Estimation of the Principal Ischaemic Stroke Growth Directions for Predicting Tissue Outcomes

Convolutional neural networks (CNN) have been widely used for the medical image analysis of brain lesions. The estimates of traditional segmentation networks for the prediction of the follow-up tissue outcome in strokes are, however, not yet accurate enough or capable of properly modeling the growth mechanisms of ischaemic stroke. In our previous shape space interpolation approach, the prediction of the follow-up lesion shape has been bounded using core and penumbra segmentation estimates as priors. One of the challenges is to define well-suited growth constraints, as the transition from one to another shape may still result in a very unrealistic spatial evolution of the stroke. In this work, we address this shortcoming by explicitly incorporating vector fields for the spatial growth of the infarcted area. Since the anatomy of the cerebrovascular system defines the blood flow along brain arteries, we hypothesise that we can reasonably regularise the direction and strength of growth using a lesion deformation model. We show that a Principal Component Analysis (PCA) model computed from the diffeomorphic displacements between a core lesion approximation and the entire tissue-at-risk can be used to estimate follow-up lesions (0.74 F1 score) for a well-defined growth problem with accurate input data better than with the shape model (0.62 F1 score) by predicting the PCA coefficients through a CNN.