Self-initialized active contours for microscopic cell image segmentation

Abstract

Level set models are suitable for processing topological changes in different regions of images while performing segmentation. Active contour models require an empirical setting for initial parameters, which is tedious for the end-user. This study proposes an incremental level set model with the automatic initialization of contours based on local and global fitting energies that enable it to capture image regions containing intensity corruption or other light artifacts. The region-based area and the region-based length terms use signed pressure force (SPF) to strengthen the balloon force. SPF helps to achieve a smooth version of the gradient descent flow in terms of energy minimization. The proposed model is tested on multiple synthetic and real images. Our model has four advantages: first, there is no need for the end user to initialize the parameters; instead, the model is self-initialized. Second, it is more accurate than other methods. Third, it shows lower computational complexity. Fourth, it does not depend on the starting position of the contour. Finally, we evaluated the performance of our model on microscopic cell images (Coelho et al., in: 2009 IEEE international symposium on biomedical imaging: from nano to macro, IEEE, 2009) to confirm that its performance is superior to that of other state-of-the-art models.

Figure 1

Example images: homogeneous image (left); inhomogeneous image (right).

Figure 2

Effect of epsilon on (a) Heaviside function, and (b) Dirac delta function.

Figure 3

Graphical representation of the proposed algorithm.

Figure 4

Comparison of proposed model with other models on the same image with five different levels of inhomogeneity: (col 1) input image with initial contour; (col 2) C-V; (col 3) LBF; (col 4) LIF; (col 5) VLSBC; (col 6) Zhang et al.; (col 7) proposed model.

Figure 5

Segmentation results with contours of different shapes and sizes at different locations on the input image. First row: input image with different contours; second row: associated results.

Figure 6

Results of the proposed model in comparison with those of other models on synthetic images: (col 1) input image; (col 2) C-V; (col 3) LBF; (col 4) LIF; (col 5) VLSBC; (col 6) Zhang et al.; (col 7) FRAGL ; (col 8) proposed model.

Figure 7

Results of proposed model in comparison with those of other models on real medical images: (col 1) input image; (col 2) C-V; (col 3) LBF; (col 4) LIF; (col 5) VLSBC; (col 6) Zhang et al.; (col 7) FRAGL ; (col 8) proposed model.

Figure 8

Top row: input image with initial contour, segmentation results for LIF, VLSBC and segmentation results for Zhang et al., respectively. Second row: segmentation results for Akram et al., Akram et al., FRAGL, and the proposed method, respectively.

Figure 9

Quantitative analysis chart showing a graphical illustration of segmentation accuracy, Dice index, Jaccard index and BF score matrices.

Figure 10

Col (1): Input image corrupted with Salt & Pepper noise levels (0.01, 0.02, 0.03, 0.04, 0.05), Col(2): Segmentation results of (col 2) C-V; (col 3) LBF; (col 4) LIF; (col 5) Adaptive; (col 6) Retinex; (col 7) FRAGL ; (col 8) proposed model.

Figure 11

Col (1): Input image corrupted with Guassian noise levels (0.01, 0.02, 0.03, 0.04, 0.05), Col(2): Segmentation results of (col 2) C-V; (col 3) LBF; (col 4) LIF; (col 5) Adaptive; (col 6) Retinex; (col 7) FRAGL ; (col 8) proposed model.

Figure 12

JS values for Figs. 10 and 11 are represented by (left) and (right), respectively.

Figure 13

Right Image: Row (1): Input image with initial contours pf different shapes at different positions; Row (2) Adaptive; Row (3) Retinex; Row (4) FRAGL; Row (5) proposed model. Left Image: Computational Cost Chart.

Figure 14

Segmentation results against time complexity (a) SPF with the Akram et al. formation (b) SPF formation with out membership function in the proposed methodology.

Figure 15

Ablation study over microscopic images database by removing different terms from the proposed function.