Segmentation of video capsule endoscopic images Using alpha-divergence based active contour method.    

Purpose Examination of the whole gastrointestinal tract represents a challenge for endoscopists due to its length and inaccessibility using natural orifices. Moreover, radiologic techniques are relatively insensitive for diminutive, flat, infiltrative, or inflammatory lesions of the small bowel. Since 1994, video capsules (VCEs) have been developed to allow direct examination of this inaccessible part of the gastrointestinal tract and to help doctors to find the cause of symptoms such as stomach pain, disease of Crohn, diarrhoea, weight loss, rectal bleeding, and anaemia. The Pillcam© video capsule designed by Given Imaging Company is the most popular of them. This autonomous embedded system allows acquiring about fifty thousand images of gastrointestinal tract during more than twelve hours of an analysis. However, acquired images are of low resolution and imaging conditions could be very poor as the movements and the speed of the VCE as illumination are not entirely controlled all along the transit of the capsule through the whole gastrointestinal tract. As a consequence segmentation of VCE images is a real challenging problem: first of all specialists need a flexible segmentation tool that makes delineations of very different structures (polyps, metastasis… to name a few) possible and second, the proposed method should be efficient even in difficult imaging conditions. In this abstract, an adapted segmentation method based on statistical active contours for VCE segmentation is proposed and tested on various types of gastrointestinal structures.

Method Among the existing proposed segmentation methods in medical imaging, active contour models have attracted extensive research in the past two decades. Originally proposed in 1988 the basic idea of the active contour is to iteratively evolve an initial curve towards the boundaries of target organ driven by the combination of internal forces determined by the geometry of the evolving curve and the external forces induced from the image. Image segmentation method using active contour is usually based on extremizing (most often on minimizing) a functional, which is so defined that for curves close to the target boundaries it has small values. The extremization process is iteratively performed until convergence is reached i.e. the value of the computed energy does not meaningfully change from an iteration to the following one. In this abstract, the region-based criterion driving the segmentation is a statistical one computed on the probability density functions (PDF) (normalized histograms) characterizing the inner and outer regions defined by the evolving curve at a given iteration of the process. More precisely, it is proposed to maximize (instead of minimizing) a particular statistical distance, namely the alpha (α)-divergence, between non-parametrically estimated PDFs of these two regions. This similarity measure is a generalization of the well-known Kullback-Leibler (KL) divergence, which is an asymptotic case (α tends to one) of the alphadivergence. Main advantage of such an approach is threefold: (i) Maximizing the proposed distance between inner and outer PDFs makes the segmentation unsupervised since no prior information on the statistical distribution is required, (ii) the non-parametric estimation of the PDFs using standard Parzen window technic makes adaptability of the method to unusual textured areas possible (which is often the case in medical images), and finally, (iii) the alpha divergence criterion is a very flexible one that can be adapted (through the choice of α value) to very different kind of imaging scenarios. Moreover, this criterion overcomes some drawbacks of usual approach like the Chan–Vese based active contour approach, which does not perform well in the particular framework of computer-aided diagnosis because of the specificity of the data. At this stage of the study, only the luminance (grey level intensities) distributions of the pixels defining inner and outer regions of the curve are used for computation of

 

1  

corresponding PDFs, and the initialization of the segmentation (a 10 pixels diameter circle centred on the structure of interest) is manually performed. Value of alpha parameter is empirically tuned to best fit the expectations of the radiologist towards the considered pathology. About the implementation scheme, the energy to maximize is embedded into a level-set framework in order to intrinsically enable topological changes of the curve if required (see Fig. 1(a) for illustration). About the data, images are acquired using the PillCam© device and are taken from the freely available database “Gastro Vidéo Web” (http://dr.dm.gastro.free.fr/). Resolution of images is 320x320 pixels. To show the flexibility of the proposed statistical active contour approach, some very different segmentation tasks are proposed (see Fig. 1 for illustration) including some demanding ones.

Results Fig. 1 shows the best results obtained for different kind of segmentation tasks using the proposed approach.

(a) α=0.4

(b) α=0.4

(c) α=0.4

(d) α=0.7

(e) α=0.7

(f) α=0.8

(g) α=0.6

(h) α=0.4

(i) α=0.6 Fig. 1: VCE image segmentation using the proposed statistical region based active contour approach. (a) and (b) Polypes gastriques (syndrome de Peutz-Jeghers) (c) Polype sigmoïdien, (d) Lymphangiectasie, (e) Métastase de l’intestin grêle, (f) radiation enteritis (g) Angiome, (h) Citerne lymphatique, (i) Papille hypertrophique du canal anal, As one can notice, the maximisation of alpha-divergence is a flexible approach that could fit to different kind of segmentation structures acquired in some challenging conditions. For

 

2  

instance, even in the case of polyp segmentation, which luminance characteristics are quite similar to the ones of the associated background (Fig. 1(a-c)), the proposed segmentation approach makes delineation of the boundaries possible whereas usual approach like the Chan and Vese one would fail. Moreover, as illustrated Fig. 1(d) and (f), even if the “pathology” is a diffused one like radiation enteritis, the alpha divergence criterion remains efficient to highlight some particular regions of interest within considered image. It is also possible to note that the best-obtained results of Fig. 1 correspond to value of α lower than one, i.e. different from the asymptotic case α=1 related to the KL divergence most often used in the particular framework of statistical based active contour.

Conclusion In this abstract, a flexible statistical region based active contour criterion has been presented for VCE image segmentation. Based on alpha-divergence similarity, this criterion, embedded into a maximisation scheme of the corresponding segmentation energy, shows a real flexibility regarding the particular type of data considered which are acquired in non usual conditions. Prospective work will now focus on two main directions: (i) the taking into account of the colour information of VCE images and (ii) on the possibility of embedding within the capsule the proposed approach in order to make it compatible with real-time processing. This last point remains a very challenging one.

 

3