Human Intestinal Condition Identification Based-on Blended Spatial and Morphological Feature using Artificial Neural Network Classifier

https://doi.org/10.17977/um018v3i12020p1-27 ©2020 Knowledge Engineering and Data Science | W : http://journal2.um.ac.id/index.php/keds | E : keds.journal@um.ac.id This is an open access article under the CC BY-SA license (https://creativecommons.org/licenses/by-sa/4.0/) Human Intestinal Condition Identification Based-on Blended Spatial and Morphological Feature using Artificial Neural Network Classifier

Colon cancer is a type of disease that attacks the intestinal walls cell of humans. Colorectal endoscopic screening technique is a common step carried out by the health expert/gynecologist to determine the condition of the human intestine. Manual interpretation requires quite a long time to reach a result. Along with the development of increasingly advanced digital computing techniques, then some of the weaknesses of the manually endoscopic image interpretation analysis model can be corrected by automating the detection process of the presence or absence of cancerous cells in the gut. Identification of human intestinal conditions using an artificial neural network method with the blended input feature produces a higher accuracy value compared to the artificial neural network with the non-blended input feature. The difference in classifier performance produced between the two is quite significant, that is equal to 0.065 (6.5%) for accuracy; 0.074 (7.4%) for recall; 0.05 (5.0%) for precision; and 0.063 (6.3%) for f-measure.
This is an open access article under the CC BY-SA license (https://creativecommons.org/licenses/by-sa/4.0/).

Keywords:
intestinal condition blended spatial morphological feature neural-networks classifier manual interpretation is the cause of the long duration of time needed by the patient to find out the results of endoscopic screening [7]. This image interpretation has a weakness; it depends on the gynecologist's expertise and experience [8]. Along with the development of increasingly advanced digital computing techniques, then some of the weaknesses of the manually endoscopic image interpretation analysis model mentioned previously can be corrected by automating the detection process of the presence or absence of cancer cells in the gut, by utilizing digital image processing techniques supported by the machine method learning. Automating the detection process can speeding up in production results and minimizing errors arising from the manual analysis model of endoscopic image interpretation.

II. Method
The identification of the intestinal condition to find cancerous colon condition in this study was divided into several steps presented in Fig. 1. The stages of Fig. 1 are explained in the following subsection points: The preprocessing aims out to ensure that the original image is ready for further processing at the feature extraction stage. Preprocessing also plays a vital role in avoiding bias in the output of a machine learning classifier. Preprocessing in this study is divided into several stages:

Convert Image RGB to Grayscale
At this stage, the conversion of RGB channel images into grayscale channels is carried out by using the formula presented in (1) by taking the red color conformity of 30%, green by 59%, and blue by 11% [2]. The RGB conversion extracts features of spatial images in the form of circularity, aspect ratio, triangularity, and cooccurence matrix values. (1) where R indicates the value of the red channel, G the green channel, and B the Blue channel, an example of the results of converting an RGB image of endoscopic images under normal conditions to grayscale is presented in Fig. 4.

Histogram Normalization
At this stage, the image histogram normalization is carried out which aims to equalize the brightness and contrast patterns that are owned by the image and the distribution of pixel intensities [10]. Histogram normalization is carried out using the formula presented in (2).  [11].
where i new is the value of pixel normalization mapping from a range [0-255] to range [0-1]; n i is the number of pixels in an image I(x,y) with gray level degrees of (i); while p(i) represents the pixel probability in gray level degrees of (i).
Feature extraction from endoscopic images is carried out to retrieve relevant information from each image so it can be used as input to the machine learning classifier. In this study, relevant information extracted from endoscopic imagery includes morphological information and spatial information. Morphological information, taken based on the size of circularity that is the result of the division between the pixel area value and the pixel perimeter on the region of interest (ROI). The circularity formula is presented in (3) [6]. (3) Whereas spatial feature information, extracted based on the co-occurrence matrix, which will produce information includes:

Energy
Energy is a measure of pixel conformity in an image. Energy reflects the degree of texture smoothness of an image. The lower the energy value, the rougher the surface texture of the image and vice versa [12]. The calculation of the energy value is presented in (4) ∑

Contrast
Contrast value is the simple comparison between foreground objects and image background.
Contrast is a unit of local image variation values [12] [13]. The calculation of the contrast value is presented in (5) ∑

Correlation
Correlation is a gray-level linearity value of two or more adjacent pixels in an image [14]. The calculation of the correlation value is presented in (6)

Homogeneity
It is a value of the distance between elements in the co-occurrence matrix in gray images [11]. The calculation of the homogeneity value is presented in (7) ∑ where P(i,j) is the elements of the co-occurrence matrix; μ i and μ j express the mean value and σ i, and σ j reflect the standard deviation in row i and column j in.
The machine learning model built in this study is an artificial neural network that utilizes the backpropagation function. The architecture of artificial neural networks is presented in Table 1. At the input layer of the artificial neural network architecture, 4 neurons and 5 neurons are used following the number of feature extractions from the colonic endoscopy image. While the hidden layer / intermediate layer uses 9 and 11 neurons following the equation stated by [12][15] that the use of a hidden layer of 2n+1 (where n is the number of input neurons) can accelerate the training process and the generalization results of neural networks. At the output layer, the binary-shot coding concept is used where 1 represents the condition of the cancerous image, while 0 represents the normal image condition [16]. 1. TP is a condition when the input "A" identified by machine learning matches the ground truth "A". 2. TN is a condition when the "non-A" input identified by machine learning matches the "non-A" groundtruth. 3. FP is a condition when input "A" is identified by machine learning as a "non-A" groundtruth. 4. FN is a condition when the input "non-A" is identified by machine learning as the "A" groundtruth.
From the indicators mentioned above, an equation can be formed, stating the accuracy presented in (8) until (11) [18].

III. Results and Discussions
The Matlab 2015 R programming platform that runs on the Windows 10 64bit operating system was used as the experimental base for this research. Endoscopic images used in this study have a total of 200 images with two categories: 100 endoscopic images of cancer category and 100 endoscopic images of the normal category. The dataset is divided into three parts to avoid bias on the results of artificial neural network training. The first part is the training dataset (70%), the second part is the testing dataset (15%), and the third part is the validation dataset (15%). The default function Matlab 2015R (dividerand) is used as a dataset divider.
Feature extraction produces two kinds of features: morphological and spatial features. Morphological features produce information on circularity values. On the other hand, spatial features product information on energy values, contrast, correlation, and homogeneity. The values of spatial and morphological features are used as input from the artificial neural network classifier. Some morphological and spatial extraction values are presented in Table 2.
Although there are many algorithm choices available in the artificial neural network training process [19] [20], this research uses Quasi-Newtonian (matlab: trainlm) algorithm because the Quasi-Newtonian algorithm can produce an optimal artificial neural network learning process and faster to achieve generalization of output values compared to training algorithms such as Scaled-Conjugate or Resilient-Propagation [21]. The parameters of the artificial neural network (ANN) training process are presented in Table 3.
Training and testing are carried out in the Matlab R2016a environment that runs on an operating system platform on Windows 10 (64-bit) with an Intel® Core i5® 4310 processor computer; 8 GB Memory; Intel HD VGA Card. For simplification purpose, there will be presented only the results of the artificial neural network training process with 5-(11)-2 architecture with blended input feature (spatial and morphological) are presented in Fig. 5 From Fig. 6 it can be seen that the training process does not experience overfitting conditions. The absence of overfitting is indicated by the blue line = train; green = validation; red = test that decreases simultaneously and does not intersect each other. A summary of the metrics for the results of artificial neural network training is presented in Table 4. A summary of the confusion matrix of the artificial neural network classifier for blended (spatial and morphological) input features is presented in Table 5. The summary of the confusion matrix of the artificial neural network classifier for the non-blended (spatial only) input feature is presented in Table 6.
The artificial neural network training process with a blended (spatial and morphological) input feature produces a regression value of 0.97232 presented in Fig. 6. Otherwise, the artificial neural network training process with a non-blended (spatial only) input feature produces 0.89151 regression value. Based on Table 5 and Table 6, it can be evaluated the performance of artificial neural network classifiers with blended and non-blended input features with indicators of accuracy, recall, precision, and f-measure. The values of the performance indicators artificial neural network classifiers with blended input feature, derived from Table 5 and Table 6 are presented in Table 7. For ease of use, a comparison from Table 7 also presented on Fig. 7.

IV. Conclusion
Identification of human intestinal conditions using an artificial neural network method with the blended input feature produces a higher accuracy value compared to the artificial neural network with the non-blended input feature. The difference in classifier performance produced between the two is quite significant, that is equal to 0.065 (6.5%) for accuracy; 0.074 (7.4%) for recall; 0.05 (5.0%) for precision; and 0.063 (6.3%) for f-measure. So it can be concluded that the use of blended features as neural network inputs sufficiently influences the results of identification of the condition of the human intestine.