Redesigning weakly supervised localization architectures for medical images

Abstract

Point localization is a task that aims to identify the location but not the extent of an object of interest in an input image. A success localization is when a model predicts a point which is within the boundary of an object of interest. This task has been successfully learnt using only image-level annotation i.e. image classification. This setting is known as weakly-supervised point localization. Weakly-supervised point localization is particularly useful in medical imaging task where not only the decision of a model is required by also the explanation. The better the explanation the more useful the model is. Usually done with class-activation map (CAM), the precision of the localization is heavily constrained by the resolution of the feature map output from a model. In this thesis, we explore approaches to scale up the output feature map of a model to accommodate high-precision weakly-supervised point localization on chest x-rays.