI-gad presents challenges during non-linear alignment [3]. Automatic localization of brain structures via machine learning (ML) offers faster and generalizable alternatives to registration approaches. However, limited studies cater ML to DBS targets while demonstrating generalizability in clinic (e.g. on MRI-gad) [4].
We validate an ML model (Figure 1) to localize surgical targets from the coordinates of salient brain landmarks [5] in patient space. Our approach is agnostic to field strength, generalizable to other brain regions, and enables more nuanced understanding of brain morphology that can be expressed in millimeters (Figure 2).

We curated and openly released [6] three imaging datasets: The SNSX dataset, 64 healthy and DBS patients imaged at 7-T; the LHSCPD dataset, same DBS participants as in SNSX, but imaged at 1.5-T MRI-gad; and the AFIDs dataset, multi-center, multi-resolution MRI data with curated anatomical landmarks across a diverse population (N = 169), including healthy, abnormal ventricular size, and neurodegenerative disease.
We leveraged coordinates of 32 anatomical landmarks (called AFIDs) from aforementioned datasets. AFIDs feature an inter-rater localization error of ~1 mm, validated across MRI field strengths and disease via 50,000+ Euclidean distances (ED) from 20 expert and novice human raters [5,6].
We computed the subthalamic nucleus center (STN) from segmentations on T2w 7-T MRI scans curated by 3 expert neurosurgeons and the lead author.
Coordinates of AFIDs were used as features to predict the STN center. Principal component analysis and correlations were used to evaluate relationships between AFIDs, subsequently a linear regression model was trained via nested 4-fold cross validation.
We employed an unseen paired dataset of 22 DBS patients imaged at both 7-T and 1.5-T to demonstrate our model's robustness across MRI field strength. We then applied a validated registration framework [7] to predict the STN, and statistically compared that to our model. To simulate the upper limit of mis-localization errors by trained raters, we augmented AFID locations anisotropically by 2 mm and evaluated prediction error. Finally, to evaluate the generalizability of our model to other brain regions we perform a leave-one-AFID out analysis where the excluded AFID was predicted from all other AFIDs.

STN prediction error on our test set is 1.01 ± 0.56 mm, outperforming non-linear alignment which failed for 6 patients (produced irregular wrap fields). Additionally, our model exhibited no statistical difference when predicting STN coordinates from 7T or 1.5T MRI-gad imaging. We leveraged our model to predict the STN on an MRI-gad scan from a DBS electrode re-implantation case, leading to more accurate targeting (Figure 1C). STN predictions from augmented AFID placements (2 mm) featured an error of 0.34 ± 0.12 mm. Finally, our brain target generalizability yielded an error of 1.1 ± 0.36 mm on 7 midbrain AFIDs.

We demonstrate a novel surgical targeting framework that accommodates for inter-patient variability and is agnostic to MRI field strength. We integrate this approach within the clinical workflow and showcase its utility in predicting surgical targets and informing DBS electrode re-implantation where conventional imaging is suboptimal.

Deep Brain Stimulation ¹

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)

Classification and Predictive Modeling ²

Subcortical Structures

Brain Atlases

Computational Neuroscience

Degenerative Disease

Machine Learning

Movement Disorder

Neurological

Open Data

Open-Source Code

Open-Source Software

STRUCTURAL MRI

Other - Deep Brain Stimulation (DBS)