Individualised Clinical Neuroimaging in the Developing Brain: Abnormality Detection

Student thesis: Doctoral ThesisDoctor of Philosophy


Perinatal neuroanatomical structure is incredibly intricate and, at time of birth, is undergoing continuous change due to interweaving developmental processes (growth, myelination and gyrification). While there is some small variability in structure and rates of development, all follow proscribed pathways with well documented milestones. Brain injury or other disruption of these processes can result in poor neurodevelopmental outcomes or mortality, making their early identification critical to estimate, and potentially forestall, negative effects. MRI is an increasingly used method of investigating suspected neonatal encephalopathies and injuries.

Identification of these injuries and malformations is more challenging in neonates compared to adults due to the brain’s continuously evolving appearance. This makes radiological review of neonatal MRI an intensive and time-consuming task which, in an ideal setting, requires a team of highly skilled clinicians and radiologists with complementary training and extensive experience. To assist this review process, some localisation method that highlights areas likely to contain tissue abnormalities would be highly desirable, as it could quickly draw attention to these locations. In addition, identifying neonates whose MRI is likely to contain some form of pathology could allow for review prioritisation.

In this thesis, I first investigated using normative models of neonatal tissue intensity for brain tissue abnormality detection. I applied voxel-wise Gaussian process (GP) regression to a training cohort of neonates with no obvious lesions, all born preterm (<37 weeks) but imaged between 28-55 weeks. Gestational age at birth (GA), postmenstrual age at scan (PMA) and sex were used as input variables and voxel intensity as the output variable. GPs output a mean value and its variance inferred from neonates within the training cohort whose demographic information most closely matched those of the prediction target. The voxel specific models were put together to form a synthesised typical image and standard deviation image derived from the variance outputs. Z-score abnormality maps were constructed by taking the difference between neonates actual MRI and GP-calculated synthetic image and scaling by their standard deviation map. Higher Z-score map values indicate voxels more likely to contain abnormal tissue intensity. Using manually delineated masks of common brain injuries seen in a subset of neonates, these abnormality Z-score maps demonstrated good detection performance using area under the curve of receiver operating characteristic scores, with the exception of small punctate lesions.

The initial voxel-wise models had substantial false positives around the edges of the brain where there is large typical heterogeneity. I next investigated if incorporating local structural information into predictive models could improve their ability to accommodate typical anatomical heterogeneity seen across individual brains and improve the accuracy of synthetic images and abnormality detection. To achieve this, voxel intensity values in a patch surrounding the prediction target were appended to the design matrix, alongside GMA, PMA and sex. The patch-based synthetic images were able to match an individual’s brain structure more closely and had lower false positives in normal appearing tissue. However, a weakness was that the centre of some larger lesions was included in the predictions (thereby classified as ‘healthy’ tissue), having a deleterious effect on their coverage, increasing false negatives. This was offset by much better coverage of smaller, more subtle lesions, to the extent that overall performance was higher compared to that seen in the earlier model.

I also investigated if the Z-score abnormality maps could be used to classify neonates with MRI positive brain injury from those with normal appearing brains.
While many machine learning algorism see frequent use in neuroimaging classification tasks, I opted for a logistic regression model due to its high levels of interpretability and simple implementation. Using the histograms of the Z-score abnormality maps as inputs, the model demonstrated good performance, being able to correctly identify neonates with injuries, but not those with subtle lesions like punctate lesions, whilst minimising false identification of neonates with normal appearing brains.

To ascertain if performance could be improved, I explored multiple classification methods. Specifically, the use of other more complex classifiers (random forest, support vector machines, GP classification) and the use of a regional abnormal voxel count, that allowed localisation of lesioned tissue rather than the more global detection ability of the Z-score histograms.

Using these innovations, I investigated their application towards a specific pathology; hypoxic ischemic encephalopathy (HIE). This is a good test for the system, as HIE has high incidence rates, multiple associated lesion types and a time dependant appearance. Further, I wanted to know if, given a positive HIE diagnosis, the Z-score abnormality maps could be used to predict long-term outcomes (normal vs poor). Several models demonstrated an excellent ability to separate HIE and healthy control neonates achieving >90% accuracy, a statistically significant result even after false discovery rate (FDR) correction (p-value < 0.05). While the outcome prediction models achieved reasonable accuracy, >70% in multiple models, none of these were statistically significant after FDR correction.

Overall, this work demonstrates how normative modelling can be used to create individual voxel-wise / image-wise estimation of tissue abnormality for neonatal MRI across a range of gestational ages. It further demonstrates that these abnormality maps can be utilised for additional tasks, in this instance, three increasingly challenging neurological classification problems. These include the separation of neonates with and without MRI positive lesions, identification of neonates with a specific pathological condition (HIE) and prediction of long-term functional outcome (normal vs poor). Within a radiological setting, these classifications task can be considered analogous to three radiological challenges, image triage, diagnostic detection and estimation of developmental prognosis, important for the clinical team but also infants and their families.
Date of Award1 Feb 2024
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorJonathan O'Muircheartaigh (Supervisor), Serena Counsell (Supervisor) & David Carmichael (Supervisor)

Cite this