AWERProcedia Information Technology & Computer Science

AWERProcedia Information Technology & Computer Science Vol 03 (2013) 1285-1292

3rd World Conference on Information Technology (WCIT-2012)

Prototype-Based Contour Detection Applied to Segmentation of Phytoplankton Images Evaldas Vaiciukynas *, Department of Electrical & Control Equipment, Kaunas University of Technology, Studentų g. 50, LT-51368 Kaunas, Lithuania. Antanas Verikas, Department of Electrical & Control Equipment, Kaunas University of Technology, Studentų g. 50, LT-51368 Kaunas, Lithuania / CAISR, Halmstad University, Box 823, S-30118 Halmstad, Sweden. Adas Gelzinis, Department of Electrical & Control Equipment, Kaunas University of Technology, Studentų g. 50, LT-51368 Kaunas, Lithuania. Marija Bacauskiene, Department of Electrical & Control Equipment, Kaunas University of Technology, Studentų g. 50, LT-51368 Kaunas, Lithuania. Sigitas Sulcius, Coastal Research and Planning Institute, Klaipeda University, Herkaus Manto 84, LT92294 Klaipėda, Lithuania. Ricardas Paskauskas, Coastal Research and Planning Institute, Klaipeda University, Herkaus Manto 84, LT-92294 Klaipėda, Lithuania. Irina Olenina, Coastal Research and Planning Institute, Klaipeda University, Herkaus Manto 84, LT92294 Klaipėda, Lithuania / Department of Marine Research, Environmental Protection Agency, Taikos av. 26, LT-91144 Klaipėda, Lithuania. Suggested Citation: Vaiciukynas, E., Verikas, A., Gelzinis, A., Bacauskiene, M., Sulcius, S., Paskauskas, R. & Olenina, I. Prototype-Based Contour Detection Applied to Segmentation of Phytoplankton Images, AWERProcedia Information Technology & Computer Science. [Online]. 2013, 3, pp 1285-1292. Available from: http://www.worldrd education-center.org/index.php/P-ITCS Proceedings of 3 World Conference on Information Technology (WCIT-2012), 14-16 November 2012, University of Barcelon, Barcelona, Spain. Received 19 January, 2013; revised 18 May, 2013; accepted 09 September, 2013. Selection and peer review under responsibility of Prof. Dr. Hafize Keser. ©2013 Academic World Education & Research Center. All rights reserved. Abstract Novel prototype-based framework for image segmentation is introduced and successfully applied for cell segmentation in microscopy imagery. This study is concerned with precise contour detection for objects representing the Prorocentrum minimum species in phytoplankton images. The framework requires a single * ADDRESS FOR CORRESPONDENCE: Evaldas Vaiciukynas, Department of Electrical & Control Equipment, Kaunas University of

Technology, Studentų g. 50, LT-51368 Kaunas, Lithuania, E-mail address: [email protected] / Tel.: +370-67642585

Vaiciukynas, E., Verikas, A., Gelzinis, A., Bacauskiene, M., Sulcius, S., Paskauskas, R. & Olenina, I. Prototype-Based Contour Detection Applied to Segmentation of Phytoplankton Images, AWERProcedia Information Technology & Computer Science. [Online]. 2013, 3, pp 1285-1292. Available from: http://www.world-education-center.org/index.php/P-ITCS

object with the ground truth contour as a prototype to perform detection of the contour for the remaining objects. The level set method is chosen as a segmentation algorithm and its parameters are tuned by differential evolution. The fitness function is based on the distance between pixels near contour in the prototype image and pixels near detected contour in the target image. Pixels “of interest correspond to several concentric bands of various width in outer and inner areas, relative to the contour. Usefulness of the introduced approach was demonstrated by comparing it to the basic level set and advanced Weka segmentation techniques. Solving the parameter selection problem of the level set algorithm considerably improved segmentation accuracy. Keywords: Contour detection, level set, trainable segmentation, differential evolution, Quadratic-Chi distance;

1. Introduction Accurate segmentation of cells in microscopy images is currently a challenging task of large practical importance in cell biology. A successful contour detection technique for cells of interest could help to precisely evaluate their size and shape, and therefore be especially useful for morphometrics – quantitative analysis of form and structure. Morphometrics allows answering biological questions in a quantitative manner. Almost all algorithms for image segmentation have their parameters that need to be optimally set to obtain a good result. Nonetheless, only a few articles deal with automatic parameter selection in image segmentation. Some solutions, that try to tackle parameter selection problem, are designed specifically for deformable model [7], while others [16][13] are more universal. These solutions typically require ground-truth segmentation, but there exist attempts to operate without any prototype, for example [1]. We propose a novel prototype-based framework for contour detection, which can be useful in selecting parameters for any segmentation algorithm. We use the recent approximation to level set method and propose tuning its parameters by differential evolution. Optimization objective here is to minimise the custom distance between contour of the prototype image and a contour, detected by the level set method. Experimental evaluation considers comparison to the level set method with possibly best combination of parameters selected beforehand and kept fixed for all images analysed. Comparison with the advanced Weka segmentation technique has also been done. Advanced Weka segmentation plug-in in Fiji [11] image processing package conveniently combines any machine learning algorithm, available in Weka [5], with a set of selected image features to produce a pixel-based trainable segmentation. Results of our experiments indicate that application of the proposed prototype-based framework can be useful for solving segmentation tasks and improving contour detection. 2. Phytoplankton images Objects of interest in this research are phytoplankton cells, namely Prorocentrum minimum (P. minimum) species. It is believed that P. minimum cells gradually change their shape when adapting to adverse biotic (for example, with increased virus pressure) conditions. To prove this hypothesis quantitative evaluation of shape changes is needed. P. minimum is a microscopic, unicellular, seasonally bloom-forming, marine nanodinoflagellate with wide distribution in coastal and estuarine environments. P. minimum cells are small, approximately 15-20 µm long and 13-17 µm wide with concavity, known as apical spine, at the wider end [14][10]. Body shape varies from triangular to oval-round including intermediate, for example heart-shaped, forms [10].

1286


Presence of chlorophyll in phytoplankton cells makes them glow under UV light and this interesting property becomes of immense help for detecting P. minimum cells. Images for the analysis were obtained from a simple RGB colour camera, providing photos of 3264 x 2448 pixels, attached to an inverted microscope with the magnification factor of 400x. View, resulting from the G channel, has been used in experiments here. Two consecutive, with a delay of a few seconds, photos (using light and fluorescence microscopy) were obtained of the same location, see Figure 1. Fluorescence microscopy helps finding light microscopy image areas containing P. minimum cells. Such areas of 250 x 250 size were saved for further processing. 263 patches, each having a single P. minimum cell in focus, were extracted from 49 light microscopy images. Patches, extracted from Figure 1, are illustrated in Figure 2. 3. Segmentation using the level set method Highly successful and widely applied techniques in image segmentation are model based methods, known as active contours. An active contour model is an interface implementing separation of image objects from the background and can be implemented in explicit (Lagrangian) approach, where resulting interfaces are known as snakes, and implicit (Eulerian) approach, where resulting interfaces are known as level sets. The level set method, introduced in [8], represents the curve implicitly as the zero level set (points, having their value equal to zero) of a function defined over a regular grid. Novel level set algorithms include region information and led to a new concept of region competition [15], where two adjacent regions compete for the common boundary, additionally constrained by a smoothness term. Conventionally, the level set implementation of the curve evolution process is based on the solution of certain partial differential equations (PDEs). Classical approach is to update level set function globally by solving PDE over the entire grid, while narrow band approach focuses on the local evolution of the curve by solving PDE only for the neighbourhood of the zero-level set.

(a) Image from the light microscopy.

(b) Image from the fluorescence microscopy.

Figure 1. Example of microscopy image containing 6 P. minimum cells.

1287


Figure 2. Cut out patches with P. minimum cells (from Figure 1).

4. Proposed prototype-based framework The level set algorithm, chosen for our experiments, is based on the region competition concept and uses a narrow band of width one. Additionally, the element switching mechanism is used, where real-valued and PDE-based calculations are replaced by integer-based ones coupled with simple sign checks, to achieve runtime efficiency. Such discrete approach, approximating curve evolution, is a novel and fast solution, overcoming the need of solving PDEs, but preserving advantages of the level set method. More details can be found in [12], while Matlab implementation was adopted from [4]. The algorithm has 4 parameters (Na, Ns, Ng, σ) and evolves through 2 alternating cycles: for Na iterations the curve is evolved according to the data-dependent term using speed derived from the region competition model; for Ns iterations the curve is evolved according to the curvature-dependent smoothing regularization, using speed derived from a Gaussian filtering process (with window size Ng and kernel width σ). The process is stopped when the curve becomes stable (no change in the level set), or the given number of iterations is reached. The prototype-based framework requires at least a single object with a ground-truth contour as a prototype to perform detection of the contour for the remaining objects. Differential evolution is configured here to tune the 4 parameters of the level set method. A fitness function was constructed to measure the distance between the prototype and the target image and is given by the sum of several components as described further. 4.1. Pixels in concentric multiple bands Pixels of interest correspond to several concentric bands of various widths in outer and inner areas of the image, with respect to the contour. Number of bands and their width can be chosen to suit the problem at hand. Use of several narrow bands instead of a single wide band could potentially help in capturing image texture specifics in the contour neighbourhood. All pixels of the band are further pooled to obtain intensity histogram. Therefore, pooling of too wide band can result in too abstract 1288


approximation and loose some information on pixel intensity dynamics near contour. We used 3 outer and 3 inner bands, each having equal width of 3 pixels. 4.2. Distance as the optimization objective We propose a custom distance D, which is given by a sum of the following 4 components: Curvature DC. At each point of a curve, we assign a number reflecting the curvature of the curve at that position. The curvature is related to the rate at which the unit tangent vector is changing with respect to arc length. Resulting values are either zero, negative or positive, depending on the bending direction. We obtain curvature values by applying the M2003 estimator [6] to the contour line. We then take an absolute value of the result and make all the values binary by converting any non-zero value into 1. This creates a variable of two-levels, namely bending (1) and non-bending (0). The binary curvature vector of the prototype image is further compared to the target image by the two-sample two-tailed binomial test. Resulting low p value indicates that there exists a significant difference between the samples – the prototype curvature differs from the target curvature significantly. The final curvature value is given by DC = 1 – p. Ratio difference DR. The total number of pixels in the outer bands is divided by total number of pixels in the inner bands. We have one such ratio R for the prototype and one for the target. The parameter DR is given by: DR = |Rprototype – Rtarget|. Outer bands distance DO. Sum of distances between histograms of each corresponding outer band. Inner bands distance DI. Sum of distances between histograms of each corresponding inner band. The normalized pixel intensity histogram is found for each outer (or inner) band. The histogram distance, used in DO (or DI), is a cross-bin generalization of the Chi-Squared (introduced in [9] as Quadratic-Chi-Squared) distance. We set bin similarity threshold to 3, which means that 3 neighbouring bins to the left and 3 neighbouring bins to the right of each bin have decreasing influence on the distance. The chosen cross-bin histogram distance measures the distance between histograms obtained from pixel intensities in corresponding bands of the prototype and target images. 4.3. Optimization by differential evolution Differential evolution (DE) is a population-based evolutionary algorithm for stochastic non-linear optimization in continuous-valued spaces [3]. We have chosen the “DE/rand-to-best/1/exp” strategy and used population of 5 members, the maximum number of iterations was 4, stepsize Fweight = 0.85, and the crossover rate Crate = 0.9. Default parameter values of the level set algorithm with an allowable interval for optimization were: Na = 20 (the interval was from 10 to 40 with a step of 5), Ns = 5 (the interval was from 3 to 15 with a step of 1), Ng = 15 (the interval was from 5 to 25 with a step of 2), σ = 15 (the interval was from 10 to 30 with a step of 5). 5. Experimental setup and results Usefulness of the proposed approach was demonstrated by comparing it to the basic level set method (with fixed parameters for all images) and advanced Weka segmentation (trainable segmentation requiring a prototype). Parameter settings for the basic level set technique were the same as the default initial values for tuning with differential evolution. The maximum number of iterations was constant during all experiments and was set to 1000.

1289


Trainable segmentation is based on the idea that a segmentation pattern can be learned from the prototype example image and applied on unseen images by providing a pattern recognition technique with feature vectors, obtained from pixels of the original image stacked with pixels resulting from various transformations of the original image. A feature vector of 88 elements was obtained and is built of (a corresponding number of components is provided in the brackets): original image (1), Gaussian blur (4), Sobel filter (5), membrane projections (6), mean filter (4), variance filter (4), minimum filter (4), maximum filter (4), median filter (4), anisotropic diffusion (8), bilateral filter (4), Lipschitz filter (5), Kuwahara linear filter (3), and neighbours (32). Random forest [2] was employed as a classifier. The tested algorithms were not able to detect contours in very noisy images, nearby objects also impeded segmentation results. Segmentation of such images often resulted either in undersegmentation (leakages of contour and inability to separate several nearby cells), or oversegmentation (jagged contour and numerous small false objects). The proposed framework provided the best results by correctly detecting contours for 71.1 % of P. minimum cells (187 out of 263). Example results of detected contours are provided in Figure 3, while accuracy of the tested segmentation techniques is summarized in Table 1. Table 1. Summary of the segmentation results. Algorithm tested Level set method of Shi & Karl [12] Proposed prototype-based framework Advanced Weka segmentation of Fiji [11]

Correct 152 187 121

Accuracy, % 57.8 71.1 46.0

(a) Result of contour detection using the level set method of Shi & Karl [12].

(b) Result of contour detection using the proposed prototype-based framework.

1290


(c) Result of contour detection using the advanced Weka segmentation of Fiji [11]. Figure 3. Example of segmentation results obtained by the algorithms.

6. Conclusions and future work The proposed prototype-based framework improves segmentation of cells in microscopy images using a segmentation algorithm of choice. The main idea is to perform parameter selection by minimizing the specified distance between the contour of the prototype image and the contour resulting from segmentation of the target image. Usefulness of such approach was demonstrated by comparing it to the basic level set method (fixed parameters for all images) and trainable segmentation (advanced Weka segmentation requiring a prototype). Parameter tuning increased flexibility of the level set method, improved precision of the resulting contour, and made segmentation more robust to noise. Further improvements could be gained by considering several prototypes. The heuristic for selecting proper examples for prototypes could be based on k-medoids clustering of images. The proposed technique could then be applied for each cluster. After all images get segmented, a fraction of the resulting best contours could become prototypes and the framework could be restarted with more prototypes. In case of several prototypes, an average of the introduced distance should be considered. Such modification could potentially improve segmentation results. Acknowledgement This research was funded by a grant (No. LEK-09/2012) from the Research Council of Lithuania. References [1] [2] [3] [4]

[5] [6]

M. A. Abdul Karim. Automated parameter selection for segmentation of tube-like biological structures using optimization algorithm and MDL. PhD thesis, Rensselaer Polytechnic Institute, 2005. L. Breiman. Random forests. Machine Learning, 45(1):5–32, October 2001. S. Das and P. N. Suganthan. Differential evolution: A survey of the state-of-the-art. IEEE Transactions on Evolutionary Computation, 15(1):4–31, Feb. 2011. T. Dietenbeck, M. Alessandrini, D.Friboulet, and O. Bernard. CREASEG: A free software for the evaluation of image segmentation algorithms based on level-set. In Proceedings of the 17th IEEE International Conference on Image Processing (ICIP), pages 665–668. IEEE, Sept. 2010. M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H. Witten. The WEKA data mining software: an update. ACM SIGKDD Explorations Newsletter, 11(1):10–18, June 2009. S. Hermann and R. Klette. A comparative study on 2d curvature estimators. In Proceedings of the International Conference on Computing: Theory and Applications (ICCTA), pages 584–589, Kolkata, India, Mar. 2007. IEEE Computer Society.

1291


[7]

[8] [9] [10]

[11]

[12] [13]

[14] [15]

[16]

S. Kumar and M. Hebert. Discriminative random fields: a discriminative framework for contextual interaction in classification. In Proceedings of the 9th IEEE International Conference on Computer Vision, volume 2, pages 1150–1157, Oct. 2003. S. Osher and J. A. Sethian. Fronts propagating with curvature-dependent speed: algorithms based on Hamilton-Jacobi formulations. Journal of Computational Physics, 79(1):12–49, Nov. 1988. O. Pele and M. Werman. The Quadratic-Chi Histogram Distance Family. In Proceedings of the 11th European Conference on Computer Vision (ECCV), page 13, Crete, Greece, Sept. 2010. S. Pertola, H. K. M. A. Faust, and G. HĆ¤llfors. Morphology of Prorocentrum minimum (Dinophyceae) in the Baltic Sea and in Chesapeake Bay: Comparison of Cell Shapes and Thecal Ornamentation. Botanica Marina, 46(5):477–486, 2005. J. Schindelin, I. Arganda-Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch, C. Rueden, S. Saalfeld, B. Schmid, J.-Y. Tinevez, D. J. White, V. Hartenstein, K. Eliceiri, P. Tomancak, and A. Cardona. Fiji: an open-source platform for biological-image analysis. Nature Methods, 9(7):676–682, July 2012. Y. Shi and W. C. Karl. A Real-Time Algorithm for the Approximation of Level-Set-Based Curve Evolution. IEEE Transactions on Image Processing, 17(5):645–656, May 2008. T. Torsney-Weir, A. Saad, T. Moller, H.-C. Hege, B. Weber, and J.-M. Verbavatz. Tuner: Principled Parameter Finding for Image Segmentation Algorithms Using Visual Response Surface Exploration. IEEE Transactions on Visualization and Computer Graphics, 17(12):1892–1901, Dec. 2011. V. Velikova and J. Larsen. The Prorocentrum cordatum/Prorocentrum minimum taxonomic problem. Grana, 38(2-3):108–112, 1999. S. C. Zhu and A. Yuille. Region Competition: Unifying Snakes, Region Growing, and Bayes/MDL for Multiband Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18:884–900, 1996. S. Zickler. Learning optimal parameters of graph-based image segmentation. Final Project Report for Machine Learning Course (School of Computer Science, Carnegie-Mellon University), 2006.

1292