Colposcopy is at present a real diagnosis tool. Colposcopy is the observation of the cervix through a colposcope. The screening focuses on two epithelium,.
CONTOUR FEATURES FOR COLPOSCOPIC IMAGE CLASSIFICATION BY ARTIFICIAL NEURAL NETWORKS Isabelle Claude*, Renaud Winzenrieth*, Philippe Pouletaut* and Jean-Charles Boulanger** *Université de Technologie de Compiègne, France ** Centre de gynécologie Obstétrique d’Amiens
Abstract This article presents colposcopic image classification based on contour parameters used in a comparison study of different artificial neural networks and the knearest neighbors reference method. In this study, significant image data bases are used (283 samples) from which a set of original parameters is extracted to characterize the attribute of contour. More precisely, we quantify the notion of sharp contours vs blurred contours in computing spatial parameters based on the number of small regions near boundaries of objects and frequency parameters based on power spectrum of lines cutting these boundaries. Experimental results show the feasibility of this study and the efficiency of the set of parameters since 95.8% of contour image set has been correctly classified.
1. Introduction Cervical cancer is the most common form of cancer in women under 35 years, worldwide. Pap smear is the primary screening method for the detection of this cancer. It consists in cytologic examination of cervical epithelium after appropriate fixation and staining. But, its accuracy is limited. Colposcopy is at present a real diagnosis tool. Colposcopy is the observation of the cervix through a colposcope. The screening focuses on two epithelium, squamous without glycogen and columnar with glycogen. Three different images are generally used: one for the observation of the cervix without any preparation; a second one after application of an acetic acid solution (dissolution of mucous and accentuation of atypical areas) and a final observation after application of a Lugol’s solution (dark brown color stain of glycogen). The given images are characterized by many attributes such as color, texture or relief. Thus, their automatic analysis is difficult. So why the existing studies have focused on a particular aspect of the whole challenge : enhancement of image quality [1], abnormal vascular patterns quantification by texture analysis [2], lesion area measurements with 3D correction [3], comparison
between colposcopic and cytological data [4], fuzzy logic for characterizing lesion morphology [5]. But, there is no global tool for gynaecologists. In this context, a project was raised in collaboration with an obstetrical center, concerning the development of a semi automatic tool for colposcopic images analysis [6]. Considering the requirements of a such computer aided diagnosis tool, we assume that learning techniques should be adapted to our problem. In particular, neural networks can acquire prior knowledge t hrough a learning process. With a suitable learning data base, they can be independent of lightening conditions. By modifying number and nature of inputs, new parameters can be taken into account. Moreover, neural networks have proved their efficiency in many medical applications [7-8]. The proposed solution consists in a global diagnosis tool comp osed by learning subsystems. Each subsystem concerns one relevant attribute for each of the three images and allows to give a part of the diagnosis. This system mimics gynaecologists’ diagnosis methodology. We already validated this method for colposcopic image classification based on color parameters [9]. The work presented here concerns the validation of this assumption as far as contours of lesions are concerned. At first, we present the methodology adopted in order to choose the set of parameters which will characterize contours of lesions. Then, we explain the creation of our learning data bases. We present also the choice of the different classifiers tested. Finally, promising technical and clinical results are given, illustrating the ability of the method to classify contours on colposcopic images.
2. Preliminary results A questionnaire send to 25 gynecologists has permitted us to know which information is essential in colposcopic diagnosis. Seven groups of parameters have been identified : color, texture, contours, vascular patterns, localization, form and relief. The results of questionnaires show also the relative importance of each observation (without preparation, with acid acetic and with lugol) and quantifies the contribution of parameters for each observation.
1051-4651/02 $17.00 (c) 2002 IEEE
a. Sharp contour
b. Blurred contour : c. Blurred contour : fuzzy edge small regions Figure 1 : sharp and blurred contours in colposcopy
It appears that color is very important for the three observations, contours for lugol observation and vascular patterns for acid acetic observation. We present here only the work on contour parameter.
3. Contour parameters used In colposcopy, observing contours on lugol image means detecting if contours of the form around orifice are either blurred or sharp. Blurred contours are either fuzzy or multiple small regions around edges. In the first case (see figure 1.b and figure 1.c), the diagnosis is often bad whereas in the second case (see figure 1.a), it would be better. Sometimes, it is very difficult to classify a contour (see figure 1.d). In order to characterize both types of contours, we determine spatial and frequency parameters on gray-level lugol images. The different steps of the method are described in figure 2 and below. The first parameter is the number of regions around edges. For its computation, we threshold image with the value separating the two modes of the histogram. Indeed, histogram of pathological lugol images are always bimodal. Then, we count 4-connexity regions obtained: this is the first parameter. After eliminating small regions, we examine edges of the form. Only a few points are necessary to characterize them. We determine frequency parameters. We draw a set of lines going through the form and cutting the edges. Each line represents a profile of the contour and its power spectrum is characteristic of the type of contours. We compute therefore four different parameters : amplitude of first peak, frequency of the end of first peak, area under first peak and area under the other peaks (see figure 3). A Principal Component Analysis has permitted to test this set of parameters and to show their good behavior to discriminate both types of contours.
d. Doubtful contour
4. Image data base We select 30 lugol images among which 13 are normal cases, 15 are pathological cases and 2 are doubtful cases from Doctor Coupez colpophotographies bank [10]. On each image, we select from 3 to 16 samples resulting in 283 samples for contour data base.
5. Learning systems used We have tested two different neural networks : - a multilayer perceptron (MLP) with three layers and six nodes per layer. The learning method was the back-propagation algorithm. - and a probabilistic network (PNN) based on Parzen estimation. This choice is justified by two reasons : the first concerns the aim of this study (to validate the use of learning systems) and the second is the interesting characteristics of both networks (simple and universal classifiers, and supervised learning) [11]. The k nearest neighbors algorithm (kNN) was our reference method. Optimization of MLP parameters (number of hidden layers neurones is equal to 4,6 and activation function is sigmo ï d/linear) and of kNN parameters (number k is equal to 1 and distance is manhattan) is performed by crossvalidation technique.
6. Results Technical validation : We choose the “Leave-one-out” technique to evaluate the different classifiers. It consists in taking N -n elements of the data base for the learning process, classifying the n elements left and doing it again 50 times. Here, n is equal to 20% of the data base. Results are given in table 1.
1051-4651/02 $17.00 (c) 2002 IEEE
Original image
Histogram computation
Determination of threshold value T
T Thresholding
Count of small regions then their removing
x Search of a point inside the form
1 Plot of lines cutting edge
2 x
Computation of power spectrum for each profile
3
1 2
Computation of contour parameters
3
Figure 2: The different steps for determining contour parameters
Power spectrum A A1 A2
F Figure 3 : frequency parameters : A is the amplitude of first peak, F the frequency of this peak, A1 area under this peak and A2 area under re st of the spectrum
1051-4651/02 $17.00 (c) 2002 IEEE
Globally, the three classifiers give good results, but MLP classifier is the best one. The percentage of true positive concerns correctly classified sharp contours. This percentage is better than the percentage of true negative concerning correctly classified blurred contours. Indeed, the notion of sharp contours is more precise than the notion of blurred contours. MLP PNN kNN % of good 95.8 85.3 68.9 classification True positive 96.3 98.6 82.1 True negative 95.3 66 49.8 Table 1 : Percentage of good classification given for each classifiers Clinical validation We perform a double blind test with an expert and MLP. Twenty images with a known diagnosis has been used and classified by both. The results obtained by both are comparable and good (90% of good classification). But, the expert which collaborates at this blind test, has a strong experience and is well-known in the French colposcopic community. Indeed, a recent study [12] has shown that, compared with real diagnosis given by histological analysis, a group of experts give 77,5% of good class ification and a group of inexperienced gynaecologists gives 55% of good classification. So, our results seems to be very promising.
[2] J. Qiang, J. Engel and E. Craine, “Classifying Cervix Pattern With Texture Analysis,” Pattern Recognition, vol. 33, pp. 1561-1573, 2000. [3] B.L. Craine, E.R. Craine, C. J. O’Toole, and J. Qiang, “ Digital Imaging Colposcopy : Corrected Area Measurements Using Shape -from-Shading,” IEEE Tr. On Medical Imaging, vol. 17, pp.1003-1010, 1998. [4] M.I. Shafi, J.A. Dunn, R. Chenoy, E.J. Buxton, C. Williams, and D.M. Luesley, “ Digital Imaging colposcopy, Image analysis and Quantification of the Colposcopic Image,” British Journal of Obstetrics and Gynaecology, vol. 101, pp. 234-238, 1994. [5] E. Binaghi, and A. Rampini, “Learning of Uncertain classification Rules in biomedical Images : The Case of colposcopic Images,” Information Processing and Medical Imaging, 12th International conference, WYE. Computer Science 511, Springer Berlin in : Lecture Notes in Computer Science, vol. 511, pp. 434-443, 1991. [6] I. Claude, P. Pouletaut, S. Huault, J-C. Boulanger, Integrated color and texture tools for colposcopic image segmentation. ICIP’2001, Int. Conf. on Image Processing, Greece, oc tobre 2001. [7] Cheng K.S. & al, The Application of Competitive Hopfield Neural Network to Medical Image Segmentation, IEEE Transactions on Image Processing, 1998.
7. Conclusion We present an efficient colposcopic image classification method based on contour parameters. These contour parameters allow to discriminate sharp and blurred contours which are, in colposcopy, a clinical sign of respectively normal and pathological cases. But the classification results for contour attribute does not give the final diagnosis of the studied case if the others parameters are not taken into account. So why, at present, contour parameter set is combined with color and texture parameters sets in an integrated software for gynaecologists [6]. The following step will concern relief attribute study.
[8] Cho J.M., Chromosome Classification Using Backpropagation Neural Networks, IEEE Engineering in Medecine and Biology, 2000. [9] I. Claude, R. Winzenrieth, P. Pouletaut, J-C. Boulanger, Classification of color colposcopic images by neural networks. CGIV’2002, Int. European Conference on Graphics, Imaging, and Vision, Poitiers, France, april 2002. [10] Coupez, ‘Colpophotographies’, Collection de la société de Colposcopie et Pathologie Cervico Vaginale, ed. Pr Boulanger, 1995. [11] Bishop C.M., Neural Network for Recognition, Oxford University Press,1995.
8. References [1] B.L. Craine, E.R. Craine, “Digital Imaging Colposcopy : Basics Concepts and Applications”, Obstet. Gynecol., vol. 82, pp.867-873, 1993.
Pattern
[12] C. Quereux, J. Gondry, J-P. Bory and J-C. Boulanger, Démarche qualité en colposcopie, JTA’2001, Paris, 2001.
1051-4651/02 $17.00 (c) 2002 IEEE
