TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 19, No. 4, August 2021, pp. 1357 1368 ISSN: 1693-6930, accredited First Grade by K emenristekdikti, No: 21/E/KPT/2018 DOI: 10.12928/TELK OMNIKA.v19i4.20292 r 1357 A machine lear ning appr oach f or the r ecognition of melanoma skin cancer on macr oscopic images J air o Hurtado, Francisco Reales Pontificia Uni v ersidad Ja v eriana, Electronics Department. School of Engineering, Bogot ´ a, Colombia Article Inf o Article history: Recei v ed Aug 20, 2020 Re vised Jan 02, 2021 Accepted Jan 20, 2021 K eyw ords: Artificial intelligence Image processing Machine learning Melanoma Skin cancer ABSTRA CT In the last years, computer vision systems for the detection of skin cancer ha v e been proposed, especially using machine learning techniques for the classification of the disease and features based on the ABCD dermatology criterion, which gi v es infor - mation on the status of the skin lesion based on static properties such as geometry , color , and te xture, making it an appropriate criterion for medical diagnosis systems that w ork through images. This paper proposes a no v el skin cancer classification sys- tem that w orks on images tak en from a standard camera and studies the impact on the results of the smoothed bootstrapping, which w as used to augment the original dataset. Eight classifiers with dif ferent topologies (KNN, ANN, and SVM) were com- pared, with and w ithout data augmentation, sho wing that the classifier with the highest performance as well as the most balanced one w as the ANN with data augmentation, achie ving an A UC of 87.1%, which sa w an impro v ement from an A UC of 84.3% of the ANN trained with the original dataset. This is an open access article under the CC BY -SA license . Corresponding A uthor: Jairo Hurtado Electronics Department Pontificia Uni v ersidad Ja v eriana. Bogot ´ a, Colombia Email: jhurtado@ja v eriana.edu.co 1. INTR ODUCTION The high le v el of sun e xposure, lo w use of sunscreens by people, and s ome en vironmental f actors lead to an increase in the number of disorders and skin diseases, including cancer . There are three main types of skin cancer: basal cell carcinoma, squamous cell carcinoma, and melanoma, being melanoma the most lethal one [1]. Rural populations in the tropics, especially in mountainous areas are particularly af fected by this disease due to the e xposition to solar radiation products of their lifestyles, skin color , and geographical location, since the UV radiation increases between 10% and 12% for each kilometer of altitude [2]. Also, as the CO VID-19 pandemic has caused limited ph ysical access to health-care pro viders, this can generate further delay treatment of melanoma producing de v astating consequences for the patients [3]. F or this reason, the adoption of computational tools in medicine is arising [4]. Melanoma has been an illness of public concern due to the rapid increase of 25.9 % between 2006 and 2016 [2] and the W orld Health Or g anization predicts that in the ne xt tw o decades, the number of people diagnosed with skin cancer will be double [5]. So that, it can be appreciated the usefulness of an algorithm that identifies malignant lesion patterns and suggests that the person go immediately to a specialist, because, if it is diagnosed early , the chance of survi ving is about 95% [6]. Besides, automatic diagnosis has sho wn to o v ercome dermatologists when recognizing either malignant and benignant lesions or a particular type of J ournal homepage: http://journal.uad.ac.id/inde x.php/TELK OMNIKA Evaluation Warning : The document was created with Spire.PDF for Python.
1358 r ISSN: 1693-6930 lesion [7]. Marco Albrecht et al. studied dif ferent computer methods for diagnosis and modeling of melanoma sho wing the helpfulness of the melanoma pattern recognition systems in order to start early treatment. In the recent years, systems oriented to the automated diagnosis of skin cancer through images ha v e been proposed [1], [7]-[17]. V ariations depends especially on the type of image that is used as input and the architecture of the system. Firstly , there are o v erall three types of images used for this purpose. Macroscopic images that are lesions tak en from standard cameras, dermatoscopic images where the images are tak en using a de vice called dermatoscope which magnifies the skin lesion making malignant patterns more visible for the dermatologist [5] and finally the least used Histo-pat hological images, which are photos of the disease using microscopic e xamination of a biopsy [18]. So that, while a system that w orks on macroscopic images may be more useful for common people, the amount of images in datasets of macroscopic lesions is v ery limited. The opposite happens with dermatoscopic images where there are man y publicly a v ailable datasets with an amount of samples of the order of thousand images. W ith re g ards to the architecture, it is used traditional machine learning with hand-crafted features or deep learning where the features are calculated automatically . In this paper an algorithm for detecting malignant patterns in a skin mole using traditional machine learning and hand-crafted features is proposed, counting with a pre-processing which reduces the shado ws in the image produced by the circularity of some parts of the body . Secondly , the skin mole is se gmented using the algorithm of unsupervised learning: Gaussian Mixture Model. After that, 70 features based on a dermato- logical criterion, which is used to diagnose melanoma skin cancer , are calculated, and finally , a classification is performed. The main contrib utions of this paper are: (i) The implementation of a no v el malignant pattern recognition system that w orks on macroscopic skin lesions images. (ii) The comparison of the performance of the Gaussian mixture model to se gment dif ferent types of skin lesions. (iii) A study on the impact of the Smoothed Bootstrap data augmentation method on the performance of dif ferent topologies of classifiers. (i v) A comparison of v arious state-of-the-art systems with dif ferent architectures and type of input images. 2. RESEARCH METHOD T o detect malignant patterns on skin lesions, the system is based on a medical criterion called the ABCD rule, this is one of the most used methods whose acron ym refers to the four parameters used in the clini- cal dermatological diagnosi s. These are Asymmetry , Border , Color , and structural dif ferences [6]. Asymmetry (A): It is generated by the uncontrolled gro wth of the lesion, because of higher le v els of melanin in dif ferent re gions and tends to ha v e an irre gular shape. Borders (B): Melanoc ytic lesions ha v e irre gular borders. In contrast, benign lesions tend to ha v e borders that f ade smoothly and are symmetric. Color (C): It is related to the e xcess melanin under the surf ace of the lesion, causing a dif ferent pigmentation in a specific re gion. Dermoscopic structur es (D): It refers to the generation of holes, points, cells, and inhomogeneity (te xture) that indicates more melanin in a gi v en re gion. The ABCD rule has been t ested in multiple studies, which ha v e documented its successful diagnostic accurac y in clinical practice. Also, has been confirmed with digital im- age analysis [19]. Ho we v er , it is a medical criterion that only can be applied to pigmented lesions, which are lesions that look lik e spots. So that, the ABCD rule can not be applied to basal cell carcinoma nor squamous cell carcinoma [1]. F or this reason, the system uses the ABCD rule aiming to recognize only benignant lesions and melanoma. In the implementation, it w as used the Dermatology Education atlas [20], which contains 173 images of macroscopic skin lesions of tw o types, melanoma (84) and benignant (89), with sizes from 154 by 186 to 1129 by 1241. This dataset w as used to train the system. Ho we v er , it is clear that it is not as lar ge to ensure statistical significance, ho we v er , since datasets of macroscopic images tend to be sm all, pre vious w orks ha v e had to deal with these situat ions with methods such as data augmentation [8]-[11]. In Figure 1 the block diagram of the proposed system is sho wn. The entire system w as implemented in Python, using the OpenCV library for the pre-processing st ep as well as for the feature e xtraction. On the other hand, for the classification, it w as used Scikit-learn and T ensorflo w 2 libraries. Each block of the diagram of Figure 1 is e xplained belo w . 2.1. Pr e-pr ocessing In this block, the shado ws, caused by the curv ature of some body parts, that can af fect the system performance are attenuated. Since there could be shado ws that resemble the color of the skin mole, making a nonrecognition between shades and mole, Figure 2. T o correct this problem, another image, obtained from the re gression of the v alues near the corners of the original image, is created. TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 4, August 2021 : 1357 1368 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1359 Figure 1. Block diagram of the system Figure 2. Inaccurate se gmentation caused by shado ws T o use this method of attenuation of shado ws, there are tw o assumptions. The shado ws change smoothly and the mole i s located in the center of the image. The first assumption ensures that a tw o poly- nomial de gree adequately fits the shado ws of the image. On the other hand, the second assumption ensures that samples can be tak en at the corners of the image without touching the mole, samples that are subsequently used in the re gression. The data tak en for t he re gression is obtained from the channel V (V alue) of the HSV color space and the tw o-de gree polynomial is presented as sho wn in (1): z ( x; y ) = p 1 x 2 + p 2 y 2 + p 3 xy + p 4 x + p 5 y + p 6 (1) The six constants that minimize the squared error gi v en as sho wn in (2) ha v e to be found, where ( i; j ) are the inde x es of all samples in the corners ’S’; Sho wn in Figure 3, V is the v alue channel of the space HSV and Z is the quadratic function that will be found. The result is sho wn in Figure 4. l = X j 2 S X i 2 S [ V ( j ; i ) z ( x j ; y i )] 2 (2) Figure 3. Samples are tak en from the corners used in the re gression (a) (b) Figure 4. These figures are; (a) v alue component of the HSV space; (b) image after re gression T o attenuate the shado ws, the v alue component of the HSV space obtained abo v e is di vided by the quadratic polynomial found and is multiplied by the ratio of the a v erage of the V v alues with the a v erage of the v alues of V / z function as sho wn in (3). Finally , after c hanging the channel V of the original image in the HSV space for the ne w one found, it is passed to the RGB color space, Figure 5. A mac hine learning appr oac h for the r eco gnition of melanoma skin cancer on ... (J air o Hurtado) Evaluation Warning : The document was created with Spire.PDF for Python.
1360 r ISSN: 1693-6930 V new ( x; y ) = V V = Z V ( x; y ) z ( x; y ) (3) (a) (b) Figure 5. These figures are; (a) original image; (b) pre-processed image 2.2. Segmentation The purpose of this step is to detect the skin mole automatically based on the color distrib ution of the image. Because images of skin lesions ha v e near tw o clusters, light and dark colors, equi v alent to back- ground and skin lesion respecti v ely , the Gaussians mi xture model (GMM) can adequately describe the color distrib ution of the image and get the parameters of each of the tw o clusters in order to perform a pix el-wise classification. Also, GMM has sho wn to be capable of recognizing skin diseases with satisf actory ef ficienc y [21]. Another reason to choose the Gaussian Mixture Model is that color -based clustering has been com- pared to other methods such as Gra p h- Cut Se gmentation and Otsu, sho wing the best classification accurac y [2]. Besides, Pedro Pereira et al. [3] has compared the performance of 39 se gmentation methods across three dif ferent lar ge datasets concluding that these methods: Local Binary P atterns Clustering, W u Quantifier , and Color Based Clustering had the best o v erall performance. After finding the tw o clusters, the one with the lar gest area is classified as background and the other one as the lesion. So that, the pix els of each cluster are labeled as 0 and 1 respecti v ely , generating the se gmentation mask, sho wn in Figure 6 (a), which after filling blackheads and dilate the image sho w in Figure 6 (b), Figure 6 (c) is generated, which gi v es shape information, and Figure 6 (d) is obtained by multiplying the mask by the original image. The first is useful for e v aluating the asymmetry of the mole and the second i t can be e v aluated, borders, v ariation in color , and te xture presence. (a) (b) (c) (d) Figure 6. These figures are; (a) GMM Clustering, (b) Filling holes, (c) Dilated mask, (d) Skin mole se gmented T o measure the accurac y of the proposed s e gme ntation method on a dif ferent type of skin lesions that apparently w ould be dif ficult for the system to recognize, such as lesions with high color v ariation (ne vus spilus, ne vus repigmented, a n d some melanomas), Caf ´ e-au-lait macule, which tends to ha v e blurred shape, and lesions containing hair , see Figure 7, the Border Error (BE), sho wn in (4), which has been used in pre vious w orks to compare the se gmentation ef ficienc y [22], is calculated. Where SM is the se gmentation mask, calculated automatically through GMM, and GT is the ground-truth which w as hand-labeled from the dataset. So that, the BE measures the percentage of the non-o v erlapping area between the se gmentation mask and the ground-truth. l B E ( S M ; GT ) = Ar ea ( S M GT ) Ar ea ( GT ) (4) TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 4, August 2021 : 1357 1368 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1361 (a) (b) (c) (d) Figure 7. These figures are; (a) Caf ´ e-au-lait macule; (b) Lesion with color v ariation; (c) Containing hair; (d) Lesion with color uniformity and sharp borders T able 1 sho ws the a v erage BE for the four dif ferent lesions of Figure 7, where the lo wer the Border Error the more accurate the se gmentation is. Therefore, it suggests that the Caf ´ e-au-lait macule is the type of lesion that the se gmentation system w orks better with. On the other hand, images that contain hairs are not correctly se gmented, due to the similarity in color between the skin lesion and hair . F or this reason, nine haired samples were remo v ed from the original dataset. So that, the classifiers were trained with 164 samples, 85 benignant lesions and 79 malignant. T able 1. Border Error (BE) for dif ferent type of lesions Measure Caf ´ e-au-lait macule Containing hair W ith color v ariation W ith color uniformity B E (%) 31.1 770.2 67.1 49.0 2.3. F eatur es based on the ABCD criterion The skin lesion characterization is made using the criterion of the ABCD, as it gi v es information on the state of pigmented skin lesions using static parameters. The process to obtain these features is sho wn belo w . 2.3.1. F eatur es based on asymmetry T o quantify the mole asymmetry , it is considered Figure 6 (c), and com pared t o ot her geom etric figures, Figure 8, as the ellipse; in purple(proposed in this paper), as if the contour or shape of the mole resembles an ellipse is less lik ely to be malignant. A comparison with the bounding box is also made; in red, to ha v e the dimensions of the lesion, and the Con v e x Hull; blue dotted. Besides, the area of the quadrants of the mole must be the same if the mole is completely symmetrical. Figure 8. Geometric parameters of the skin lesion The parameters used are, b p and a p ; minor and major ax es. A p ; A c ; A b and A e ; areas of the lesion, con v e x hull, bounding box and ellipse, respecti v ely , and P p ; P ; A b and P e ; perimeter of the lesion, con v e x hull, Bounding Box, and ellipse respecti v ely . On the other hand, A 1 and A 2 represent the areas of each di vision of the axis a p . Similar to B 1 and B 2 for the axis b p . The asymmetry features are presented in T able 2. F or the features not to depend on the size and resolution of the image, those ha ving units of area were di vided by the area of the Bounding Box A b and with length units were di vided by its perimeter P b . A mac hine learning appr oac h for the r eco gnition of melanoma skin cancer on ... (J air o Hurtado) Evaluation Warning : The document was created with Spire.PDF for Python.
1362 r ISSN: 1693-6930 T able 2. Features of asymmetry Lesion area Solidity Equi v alent diameter Con v e x Hull Area A p A p A c q 4 A p A c Circularity Lesion perimeter Aspect ratio Con v e x Hull perimeter 4 A p P p 2 P p b p a p P c Aspect ratio Elliptic area rate Elliptic rate Rate of areas b p b b a b A p A e P p P e ( B 1 B 2) A p Rate of areas a p Rate form b p Rate form a p ( A 1 A 2) A p B 1 B 2 A 1 A 2 2.3.2. T extur e v ariation, darkness and color inf ormation image F or features based on borders, color uniformity , and dermoscopic structures, a ne w image ( I N ) , with three channels, obtained from the original image is created 4. The first channel pro vides information on the te xture v ariation, the second on the skin darkness, and the third, information on the color v ariation, I i N ( i = 1 ; 2 ; 3) . In order to calculate the te xture v ariation channel ( I 1 N ) , the brightness image L is obtained from as sho wn in (5), where the three channels of the original image I C are a v eraged. l L ( x; y ) = P 3 i =1 I C i 3 (5) The te xture ( x; y ; ) is defined by as sho wn in (6) where S ( x; y ; ) = L ( x; y ) G ( ) is the bright- ness smoothed by a Gaussian filter with standard de viation . l ( x; y ; ) = L ( x; y ) S ( x; y ; ) L ( x; y ) (6) The te xture image ( x; y ; ) is calculated for dif ferent v alue s of = ( 1 ; : : : ; N ) , and it is selected for each pix el the highest te xture among all scales, this is sho wn in (7). F or this paper , the standard de viation w as chosen as = 1 ; 11 7 ; : : : ; 43 7 with a windo w of 7 by 7 . The v alues of this parameter were suggested by Ca v alcanti et al. [10] due to the a v erage size of the images in the dataset [5]. ( x; y ) = max [ ( x; y ; )] (7) The te xture v ariation channel will be obtained from the normalization of ( x; y ) , sho wn in (8), where the minimum among all v alues is subtracted from the te xture image, and then di vided by the dif ference between the maximum and minimum, causing that all data is in t he interv al [0, 1]. The original image as sho wn in Figure 9 (a) and the result of I 1 N is sho wn in Figure 9 (b). I N 1 = ( x; y ) min max min (8) F or the darkness information image ( I 2 N ) , since health y skin tends to be reddish, when the red chan- nel of the original image is brighter means that it is part of the fund and if it is dark er injury . So that, the darkness is gi v en by the complement of the red channel of the original image, as sho wn in (9). The result of I 2 N is sho wn in Figure 9 (c). l I N 2 = 1 I C 1 (9) In the color information channel ( I 3 N ) , the three color channel s of the original image I C are repre- sented in a single channel I 3 N using PCA (Principal Components Analysis), and the absolute v alue is tak en creating the image C ( x; y ) . which is then normalized as sho wn in (10). The result of I 3 N is sho wn in Figure 9 (d). TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 4, August 2021 : 1357 1368 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1363 I N 3 = C ( x; y ) C min C max C min (10) (a) (b) (c) (d) Figure 9. These figures are; (a) original image, (b) te xture v ariation, (c) darkness information, (d) color information 2.3.3. F eatur es based on borders T o quantify the v ariation of the intensity in the borders of the lesion, the magnitude of the gra dient v ector has to be calculated, as it measures if the border is sharp or soft. First, the border is obtained by subtracting the mask contracted from the mask dilated, it is important that the border is thik enough to ensure that it contains the change from lesion to the background. Then the gradient it i s calculated for the v alues on the border B . So that, the features containing information of the i ntensity v ariation of the image in the borders are calculated from the mean and v ariance of the v alues of the gradient magnitude, as sho wn in (11) and (12). l = 1 N X ( x;y ) 2 B jr I ( x; y ) (11) 2 = 1 N X ( x;y ) 2 B ( jr I ( x; y ) j ) 2 (12) Where r I ( x; y ) is the gradient v ector for the scalar field I ( x; y ) . Due to the dependenc y of the gradient v ector magnitude on the skin color of the original image, the v alue of the magnitude w ould be less if the skin color w as dark er . F or this reason, the pre viously found channels I 1 N ; I 2 N and I 3 N are used, due to the less dependenc y on this parameter . T o approximate the gradient of the image, the Sobel operator is calculated, Figure 10. The color information channel is also di vided into eight pieces, whose principal ax es are oriented in the direction of the mole, this is ensured by using the eigen v ectors, and the a v erage and v ariance of the mean of the gradients are obtained in each fragment that belongs to the border of the lesion, thus obtaining tw o ne w features for each channel. This idea is similar to the proposed in Figure 11. The features based on the borders of the skin mole are sho wn in T able 3. Figure 10. Borders and Sobel operator for I 1 N ; I 2 N e I 3 N Figure 11. Pieces of the color information channel A mac hine learning appr oac h for the r eco gnition of melanoma skin cancer on ... (J air o Hurtado) Evaluation Warning : The document was created with Spire.PDF for Python.
1364 r ISSN: 1693-6930 T able 3. Features based on borders Name Notation Quantity A v erage gradient for each channel. r I i N 3 Gradient v ariance for each channel. r I i N 3 A v erage mean gradient for the eight fragments per channel. av g ( r I N i ) 3 V ariance of a v erage gradient for the eight fragments per channel. v ar ( r I i N ) 3 2.3.4. F eatur es based on color unif ormity Color features are obtained from the images of the se gmented lesion, sho wn in Figure 6 (d), and the color information channel sho wn in Figure 9 (d). In order to remo v e color noise, both images are smoothed using a Gaussian filter . T able 4 is al so used where the tones of i nterest are sho wn. These tones are the most common colors found in dif ferent types of lesions and allo w the system to recognize particular color patterns. So that, six counte rs (C) that are increased depending on the Euclidean distance of a gi v en pix el color to the tones in T able 4 are proposed. Thus, if the color of a gi v en pix el is the closest to one of the tones of interest, the respecti v e tone counter is incremented. Also, it is proposed in this paper , adding the mean and v ariance of a ne w image obtained by calculating the Euclidean distance between the color of the original image for each pix el and the tones of interest. These ne w parameters gi v e information on ho w f ar the colors of interest are on a v erage and ho w much the y de viate from the v alues of the original image. T able 4. T ones of interest in skin lesions. Alcon et al. [23] Colour Counter (C) Red channel Green channel Blue channel White (W) c W 1 1 1 Red (R) c R 0.8 0.2 0.2 Light bro wn (LB) c LB 0.6 0.4 0 Dark Bro wn (DB) c D B 0.2 0 0 blue-gray (BG) c B G 0.2 0.6 0.6 Black (BBL) c B L 0 0 0 In order to ha v e information about the non-uniformity of the mole distrib ution color , features that depend on the location of the gi v en color distrib ution in the image are proposed. F or this, the channel of color information I 3 N is di vided into eight pieces, sho wn in Figure 11, whose main ax es are oriented in the mole direction (this is ensured by using the eigen v ectors) and the a v erage of the mean of the v alues of intensity in each fragment belonging to the lesion and its v ariance are calculated. So that, in T able 5, the features that gi v e information on the color distrib ution and uniformity of the skin mole are sho wn. Being R, G, and B the color channels of the original image whose pix els are part of the lesion. Lik e wise, for I 3 N , there will be tak en only the v alues belonging to the lesion. On t he other hand, the mean and v ariance of the data are represented with the symbols and respecti v ely . T able 5. Color distrib ution features max ( R ) max ( G ) max ( B ) min ( R ) min ( G ) min ( B ) ( R ) ( G ) ( W ) ( R ) ( G ) ( B ) ( j I c W j ) ( j I c R j ) ( j I c LB j ) f ont siz e : 14 px ( j I c D B j ) ( j I c B G j ) ( j I c LB j ) ( j I c W j ) ( j I c R j ) ( j I c LB j ) f ont siz e : 14 px ( j I c D B j ) ( j I c B G j ) ( j I c LB j ) max ( I 3 N ) ( I N 3 ) ( I N 3 ) ( R ) = ( G ) ( R ) = ( B ) ( G ) = ( B ) c LB c R c D B c B G c B L c W ( I N 3 ;; 8 ) ( I N 3 ;; 8 ) 2.3.5. F eatur es based on dermoscopic structur es Although dermoscopic structures can only be measurable using a dermatoscope, which is a de vice that enables dermatologists to ha v e a closer vie w of the skin lesion, dif ferences between benignant and malignant skin moles can be measured through macroscopic images using features based on the skin mole te xture [4]. F or this reason, the te xture channel I 1 N , Figure 9 (b),which gi v es information of the mole rugosity (Holes, points and inhomogeneity), is used in order to obtain four more features which are the maximum, minimum, mean, and v ariance of the te xture v ariation channel calculated inside the lesion. These features are sho wn in T able 6. TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 4, August 2021 : 1357 1368 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1365 T able 6. T e xture Features Feature Description min ( I 1 N ) Minimum te xture max ( I 1 N ) Maximum te xture mean ( I 1 N ) A v erage te xture v ar ( I 1 N ) V ariance of the te xture channel 2.4. Data A ugmentation In the abo v e processes, 70 features based on the medical criterion of the ABCD were calculated. W it h this, the system has enough information related to the status of the skin mole to perform a classification with labels (cancer , not cancer). Ho we v er , as the dataset used for training and testing only has 164 samples, there is not supported statis tical significance. F or this r eason, the smoothed bootstrap data cloning is used. This (re)sampling technique is based on the idea that ne w data samples can be added if these samples are distrib uted according to the same probability densit y as the real data set, resulting in a greater statistical significance [24]. In order to clone data through the smoothed bootstrapping method, a Gaussian distrib ution is a reason- able assumption [10]. This is adding to each sampl e a Gaussian error with zero mean and standard de viation ten times smaller than the de vi ation of each feature. Each ne w sample is gi v en by as sho wn in (13), where ! Y k ;n ; ! X n ; ! G k 2 R 70 and ! G k is a random v ector with normal distrib ution, ! Y k ;n is the ne w sample which is obtained from adding Gaussian noise to the current sample ! X n . ! Y k ;n = ! X n + ! G k (13) This data augmentation technique has been used pre viously by Ca v alcanty et al. [4]. Ho we v er , in their w ork the entire dataset w as augmented v e times, making that there w as al w ays in the test partition a similar sample in the training partition. This w ould mak e k-nearest neighbors the classifier with the best accurac y , ne v ertheless, those results w ould be biased due to the e xtended dataset. T o a v oid this situation, only the training partition is augmented tw o times and it is tested o n the remaining samples. Also, the results of this method for dif ferent classifiers as well as the results without data augmentation are studied. 2.5. Classification system F or this research area, there are specially three classifiers used, K-nearest neighbors (KNN), artificial neural netw ork (ANN), and support v ectors machine (SVM). These classifiers ha v e been used in pre vious w orks sho wing a dermatologist le v el accurac y [2], [4], [6], [7], for this reason, the performance of the three said classifiers, both with and without data augmentation, are compared in order to see ho w the beha vior of these classifiers change with the Smoothed Bootstrapping Data Cloning. 3. RESUL TS AND DISCUSSION Performance metrics of the classification systems are presented in T able 7. These results are generat ed using a 10-fold cross-v alidation, sho wing the accurac y; which is the o v erall system performance, specificity; ability to recognize malignant lesions and sensiti vity , ability to recognize benign lesions. Specificity becomes the most important par ameter to consider because if the system does not suggest to visit a specialist, since the image is a malignant lesion, it can endanger the patient. Therefore, specificity should be the highest possible. On the other hand, it is better if the sensiti vity has a high v alue, ho we v er , it does not represent an imminent danger to the patient. T able 7. Comparison of the performance measures of dif ferent classification systems Classifier Specificity (%) Sensiti vity (%) Accurac y (%) SVM with k ernel grade 5 15.9 100 70.6 SVM with k ernel grade 5 augmented 15.9 100 70.6 KNN, Euclidean distance, k=2 37.6 83.1 66.7 KNN, Euclidean distance, k=2 augmented 56.3 76.7 66.7 KNN, Mahalanobis distance, k=2 47.4 90 72.5 KNN, Mahalanobis distance, k=2 augmented 71.7 87 80.4 ANN 200-180-150-100-50-20-1 80.2 94.4 88.2 ANN 200-180-150-100-50-20-1 augmented 86.9 87.8 86.3 A mac hine learning appr oac h for the r eco gnition of melanoma skin cancer on ... (J air o Hurtado) Evaluation Warning : The document was created with Spire.PDF for Python.
1366 r ISSN: 1693-6930 T able 7 sho ws that the neural net w ork has the best accurac y and with the original dataset the sys tem achie v es a hi g h sensiti vity le v el. In contrast, when the dataset is augmented the sensiti vity decreases while the specificity increases making it a more balanced classifier . The problem with the measures tak en in T able 7 is that while it gi v es an estimate of the classifier performance in one (specificity , sensiti vity) point, it is better to compare the entire curv e. This method to measure the performance is called the R OC curv e, which stands for Recei v er Operating Characteristic Curv e and sho ws the dependence between sensiti vity and specificity for each classifier , making possible a comparison not only for one point b ut for the entire spectr um of v alues of sensiti vity and specificity . T o c ompare the performance of the classifier using the R OC curv e, the A UC is used, which is the area under the R OC curv e, and measures the ability of the system to recognize benignant lesions as benignant and malignant lesions as malignant. So that, the closer the A UC to 100% the more accurate the system is. The R OC curv e and A UC for the three topologies compared (SVM, KNN, and ANN) are sho wn in Figure 12, which suggests that the o v erall performance of the neural netw ork and SVM increases with data augmentation. Ho we v er , being consistent with T able 7, the neural netw ork has the best performance among the topologies compared. On the other hand, both T able 7 and Figure 12 sho w that the KNN tends to w ork better with the Mahalanobis distance than the Euclidean distance for this specific task. This is important because most approaches to skin cancer classi fication that use KNN are based on the Euclidean distance while the y could increase their performance using a Mahalanobis distance-based KNN. On the other hand, Figure 12 mak es clear that when the KNN with Mahalanobis distance i s based on augmented data, the o v erall performance drops, the sensiti vity increases and the specificity decreases. Figure 12. R OC Curv es for the three toplogies: ANN, KNN and, SVM On the other hand, o v er the years traditional machine lea rning and deep learning systems ha v e been proposed. Ho we v er , while It is not possible to ha v e a di rect comparison of these systems’ performance, since the y are trained o n dif ferent datasets, comparing the accurac y with pre vious systems gi v es information on whether the proposed approach is viable or not. Also, mak es it possible to recognize patterns present among dif ferent systems, datasets, arquitectures, and types of im ages. So that, a comparison of dif ferent state-of-the-art systems is sho wn in T able 8. W ith re g ards to traditional machine learning usually counting with pre-processing, se gmentation, feature e xtract ion, and classification. Y uheng et al. [9] in 2019 proposed a system based on an SVM classifier and 143 macroscopic images using polarization speckle imaging which allo wed the system to increase performance. Another interesting approach w as made by V erosha et al. [8] in 2019. Their system used 170 macroscopic images of skin lesions for trai ning and testing and implemented both the KNN with k=5 and SVM classifiers. Also, traditional machine learning has been used in histo-pathological images, for e xample, T akruri et al. [12] implemented a PSO-SVM h ybrid system that optimizes SVM parameters and then performs a classification. On the other hand, deep learning systems w orking on the recognition of skin cancer ha v e seen important impro v ements o v er the years. P acheco et al. [1] proposed t hat the input of a con v olutional neural netw ork (CNN) w as not only the image of the skin lesion b ut also clinical information such as age, location of the lesion and if it had bled, impro ving the a v erage accurac y in o v er 7%. Maron et al. made a comparison between 112 dermatologists and a deep learning system, sho wi n g that the con v olutional neural netw ork w as more accurate in both binary class problem (Melanoma, benignant) and multiclass (T ype of skin disease). T able 8 sho ws the result s obtained by the 112 dermatologists in order to ha v e a reference to compare the systems. Ade gun et al. [15] proposed an encoder -decoder architecture which g a v e the system the possibility of e xtracting TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 4, August 2021 : 1357 1368 Evaluation Warning : The document was created with Spire.PDF for Python.