| Summary: | <p>This thesis presents methods that address three fundamental tasks in the field of microscopy image analysis: detection of instances of an object, counting instances of an object when individual detection is not possible, and exploration of large datasets of microscopy images. We address these areas while strongly considering the constraints related to the usability and flexibility of a real practice. Through the use of computer vision and machine learning, we aim to deal with minimalistic user annotations, interactivity and intuitive data visualization.</p> <p>Firstly, a novel framework is presented to detect all the instances of an object of interest in microscopy images, where they may be partially overlapping and clustered. To this end, a tree-structured discrete graphical model is introduced, that is used to select and label a set of non-overlapping candidate regions in the microscopy image by a global optimization of a classification score. Additionally, it is shown how to learn, from the images, the generation of a surface that can improve the collection of the candidate regions. The learning for the object model and optimal surface only require simplistic annotations -- a dot on each instance of the object.</p> <p>Secondly, it is considered the scenario where individual instances of an object may not be discernible due to extreme overlap, but the count of instances over regions of interest is still useful quantitative information. A novel object density estimation approach is proposed, which is also trained from simple dot-annotations, and achieves a similar accuracy to current state-of-the-art methods while being considerably faster to train. The application to counting in time-lapse microscopy sequences is illustrated. Moreover, by taking advantage of the possibility to learn on-the-fly from simple annotations, an interactive counting system is developed which allows a user to quickly obtain object counts on simple cases, or to bootstrap dataset annotations for more complex scenarios. </p> <p>Finally, the exploration of microscopy datasets coming from large exploratory studies is considered. An unsupervised pipeline is proposed, which allows the discovery and visualization of the effect of external perturbations such as drugs over a target of interest (e.g. a cell protein). This framework enables a user to perform large scale image-dataset exploration in an intuitive way using a novel 2D visualization tool.</p>
|
|---|