Object-level analysis of changes in biomedical image sequences

In many applications, biomedical and otherwise, voluminous data sets must be analyzed objectively, quantitatively, and intelligently to detect important changes at short and long time scales, while rejecting uninteresting/nuisance changes, such as illumination fluctuations and imaging system artifacts. Another compelling need is to generate concise and high-level descriptions of the image content. In this thesis, we present a broadly-applicable framework for automated, high-level description of changes in image sequences, going significantly beyond pixel-level change detection. In this framework, changes are expressed in terms of known objects and their behaviors, as would be produced by a domain expert analyzing the same data. This is accomplished by formulating change as a statistical model selection problem in a modular framework integrating (1) a library of behavioral models, (2) algorithms for automated segmentation of key biological structures, (3) illumination-insensitive change detection, and (4) image registration. We built instances of the framework for two different real-world applications. The first example is the analysis of time-lapse image sequences of cultured neurons to detect short-term events such as neurite growth, shrinkage, merging, and splitting, as well as long-term events such as axonal specification and apoptosis. The second example is the analysis of time-lapse image sequences of cultured neuronal stem cells to track multiple cells as they move, deform, and divide, and to reconstruct lineage trees relating every derived cell to its progenitor. We developed a novel curve-distance measure, the integral area distance, to assess similarity between curve segments in a way that enables us to find the most likely changes between two curves. In addition, we developed a novel multi-object assignment algorithm which was necessary to track multiple cells as they move, deform, and divide. When tested on 8 representative recordings of live cultured neurons 60--200 frames long, the framework achieved labeling accuracies ranging from 85--100% of all the short-term changes, and correctly labeled all the long-term ones. When tested on 5 representative recordings of live cultured neuronal stem cells 750--2100 frames long, all lineage trees were reconstructed with no errors. The testing results were compared to hand-labeled ground truth for validation.