Finding dots in microscopic images
Extracting and counting numerous ‘dots’, i.e. small round regions, in large microscopic images is encountered in a wide range of medical and scientific research, from studying human cells for cancer prognosis to counting silicon wafer defects for solar cells improvement. Extracting dots is a challenging segmentation problem: faint boundaries and low contrast between regions, large intensity variations within the regions and conjoined clusters of dots make some dots hard to tease apart even upon close inspection. Background clutter or a variety of different shapes in the background can increase the challenge even further.
This thesis presents a constrained spectral graph partitioning framework to deal with the fine granularity of these small structures together with the ongoing challenge of dealing with the complexity associated with increasing image sizes. The segmentation of the entire image is obtained from a set of patch segmentations which are independently derived but subject to stitching constraints between neighboring patches. The constraints come from mutual agreement analysis on patch segmentations from a previous round.
For each individual segmentation, we introduce our ‘Finding Dots’ model to popout dots simultaneously as many disconnected components of one common foreground. We note that may applications do not require precise segmentation and model this by viewing dot boundaries as flexible regions of their own. By distancing ourselves from all traditional image segmentation methods that emphasize precision of boundary locations and shapes, we obtain a solution that is paradoxically closer to the desired segmentation. The features we use are a pixel-centric relational representation that encode local geometry. We introduce two types of grouping cues: short-range attraction based on feature similarity and long-range repulsion based on feature dissimilarity. Repulsion is at the basis of the dots popout: it plays an active and complementary role to local attraction as it operates at a different (larger) spatial range. Our work is in fact the first successful application of attraction and repulsion to real segmentation problems.
Finally, we exploit the complementary information given by region segmentation and contour grouping to present another way to incorporate local shape information to segment objects with faint boundaries along regions of low contrast. The information of the most salient region segments is combined together with the edge map obtained from the responses of an oriented filter bank. This enables us to define a new contour flow on the graph nodes, which captures region membership and enhances the flow in the low contrast or cluttered regions. The graph setup and our proposed region based normalization give rise to a random walk that allows bifurcations at junctions arising between region boundaries and favors long closed contours. Junctions become key routing points and the resulting contours enclose globally significant regions.
0984: Computer science