Preview

PDF, English Download (23MB) | Terms of use

Abstract

The fate of contemporary scientific research in biology and medicine is bound to the advancements in computational methods. The unprecedented data explosion in microscopy and the crescent interest of life scientists in studying more complex and more subtle interactions stimulate the research for innovative computational solutions on challenging real world applications. Extensions and novel formulations of generic and flexible methods based on learning/inference are necessary to cope with the large variety of the produced data and to avoid continuous reimplementation and heavy parameter tuning. This thesis exploits cutting edge machine learning methods based on structured probabilistic models and weakly supervised learning to provide four novel solutions in the areas of large-scale microscopic imaging and multiple objects tracking. Chapter 2 introduces a weakly supervised learning framework to tackle the problem of detecting defect images while mining massive microscopic imagery databases. This thesis demonstrates accurate prediction with low user annotation effort. Chapter 3 presents a learning approach for counting overlapping objects in images based on local structured predictors. This problem has numerous applications in high throughput microscopy screening such as cells counting for drug toxicity assays. Chapter 4 develops a deterministic graphical model to impose temporal consistency in objects counts when dealing with a video sequence. This Chapter shows that global (temporal and spatial) structural inference consistently improves over local (only spatial) predictions. The method developed in Chapter 4 is used in a novel downstream tracking algorithm which is introduced in Chapter 5. This Chapter tackles, for the first time, the difficult problem of tracking heavily overlapping, translucent and indistinguishable objects. The mutual occlusion event handling of such objects is formulated as a novel structured inference problem based on the minimization of a convex multi-commodity flow energy. The optimal weights of the energy terms are learned with partial user supervision using structured learning with latent variables.To support behavioral biologists, we apply this method to the problem of tracking a community of interacting Drosophila larvae.