Novelty, distillation, and federation in machine learning for medical imaging
MetadataShow full item record
The practical application of deep learning methods in the medical domain has many challenges. Pathologies are diverse and very few examples may be available for rare cases. Where data is collected it may lie in multiple institutions and cannot be pooled for practical and ethical reasons. Deep learning is powerful for image segmentation problems but ultimately its output must be interpretable at the patient level. Although clearly not an exhaustive list, these are the three problems tackled in this thesis. To address the rarity of pathology I investigate novelty detection algorithms to find outliers from normal anatomy. The problem is structured as first finding a low-dimension embedding and then detecting outliers in that embedding space. I evaluate for speed and accuracy several unsupervised embedding and outlier detection methods. Data consist of Magnetic Resonance Imaging (MRI) for interstitial lung disease for which healthy and pathological patches are available; only the healthy patches are used in model training. I then explore the clinical interpretability of a model output. I take related work by the Canon team — a model providing voxel-level detection of acute ischemic stroke signs — and deliver the Alberta Stroke Programme Early CT Score (ASPECTS, a measure of stroke severity). The data are acute head computed tomography volumes of suspected stroke patients. I convert from the voxel level to the brain region level and then to the patient level through a series of rules. Due to the real world clinical complexity of the problem, there are at each level — voxel, region and patient — multiple sources of “truth”; I evaluate my results appropriately against these truths. Finally, federated learning is used to train a model on data that are divided between multiple institutions. I introduce a novel evolution of this algorithm — dubbed “soft federated learning” — that avoids the central coordinating authority, and takes into account domain shift (covariate shift) and dataset size. I first demonstrate the key properties of these two algorithms on a series of MNIST (handwritten digits) toy problems. Then I apply the methods to the BraTS medical dataset, which contains MRI brain glioma scans from multiple institutions, to compare these algorithms in a realistic setting.