Figure 3: Synthetic results with different initialization methods and the robustness of the model to noise.
From: Accurate Morphology Preserving Segmentation of Overlapping Cells based on Active Contours
(a) Evaluation of segmentations with different initialization methods. “Neutral” means the initial ϕ was set to a Gaussian white noise with mean λf/αf and small variance (no initial objects are needed). “Seeded” means that small circles inside the objects were used to initialize the phase field. (b) Segmentation results with neutral initialization using 3 layers. (c) Segmentation results with seeded initialization. (d) Synthetic image with overlapping circles and increasing levels of signal-to-noise ratio (+20 dB to −5 dB). (e,f) Evaluation of the results of the proposed model. The segmentation results were evaluated on 50 test images per noise level. Note that different data term weights γd were used for different noise levels. (e) Average precision and recall values of segmentation results for different signal-to-noise ratios. (f) Box and whisker plot of segmentation accuracy for different levels of signal-to-noise ratio. The bottom and top edges of the box indicate the first and third quartiles; the line inside the box indicates the median; the whiskers (lines protruding from the box) indicate the smallest and largest data points whose distance from the box is not greater than 1.5 times the interquartile range.
