How to use level set methods to accurately find boundaries of cells in biomedical images ? Evaluation of six methods paired with automated and crowdsourced initial contours

Level set methods are popular tools for automatically collecting accurate outlines of biological objects in biomedical images and videos. The two main challenges with successfully applying these methods are identifying which among the many options will work well for a particular image set and choosing an initial contour that the method will successfully evolve to the desired final boundary. Little is known about the comparative performance resulting from different initial-contour method pairings. To examine the practical impact of this concern for biomedical applications, we compared six freely available level set methods with 12 different initializations on fluorescence and phase contrast images showing cells. The studies revealed that none of the initial-contour method pairings performed well for phase contrast images. These results motivated us to suggest using internet workers to draw estimates of cell boundaries. These crowdsourced boundaries can then serve as initial boundaries for a level set method to produce results closer to the true boundaries. We found that pairing segmentation algorithms with crowdsourced initial-contours yields over 50 percent points better performance than the other pairings for phase contrast images. Our results yield recommendations for initial-contour method pairings based on image modality and highlight the benefit of engaging non-expert internet workers to successfully leverage level set methods for biomedical images. We invite extensions of this work by sharing all code.