Unsupervised Hippocampus Segmentation: Translating Deep Learning Models from Research to Clinic

1994 

Abstract Submission 

Hoda Kalabizadeh¹, Ludovica Griffanti², Natalie Voets³, Grace Gillis², Clare Mackay², Ana Namburete¹, Nicola Dinsdale¹

¹Department of Computer Science, University of Oxford, UK, ²Department of Psychiatry, University of Oxford, UK, ³Nuffield Department of Clinical Neurosciences, University of Oxford, UK

Hoda Kalabizadeh  
Department of Computer Science, University of Oxford
UK

Ludovica Griffanti  
Department of Psychiatry, University of Oxford
UK

Natalie Voets  
Nuffield Department of Clinical Neurosciences, University of Oxford
UK

Grace Gillis  
Department of Psychiatry, University of Oxford
UK

Clare Mackay  
Department of Psychiatry, University of Oxford
UK

Ana Namburete, Professor  
Department of Computer Science, University of Oxford
UK

Nicola Dinsdale  
Department of Computer Science, University of Oxford
UK

Hippocampus volume (HV), typically measured via MRI segmentation, is a well-established biomarker for Alzheimer's disease (AD) [1]. Most automated segmentation methods have been developed and validated for research datasets. However, domain shifts [2] between clinical and research datasets often significantly degrade model performance. Due to differences in imaging acquisitions (scanner shift) or disease severity (population shift), current segmentation methods are often unsuitable for clinical populations. We investigated the performance of popular segmentation methods on a clinical AD dataset. Fig 1a shows a schematic of the current data setting.

For the clinical dataset we used a sample of 29 patients from the Oxford Brain Health Clinic (BHC) [3] with labels only used for evaluation. The HarP [4] dataset was used as a research dataset, consisting of 133 labelled MRI volumes. We explored the impact of increasing the number of inputs by splitting images into hemispheres, and also of mirroring each hemisphere. Both datasets included cognitively unimpaired individuals, Mild Cognitive Impairment (MCI) patients and dementia patients. Fig 1b compares the true left HV between HarP and BHC populations.
We tested several publicly available out-of-the-box (OOB) tools for automatic segmentation: FSL FIRST [5], FreeSurfer [6] and SynthSeg [7] on the clinical dataset. For a deep learning (DL) baseline, we used a UNet [8] trained on the HarP data as our labelled reference. To investigate generalisation to the unlabelled clinical data, we expanded the UNet with data augmentation and unsupervised domain adaptation (UDA) approaches. Augmentation is an effective technique for improving a model generalisability [9] by increasing robustness to likely data variations. We explored both standard (affine transforms, flips, noise, intensity changes) and MRI-specific augmentation (motion, bias field). We implemented a UDA model [10] by including a domain classifier to enable adversarial training, which aims to do the main task while learning domain invariant features. The domain classifier does not require segmentation masks and was trained using unlabelled clinical data.

Fig 1c shows the Dice scores (DSC) for the segmentation methods, tested on the HarP and BHC datasets. The increase in training data size improved most models. OOB models, which are commonly trained and/or validated against research populations, all struggled with our clinical population. As shown on the violin plot in Fig 1d, FreeSurfer and FIRST, had instances of failing (DSC= 0). The UNet models with augmentations improved performance compared to OOB methods. However, the performance on BHC data was worse than on HarP data, with some particularly low-performing outliers. UDA was comparable to the augmentation methods for most participants, however, the UDA method led to particularly low DSC for certain individuals, possibly due to correlation between disease state and scanner leading to the removal of important information during DA. The UNet with basic augmentations was the best-performing model and used for further analysis. Fig 1e shows a positive correlation (r=0.70, p=1.31×10^-9) between true HV and achieved DSC, showing that larger volumes had higher DSC. This may be expected as individuals in HarP are on average younger and healthier than BHC, thus larger BHC hippocampi are more similar to those in the HarP training data. Performance was worse for MCI/dementia patients. Fig 2 shows the brains corresponding to (a)highest and (b)lowest DSC; brains in the latter have visibly larger ventricular atrophy.

Our findings highlighted that domain shifts between research and clinical data extend beyond acquisition differences. Despite model generalisation or UDA techniques, population shifts due to disease severity and extent of atrophy may also cause challenges for translating research tools to clinical practice and should be considered in future model development.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Segmentation and Parcellation ¹

Neuroanatomy Other

Anatomical MRI

Aging

Degenerative Disease

Machine Learning

MRI

Segmentation

STRUCTURAL MRI