Decision Trees for the detection of skin lesion patterns in lower limbs ulcers

Size: px

Start display at page:

Download "Decision Trees for the detection of skin lesion patterns in lower limbs ulcers"

Sylvia West
5 years ago
Views:

1 2016 International Conference on Computational Science and Computational Intelligence Decision Trees for the detection of skin lesion patterns in lower limbs ulcers José Luis Seixas Jr. Computer Science Department State University of Paraná (UNESPAR) Apucarana - PR, Brazil jlseixasjr@gmail.com Rafael G. Mantovani Sciences Institute of Mathematics and Computers (ICMC) University of São Paulo (USP) São Carlos - SP, Brazil rgmantov@icmc.usp.br Abstract Misleading diagnosis of skin diseases can result in complications during the healing process. Skin images provide important information for the medical staff for information storage and exchange, to trying to prevent this misdiagnosis from happening. For such, a good segmentation process is needed. The segmentation of these images is already being used and has been an effective tool for skin diseases recognition. This paper presents a method for targeting seeds for region growing algorithms, as several of region growing algorithms have good clustering results, but are sensitive to seed. Machine learning were use to create the seed for segmentation of medical images of skin ulcers in the lower limbs. For machine learning, decision tree algorithms were used, which bring a more intuitive approach. The results were compared with gold standard obtained with the help of experts, the results were good and opened paths that can be followed for further work since, even though good results, they can still be improved. I. INTRODUCTION Incorrect diagnosis of skin diseases may lead to complications during the healing process. Weenig et al. [1] cites the case of pyoderma gangrenosum which has several causes and may have different treatments which would not be recommended for patients with other diseases. Lesions recognition by characteristics such as size, severity, presence of infection and vascularization has an important impact on the determination of the most appropriate treatment for each patient [2]. Many skin lesion descriptors presented by Frykberg [2] can be obtained from segmented images analysis of the skin lesion, but a good segmentation is required to work with accurate and reliable data. There is a large volume of information that a medical image can offer, but it requires sophisticated analytical techniques so it can help the medical staff make decisions [3]. Analysis that can determine an injury status and evolution, but that depends on its characteristics cited by Frykberg [2]. These characteristics can be obtained by image analysis, but before that, an image segmentation process is necessary to evidence of lesion area. Segmentation is the first step to identify the injury condition, also enables automation in getting information on a large scale and final analyses standardization, which is no longer subjective, increasing the accuracy of comparative methods. Color is a low-level feature that can transmit relevant information, as described by Frykberg [2] of a given image according to Fu et al. [4], so many algorithms turn to colors in targeting areas of interest. Cheng et al. [5] describe various algorithms and different proposals on targeting colors and cites regions based approaches as a good and not noise sensitive option. Many problems in medical tests depend on or can be helped using digital images. These photos depend on experts to describe each one of them. In some cases, doctors may not be with the patient, in these cases, images are options to send or save the patient s history. However, an analysis done entirely by doctors will still be required. According to Cheng et al. [5], there are many different techniques for segmentation, such as histogram, fuzzy logic, regions approaches, borders and artificial intelligence. Bhattacharyya [6] mentions techniques based on wavelet and genetic algorithms. In Kekre et al. [7], we see that image segmentation is a process that requires a great human intervention or the system will be very specific, in other words, a problem-oriented application. Trying to avoid human intervention, our approach uses a decision tree to segment medical images from chronic leg ulcers caused by venous insufficiency. This study was approved by the Ethics Committee of the State University of Londrina (UEL) according to law 315/06 of 02/02/07 (Amendment 1/2010). Individuals were informed of the nature of this research and signed consent documentation. The survey was conducted in CISMEPAR (Consórcio Intermunicipal de Saúde do Médio Paranapanema) in partnership with the University Hospital (UH) of Londrina / UEL.. II. PROPOSED APPROACH Images were acquired from eight different patients at different stages of lesion with a Sony, DSC-S1900 model which allows a good portability, it is a small camera and also accepts manual settings. The settings were used in manual mode to control similarity among images, since auto settings would change depending on external condition changes, leaving only focus to be done automatically. Thus, camera could be configured only once and the medical team itself could make further acquisitions, with no need to have some expert on image acquisition /16 $ IEEE DOI /CSCI

The set with the best features that fit this work purpose was as following: Exposure Value: -0.3; F-Number 3.1; ISO: 400; Flash: disabled.

2 Images sets were acquired to adjust the camera settings, which would be used in all further acquisitions. These settings has as main purpose, to maintain color quality and avoid light overexposure, thus avoiding reflections that could damage colors sheet. The set with the best features that fit this work purpose was as following: Exposure Value: -0.3; F-Number 3.1; ISO: 400; Flash: disabled. Images acquired with these previously mentioned settings, had lesion area limits highlighted by doctors in hospital so it would be possible to compare what may be considered lesion for future results. Fig. 1. Image obtained in cited conditions. Figure 1 shows an acquired image example with previously described configurations. Settings were kept on all images, in order that, since images are from the same type of injury, similarity would be kept among all images. Thirty-three images were acquired in bitmap format. As our purpose did not require high resolution images, they were reduced and kept at 400 x 225 pixels resolution, in addition, images with very high resolution require unnecessary processing, but would not have a significant difference in efficacy. The white square presented in Figure 1 is a marker for area calculation, because segmentation process is one of the steps for lesion characteristics extraction. Thus, an object of easy detection of known size was positioned in image as a reference. Moreover, the square can also be used as a good object to camera s white balance, which is a color adjustment made to compensate for effect caused by environment illumination temperature, approaching image colors to real object color [8]. In Figure 2 is shown a region marked by doctors from lesion area found in Figure 1. These images were obtained with the help of medical experts who identified regions that would be considered injury of each image. Following this marking which were made on the boundaries of the blue area, marked area was filled for easy viewing and subsequent comparisons. Filling also assisted on the automatic identification of lesion area. Fig. 2. Image with region bounded by doctors. Ten of the thirty three images acquired were used to build the data set, i.e., nine hundred thousand entries. In [9], authors state that the most famous form of diagnosis is the ABCD rule (Asymmetry, Border, Color, Diameter) and new trends for skin image segmentation are through colors and textures methods. Some studies seek a different color systems to analyze uniformity of color and texture by the dependence of light intensity found on the RGB system [10]. However, RGB system is a color space that has a very close representation of human vision perception [11]. Thus, RGB system was chosen as a target for our experiments. Evaluation strategy will use images from experts markings as a gold standard. So, as we seek the final decision from medical staff to aid lesion diagnosis, this markings will be used to certify what is or is not part of the injury. As constructional features of the instances of Weka software were attributes used color RGB image, with the label on which the pixel area in question belongs. Examples of input lines for the algorithm are: 54, 41, 35, false; 27, 23, 24, true; 26, 22, 23, true; 225, 206, 192, false; Where first value is the red value (R), second green (G), third blue (B), from RGB system and last value is the corresponding background relationship, where will be used true when the presented pixel value belongs to a pixel which is part of background, and false for pixels that belong to the lesion area. Experiments were performed with software Weka function set importing its methods directly in Java code. Values of DT attributes were used as follow: Binary Splits: disabled; Confidence Factor: 0.25; Minimum number of instances per leaf: 2; Set number of folds: 3; Seed: 1; Subtree Raising: enabled; Use unpruned tree: disabled; Laplace Smoothing: disabled;

Weka software was used as a time resource, since it has the chosen DT implementation, in addition, the software, which is used in scientific research, provides its direct use in source code, allowing

3 Weka software was used as a time resource, since it has the chosen DT implementation, in addition, the software, which is used in scientific research, provides its direct use in source code, allowing their functions to be used directly in other implementation, facilitating automation for large-scale projects. This feature is also found in Java language, where We can simply import Java Archive (.jar) in our project to import and execute the wanted functions. Decision Tree (DT) induction algorithms is popularly used as classifiers due to its resemblance to human reasoning [12], which makes a good target to deal with problems that affect other areas, not only Computer Science, since researchers can understand its operation. Some advantages of DTs presented by Mantovani et al. [13] indicates that DTs can be strongly recommended for a wide range of classification problems. Among these advantages, we can find robustness to noise, low computational cost and the capacity to work with redundant attributes that makes DTs work fine even with simple data sets. III. RESULTS AND DISCUSSIONS Results from DT classification can be seen in Table I, which brings the percentage in each indicators for a binary classification test using J48 algorithm from Weka. TABLE I PIXEL RECOGNITION PERCENTAGE. Index True False Positive Positive False True 3 88,65% 2,38% Negative Negative 6 95,23% 0,73% 1,12% 7,85% 10 98,20% 0,82% 1,58% 2,47% 11 94,33% 3,61% 0,51% 0,48% 16 92,69% 1,29% 0,12% 1,95% 17 93,06% 1,51% 0,93% 5,08% 18 92,60% 2,30% 1,25% 4,17% 21 92,86% 0,75% 0,47% 4,64% 22 91,59% 3,61% 0,50% 5,89% 23 91,15% 0,60% 0,59% 4,22% 24 87,38% 4,03% 0,41% 7,83% 25 85,29% 6,66% 1,98% 6,61% 27 93,80% 1,89% 0,45% 7,60% 29 94,18% 1,03% 0,14% 4,18% 30 93,45% 0,83% 0,08% 4,71% 31 82,88% 0,46% 0,44% 5,28% 32 94,44% 0,77% 2,92% 13,74% Mean 91,87% 1,96% 0,78% 4,01% 0,84% 5,34% As a reminder, not all indexes are presented in Table I because some of them were used to assemble the DT classification model. Images used to DT construction were chosen based on different phases of lesion and different patients. They were used at least one image of each of the eight patients at different stages and state of injury. In Table I, we show the image index, which is just a number to label each image without needing to explicit any information about the patient it came from. Images used in training were got from any number sequence, because diversity of patients and regions were wanted. This percentage were obtained by summing all the answers got from the algorithm and dividing it by the total possible values (all pixels in each image). Fig. 3. DT recognition percentage. Values were ordered by true positives for better monitoring of algorithm correctness ratio for each image represented by indexes. For better visualization, obtained results are represented in Figure 3. Indicators from Table I and Figure 3 represents as follow: True Positive: Any pixel the DT model indicated as background and was truly background (white in Figure 2); False Positive: Any pixel the DT model indicated as background, but it was part of the lesion; False Negative: Any pixel the DT model indicated as part of the lesion, but it was background; True Negative: Any pixel the DT model indicated as lesion and was truly lesion (blue in Figure 2); Thus, on average, 97.20% of pixels of each image was classified correctly background and region of interest. Representing an average accuracy of 97.21% and average accuracy of 97.91%. As the image consists mostly of background, True Positive, by itself, is not a good indicator for good results. But when using True values and high accuracy, this results become a more reliable indicator. Figure 4, shows four examples of segmentation made by DY model. Where Figures 4 (b) and (c) are from same patient at different acquisition times. Figures 4 (a) and (d) are from two other patients. For comparison, Figure 5 shows the same images example, but obtained from doctors markings, i.e., how images from Figure 4 should look in case algorithm could obtained 100% accuracy

4 Fig. 4. Segmentation made by J48 algorithm. IV. CONCLUSION Machine Learning (ML) algorithms have already been shown able to work in pattern recognition and clustering tasks, which impact directly in image segmentation problems. Its use in a particular skin lesion image segmentation task was equally promising and may also be improved or cover similar tasks. Seixas Jr. et al. [14] bring a proposal to search, using signal processing, for a region that is surely lesion which can be used as a preprocessing operation to a continuous learning processes, avoiding need to have specific images set for training. Therefore, the developer can use known algorithms and already implemented or propose a new solution using the data that can be extracted from the region found. These characteristics mean that even people who are not part of the computer science context to use this strategy. Similarly, DT has proved a very effective alternative to the problem of segmentation. In specific case of decision trees, a few steps after processing can be incorporated to refine output. As an example, it could be used a fill algorithm to cover small regions avoiding holes in lesion area. Even considering sizes, algorithms which eliminates very small unconnected areas may be considered. Other ML techniques may also be used in this pattern recognition task, perhaps, even hybrid models could be considered. Since this paper were made up of simple information that can be intuitive and fast processing, further studies may be able to show better results. For example, rather than filling regions, it could be used a neural network filter the classification made by decision trees. Finally, studies can also be made with any one of the techniques in attempts to separate different areas of lesion, as presence of infection, epithelialized area, necrosis, or any other area that may constitute a lesion. We hope that this work can help other researchers from different areas of how to use these techniques in resolutions of their respective segmentation problems. And especially that this work can briefly lead to applications that are able to assist medical teams and thus benefit the community and leverage more research that can bring further benefits to the community and medicine. REFERENCES [1] R. H. Weenig, M. D. P. Davis, P. R. Dahl, and D. Su, Skin ulcers misdiagnosed as pyoderma gangrenosum, The New England journal of medicine, vol. 347, no. 18, pp , [2] R. G. Frykberg, Diabetic foot ulcers: Current concepts, The Journal of Foot and Ankle Surgery, vol. 37, no. 5, pp , [3] M. Wernick, Y. Yang, J. Brankov, G. Yourganov, and S. Strother, Machine learning in medical imaging, Signal Processing Magazine, IEEE, vol. 27, no. 4, pp , [4] K. Fu, C. Gong, J. Yang, Y. Zhou, and I. Y.-H. Gu, Superpixel based color contrast and color distribution driven salient object detection. Sig. Proc.: Image Comm., vol. 28, no. 10, pp , [5] H.-D. Cheng, X. Jiang, Y. Sun, and J. Wang, Color image segmentation: advances and prospects. Pattern Recognition, vol. 34, no. 12, pp , [6] S. Bhattacharyya, A brief survey of color image preprocessing and segmentation techniques, Journal of Pattern Recognition Research, vol. 1, pp ,

5 Fig. 5. Segmentation made with doctors markings. [7] H. B. Kekre, T. K. Sarode, and B. C. Raul, Color image segmentation using kekre s fast codebook generation algorithm based on energy ordering concept, in Proceedings of the International Conference on Advances in Computing, Communication and Control, ser. ICAC3 09. New York, NY, USA: ACM, 2009, pp [8] Y.-C. Liu, W.-H. Chan, and Y.-Q. Chen, Automatic white balance for digital still camera, IEEE Trans. on Consum. Electron., vol. 41, no. 3, pp , Aug [Online]. Available: [9] C. Serrano and B. Acha, Pattern analysis of dermoscopic images based on markov random fields, Pattern Recognition, vol. 42, no. 6, pp , [10] E. Littmann and H. Ritter, Adaptive color segmentation-a comparison of neural and statistical methods, Neural Networks, IEEE Transactions on, vol. 8, no. 1, pp , [11] F. M. Lopes and L. A. Consularo, A rbfn perceptive model for image thresholding, in Computer Graphics and Image Processing, SIBGRAPI th Brazilian Symposium on. IEEE, 2005, pp [12] R. Barros, M. Basgalupp, A. de Carvalho, and A. Freitas, A survey of evolutionary algorithms for decision-tree induction, in IEEE Transactions on Systems, Man and Cybernetics, Part C: Applications and Reviews, May 2012, pp [13] R. G. Mantovani, T. Horváth, R. Cerri, J. Vanschoren, and A. C. P. L. F. d. Carvalho, Hyper-parameter tuning of a decision tree induction algorithm, in 2016 Brazilian Conference on Intelligent Systems (BRACIS), Oct 2016, pp [14] J. L. Seixas, S. Barbon, C. M. Siqueira, I. F. L. Dias, A. G. Castaldin, and A. S. Felinto, Color energy as a seed descriptor for image segmentation with region growing algorithms on skin wound images, in e-health Networking, Applications and Services (Healthcom), 2014 IEEE 16th International Conference on, Oct 2014, pp

Automatic Licenses Plate Recognition System

Automatic Licenses Plate Recognition System Garima R. Yadav Dept. of Electronics & Comm. Engineering Marathwada Institute of Technology, Aurangabad (Maharashtra), India yadavgarima08@gmail.com Prof. H.K.