ILEE shows superior accuracy and stability over classic thresholding approaches. The manually portrayed binary ground truth of a diverse image set (n = 7, 400*400) was compared with corresponding binary images rendered by ILEE, MGT, four global thresholding algorithms (Otsu, Triangle, Li, Yan), and two local thresholding algorithms (Niblack and Sauvola). (a) Visualized demonstration of ILEE versus other approaches. Pixels with different colors are defined as green: match of rendered binary image and ground truth; blue: false-negative, the pixels omitted by the algorithm; red: noise-like false-positive, the pixels that are rendered by the algorithm but not in the ground truth without a shape of filament; brown: actin-like false-positive, the false-positive pixels within a filament-like component, which cannot be judged as false result at high confidence. Scale bars are marked as red solid line. (b) Quantitative comparison of pixel rendering accuracy. The seven images annotated with single-pixel ground truth are tested. ILEE has a superior accuracy and with the highest match rate and a stably low error rate; MGT and Li also have acceptable performance. Error bar = standard deviation. (c) Comparison of the distribution of distance transformation error. Single pixel errors of all the images were merged (n = ∼2.5e5) and summarized as a violin plot. Red dashed line indicates zero error, or results identical to the ground truth. ILEE has a symmetric and centralized distribution, indicating an accurate and unbiased filament segmentation. (d) Comparison of computational accuracy of cytoskeletal indices. Nine biologically interpretable cytoskeleton indices computed using the binary images rendered by different algorithms were compared with the ground truth. The index values were normalized to the fold of ground truth. Red dashed line indicates 1-fold, or identical to the ground truth. Error bar = standard deviation. Routine significance tests do not apply to b, c, and d because both mean and standard deviation serve as independent quantitative metrics of algorithm performance rather than indicators of significant differences. Related results are shown in Figs. S12, S13, S14, and S15.
Figure 4.

ILEE shows superior accuracy and stability over classic thresholding approaches. The manually portrayed binary ground truth of a diverse image set (n = 7, 400*400) was compared with corresponding binary images rendered by ILEE, MGT, four global thresholding algorithms (Otsu, Triangle, Li, Yan), and two local thresholding algorithms (Niblack and Sauvola). (a) Visualized demonstration of ILEE versus other approaches. Pixels with different colors are defined as green: match of rendered binary image and ground truth; blue: false-negative, the pixels omitted by the algorithm; red: noise-like false-positive, the pixels that are rendered by the algorithm but not in the ground truth without a shape of filament; brown: actin-like false-positive, the false-positive pixels within a filament-like component, which cannot be judged as false result at high confidence. Scale bars are marked as red solid line. (b) Quantitative comparison of pixel rendering accuracy. The seven images annotated with single-pixel ground truth are tested. ILEE has a superior accuracy and with the highest match rate and a stably low error rate; MGT and Li also have acceptable performance. Error bar = standard deviation. (c) Comparison of the distribution of distance transformation error. Single pixel errors of all the images were merged (n = ∼2.5e5) and summarized as a violin plot. Red dashed line indicates zero error, or results identical to the ground truth. ILEE has a symmetric and centralized distribution, indicating an accurate and unbiased filament segmentation. (d) Comparison of computational accuracy of cytoskeletal indices. Nine biologically interpretable cytoskeleton indices computed using the binary images rendered by different algorithms were compared with the ground truth. The index values were normalized to the fold of ground truth. Red dashed line indicates 1-fold, or identical to the ground truth. Error bar = standard deviation. Routine significance tests do not apply to b, c, and d because both mean and standard deviation serve as independent quantitative metrics of algorithm performance rather than indicators of significant differences. Related results are shown in Figs. S12, S13, S14, and S15.

This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal