Resumen de Efficient interaction with large medical imaging databases
Everyday, a wide quantity of hospitals and medical centers around the world are producing large amounts of imaging content to support clinical decisions, medical research, and education. With the current trend towards Evidence-based medicine, there is an increasing need of strategies that allow pathologists to properly interact with the valuable information such imaging repositories host and extract relevant content for supporting decision making. Unfortunately, current systems are very limited at providing access to content and extracting information from it because of different semantic and computational challenges. This thesis presents a whole pipeline, comprising 3 building blocks, that aims to to improve the way pathologists and systems interact. The first building block consists in an adaptable strategy oriented to ease the access and visualization of histopathology imaging content. The second block explores the extraction of relevant information from such imaging content by exploiting low- and mid-level information obtained from from morphology and architecture of cell nuclei. The third block aims to integrate high-level information from the expert in the process of identifying relevant information in the imaging content. This final block not only attempts to deal with the semantic gap but also to present an alternative to manual annotation, a time consuming and prone-to-error task. Different experiments were carried out and demonstrated that the introduced pipeline not only allows pathologist to navigate and visualize images but also to extract diagnostic and prognostic information that potentially could support clinical decisions.
