A novel approach to classify pharyngitis using potential features based on doctor’s perspective

Abstract

Pharyngitis is an inflammation of the oropharynx’s mucous membranes. It is typically brought on by a bacterial illness. The outburst of latest technologies has created the need for remote care of detecting diseases like pharyngitis through images of throat taken with help of smart camera. In recent years, research has forwarded with help of deep learning in classifying pharyngitis. But deep learning models require at least one hour training and requires considerably large data set to get a good accuracy. In this paper, we focused on this time constraint and are proposing a novel approach PFDP to classify pharyngitis through detection of potential features based on doctor’s perspective. We have extracted the tiny portions of image which the doctor observes them as infected and calculated frequencies of the occurrences of these portions and are given to custom made decision rules. The classification results showed significant improvement in performance in terms of time taken to reach average accuracy of 70%. It has taken only 5 minutes to extract counts of infected patterns and 1 more minute to get classification results by decision rules of if-then-else rules. We have conducted the experiment on set of 800 images. Though accuracy is lesser than that of what other works achieved but time taken to extract features is significantly lower than that of previous works. Also our approach does not require training and can be applied where scarcity of dataset exists. We assure that our approach is a new direction of research and can compete with more state of the art works in future.

Keywords

Pharyngitis potential feature extraction classification machine learning

1. Introduction

The latest technological developments highlighted the essential role of remote healthcare. These solutions offer patients safe access to medical services from their homes, mitigating the risk of viral spread and ensuring ongoing care, especially for those with throat related diseases. With the widespread use of smartphones, leveraging their capabilities for capturing throat or mouth images is feasible. These snapshots can be input into specialized processing systems or software, enabling the classification of conditions like pharyngitis, marking a promising approach to accessible healthcare diagnostics.

The contagious nature of the virus, coupled with the vulnerability of ill patients, necessitates innovative approaches to minimize risks. Remote healthcare, facilitated by advancements in technology, offers a crucial solution. In the case of diseases like pharyngitis, where symptoms may overlap with COVID-19, remote diagnosis and treatment play a pivotal role. Pharyngitis, often associated with respiratory infections, shares symptoms with COVID-19, such as sore throat and difficulty swallowing. Utilizing remote health solutions, individuals can capture images of their throats using smartphones, allowing healthcare providers to assess and diagnose pharyngitis symptoms without direct physical contact.

This not only minimizes the risk of viral spread but also addresses the challenges posed by symptom overlap between respiratory illnesses. Remote healthcare proves crucial in promptly identifying and treating pharyngitis, preventing its progression, and reducing the burden on healthcare facilities. Advanced technologies, including machine learning algorithms, enhance diagnostic accuracy, making remote pharyngitis diagnosis more efficient and reliable. Overall, in the COVID-19 context, remote healthcare acts as a vital tool for both preventing and managing pharyngitis, ensuring timely intervention and safeguarding public health.

Pharyngitis manifests with symptoms such as sore throat, difficulty swallowing (dysphagia), and inflammation of the tonsils and the back of the throat. It is commonly caused by viral infections like the flu or the common cold, though bacterial infections, particularly by Streptococcus pyogenes, can also be responsible. The condition may lead to discomfort, pain, and irritation in the throat, impacting daily activities. In more severe cases, untreated bacterial pharyngitis can result in complications such as the spread of infection to neighboring areas, like the middle ear or sinuses.

According to the Health Report of U.S National Statistics, streptococcal pharyngitis, commonly known as strep throat, constitutes a significant driver for patient admissions to hospital emergency departments in the United States [1]. This bacterial infection, primarily attributed to Group A beta-hemolytic streptococcus, poses a substantial health risk for both children and adults [2, 3, 4]. In 2010, 1.814 million people with Pharyngitis, with 692,000 children under 15, visited the emergency room. The majority of pharyngitis cases occur in children under the age of five.

Delayed diagnosis of strep throat elevates the potential for rheumatic fever, a precursor to chronic rheumatic heart disease, resulting in approximately 320,000 global fatalities annually [5, 6, 7]. Consequently, early identification of strep throat assumes paramount importance, particularly in medically underserved remote areas, as it serves as a pivotal preventive measure against fatalities stemming from rheumatic heart disease. Furthermore, an erroneous diagnosis of strep throat may precipitate unwarranted antibiotic treatments, contributing to the peril of bacterial resistance [8, 9].

Figure 1.

Overall architecture of pharyngitis detection.

Chronic pharyngitis, characterized by persistent inflammation, may contribute to the development of more serious conditions, including tonsillitis or contribute to the risk of developing respiratory issues. Regular medical assessment and appropriate treatment are essential to manage symptoms effectively and prevent potential complications associated with pharyngitis. The overall architecture is depicted in Fig. 1. The various sections of this paper are arranged as follows. In Section 2, we proposed the literature survey. In Section 3, we discussed proposed methodology. Section 4 outlines the results of proposed approach. Section 5 concludes the paper.

2. Literature survey

The growing need for remote healthcare, fuelled by the vulnerability of ill patients and the recent challenges posed by the COVID-19 pandemic, has spurred research into efficient methods for diagnosing and treating diseases remotely. In the context of pharyngitis, a common upper respiratory tract infection affecting millions, leveraging smartphones for remote diagnostics has gained attention.

The prevailing diagnostic approach involves clinical decision-making employing the Centor score, derived from a comprehensive set of criteria encompassing parameters such as coughing and fever [2, 3, 5, 7, 8, 10]. An alternative clinical diagnostic method for detecting streptococcal pharyngitis is throat culture, wherein a sample of throat cells is introduced into a medium conducive to bacterial growth, facilitating disease identification [9, 11, 12, 13, 14, 15, 16]. A positive bacterial growth signifies a bacterial infection with a diagnostic accuracy of 98%, while the absence of growth indicates the absence of a bacterial infection [15]. Touch spray ionization mass spectrometry has also been employed for strep throat diagnosis [14]. Nevertheless, these diagnostic modalities necessitate the expertise of trained physicians or specialists, presenting an enduring challenge in achieving timely and universally accessible diagnoses for all patients.

Research endeavours have harnessed color intensity values to discern a spectrum of maladies, ranging from diabetes [17, 18] and internal-organ disorders [19, 20, 21] to afflictions of the heart and kidneys [17, 18, 22, 23, 24, 25, 26, 27, 28, 29]. These methodologies, predicated on color intensity values, have been intricately amalgamated with machine learning paradigms, including naive Bayes, Bayes net, and sequential minimal optimization (SMO) [30, 31, 32]. In these investigations, 21 distinct properties were derived from tongue color intensity values to effectively diagnose a diverse array of 23 diseases. Despite the discerning capacity exhibited in diagnosing various afflictions through tongue color features, certain limitations persist in the identification of syndromes, discrimination of color features, and accurate disease classification [17, 22, 23, 24]. Evidently, factors such as disparate lighting conditions, color spaces, and devices, as emphasized by Zhang and Kim et al., introduce variability that compromises the reliability of these aforementioned methods in the accurate diagnosis of corresponding diseases [17, 33, 34].

The previous study aimed to capture throat images by combining smartphones with supplementary equipment and utilized the k-nearest neighbor algorithm in the color component distribution space to categorize streptococcal tonsil images [35]. Streptococcus pyogenes is the leading bacterial cause of pharyngitis, accounting for 20–30% of sore throat cases [36, 38]. In another study [3], CycleGAN-generated synthetic images exhibited practical pharyngitis characteristics. A deep learning model notably improved pharyngitis diagnosis accuracy using these synthetic throat images. In the evaluation of pharyngitis identification on the test dataset, the ResNet50 model with GAN-based augmentation achieved the highest ROC-AUC of 0.988. Additionally, using the ResNet50 model in a four-fold cross-validation approach resulted in a maximum classification accuracy of 0.953 and a ROC-AUC of 0.992.

In contemporary pursuits within the realm of image classification or categorization, there has been a notable integration of features rooted in the Bag of Visual Words (BoVW) paradigm, as evidenced by various scholarly endeavours [40, 41, 42, 43]. A subset of these initiatives has strategically employed the principles of transfer learning. Transfer learning, in the context of machine learning, denotes a paradigmatic approach wherein a pre-existing model, having undergone prior training, is harnessed to enhance performance on a cognate yet distinct task [44, 45, 46].

Figure 2.

Proposed Methodology of PFDP.

3. Proposed methodology-potential features based on doctor’s perspective (PFDP)

The framework of proposed methodology PFDP is given in Fig. 2. As part of proposed approach, we need to identify 5 potential candidate images of pharyngitis. Then, we need to extract small infected portions from these images and call these portions as potential feature templates and these are given in Step-1 of Fig. 2. Then, we need to find minimum pixel value and maximum pixel values for each of feature template as given in step-2. As part of step-3, we extract specific portion of image and let it be ImageP. Then, we check whether the pixel values of ImageP fall in the range of minimum pixel value to maximum pixel value of particular feature template. If so, the match count for that feature template is incremented. This process is given in step-4. Then, as part of step-5, we finalize the match counts for each of feature templates. Then, these match counts are given to simple if-then-else rules for classification. As shown in Fig. 3(e), the potential feature templates are extracted by snipping a squared portion from all 5 candidate images for Pharyngitis. In this same fashion, all 5 feature templates of pharyngitis class were extracted from 5 potential candidate images and are shown in Fig. 4. All these feature templates were of size 13 $\times$ 13 and are shown as zoomed in Fig. 4.

Figure 3.

Potential candidate images for Pharyngitis.

Figure 4.

Potential feature template images for Pharyngitis.

Consider the sample image given in Fig. 5. As teeth color is matching with the color of white infected portions, we snip the image in the form of 20 columns as left margin and 96 columns as right margin and also of 20 rows as top margin and 96 rows as top margin as shown in Fig. 5. This is the portion ImageP which we discussed earlier.

Algorithm 1: find_min_max_values_PFDP(path, numtemplates, minvaluesbgr, maxvaluesbgr):
	Input:
	path:path of feature template images,
	numimages:number of images in path path,
	Output:
	minvaluesbgr:minimum pixel values of 3 channels of 5 features,
	maxvaluesbgr:maximum pixel values of 3 channels of 5 features
1.	maxvaluesbgr $\leftarrow$ np.zeros((5,3))
2.	minvaluesbgr $\leftarrow$ np.zeros((5,3))
3.	templates $\leftarrow$ list of files from path path.
4.	for i in range(0,numtemplates) a. image $\leftarrow$ read image at templates[i] b. if image is None c. print Image is empty. d. templateb $\leftarrow$ image[:,:,0] e. templateg $\leftarrow$ image[:,:,1] f. templater $\leftarrow$ image[:,:,2] g. maxvaluesbgr[i,0] $\leftarrow$ np.max(templateb) h. maxvaluesbgr[i,1] $\leftarrow$ np.max(templateg) i. maxvaluesbgr[i,2] $\leftarrow$ np.max(templater) j. minvaluesbgr[i,0] $\leftarrow$ np.min(templateb) k. minvaluesbgr[i,1] $\leftarrow$ np.min(templateg) l. minvaluesbgr[i,2] $\leftarrow$ np.min(templater)

The proposed work comprises of three algorithms. The process of step-1 of Fig. 2 is given in Algorithm-1 whose name is find_min_max_values_PFDP. In Algorithm-1, the minimum and maximum pixel values of blue, green and red color channels of each template feature are calculated. In 1^st and 2^nd lines, the arrays minvaluesbgr and maxvaluesbgr are initialized to zeros. In 3^rd line, the list of file names were extracted from the path where feature templates are stored. In lines numbered 4 to 16, we read each image of feature template and extract blue, green and red components into templateb, templateg and templater respectively. Next, we extract max and min values from each of these channel specific template matrices and store it in maxvaluesbgr array and minvaluesbgr array.

Figure 5.

Selecting only the particular portion of source image.

Algorithm 2: find_num_matches_phary_nonphary_PFDP(path,numimages,vector, labels,classnum,width,height):
Input: path:path of test data set of phayrngitis or non pharyngitis,
numimages:number of images in specified path,
vector: vector of dimension numimages X 5 initialized to zeros.
labels $\sim$ variable to hold labels
classnum $\sim$ class numer(1 for pharyngitis and 0 for non pharyngitis)
width $\sim$ The width of portion of image in which feature template is to be searched.
height: The height of portion of image in which feature template is to be searched.
Output: Classification result parameters such as Accuracy, Precision
1.	files $\leftarrow$ list of files from path path.
2.	for $i=1$ to numimages do
3.	image $\leftarrow$ read image at files[i]
4.	if image is None
5.	print Image is empty.
6.	labels[i] $\leftarrow$ classnum.
7.	imageb $\leftarrow$ image[20:160,20:160,0]
8.	imageg $\leftarrow$ image[20:160,20:160,1]
9.	imager $\leftarrow$ image[20:160,20:160,2]
10.	prend $\leftarrow$ 140-width
11.	pcend $\leftarrow$ 140-height
12.	for $j=$ 0 to 4 do
13.	for $r=$ 0 to prend do
14.	for $c=$ 0 to pcend do
15.	rend $\leftarrow$ $r+$ width
16.	cend $\leftarrow$ $c+$ height
17.	portionb $\leftarrow$ imageb[r:rend,c:cend,0]
18.	portiong $\leftarrow$ imageb[r:rend,c:cend,1]
19.	portionr $\leftarrow$ imager[r:rend,c:cend,2]
20.	if(np.minf()(portionb) $\geqslant$ minvaluesbgr[j,0]and a. np.maxf()(portionb) $\leqslant$ maxvaluesbgr[j,0] and b. np.minf()(portiong) $\geqslant$ minvaluesbgr[j,1] and c. np.maxf()(portiong) $\leqslant$ maxvaluesbgr[j,1] and d. np.minf()(portionr) $\geqslant$ minvaluesbgr[j,2] and e. np.maxf()(portionr) $\leqslant$ maxvaluesbgr[j,2]) then
21.	vector[i,j] $\leftarrow$ vector[i,j]+1
22.	endif
23.	endfor
24.	endfor
25.	endfor
26.	endfor

The Algorithm-2 is given the name find_num_matches_phary_nonphary_PFDP that comprises of steps to extract final counts of cases where feature templates are matched inside the image. The first step is to read all set of file names of images from a train or test pharyngitis or non-pharyngitis images. Then pick the first image and read into array format. If image read is empty, then report it as empty. The “classnum” parameter is for class labels being 1 for pharyngitis and 0 for non-pharyngitis. The labels variable for ith image is set accordingly. Then lines 7, 8 and 9 extract blue, green and red components of ith image with portion of [20:160][20:160] for rows and columns respectively. In line numbers 10 to 17, for each feature template, we are extracting the portion of width and height from which the pixel values are retrieved and searching if these pixel values fall between the range of minimum value and maximum value of feature template. This is done for all three color channels of blue, green and red. We count as 1 occurrence of feature matching when all these pixel values in selected portion of the image are in range of minimum and maximum values of all three color channels of feature template.

Algorithm 3: Mainfunction_PFDP():
1.	test_vector_phary $=$ np.zeros((800,5))
2.	test_vector_nonphary $=$ np.zeros((800,5))
3.	test_phary_labels $=$ np.zeros((800))
4.	test_nonphary_labels $=$ np.zeros((800))
5.	testphary $=$ “./gdrive/MyDrive/Data_Pharyngitis/PHARYNGITIS/PHARYNGITIS_UNZIPPED/”
6.	testnonphary $=$ “./gdrive/MyDrive/Data_Pharyngitis/NO_PHARYNGITIS/NO_PHARYNGITIS_UNZIPPED/”
7.	find_num_matches_phary_nonphary_PFDP(testphary,200,test_vector_phary, test_phary_labels,1)
8.	find_num_matches_phary_nonphary_PFDP(testnonphary,200,test_vector_nonphary, test_nonphary_labels,1)
9.	finalfeatures=np.concatenate((test_vector_phary,test_vector_nonphary),axis $=$ 0)
10.	testlabels $=$ np.concatenate((test_phary_labels,test_nonphary_labels),axis $=$ 0)
11.	testcalclabels $=$ np.zeros((1600,1))
12.	for i in 0 to 400
13.	count $=$ 0
14.	if(finalfeatures[i,0] $>$ 0 or finalfeatures[i,1] $>$ 0 or finalfeatures[i,2] $>$ 0 or finalfeatures[i,3] $>$ 0 or finalfeatures[i,4] $>$ 0 or finalfeatures[i,5] $>$ 0)
15.	count $=$ count $+$ 1
16.	if(count $\geqslant$ 1)
17.	testcalclabels[i] $=$ 1
18.	else
19.	testcalclabels[i] $=$ 0
20.	endif
21.	endfor
22.	correct $=$ 0
23.	for i in 0 to 1600
24.	if(testlabels[i]==testcalclabels[i] )
25.	correct $=$ correct $+$ 1
26.	endif
27.	endfor
28.	accuracy $=$ correct/1600
29.	print accuracy

Before we proceed to main functio, we assume that we called the function find_min_max_values_PFDP with necessary arguments supplied and calculated the min and max pixel values of feature templates. Main function with name Mainfunction_PFDP for the proposed work is given in Algorithm 3 above. The first two lines in Algorithm 3 are used to initialize vectors holding feature counts. The 3rd and 4th lines are used to initialize values for labels. The fifth and sixth lines are used to hold the paths for pharyngitis and non-pharyngitis images. The 7th and 8th lines call the function find_num_matches_phary_nonphary_PFDP() function with specified arguments supplied. The 9th line concatenates the both features of pharyngitis and non-pharyngitis images and 10th line concatenates the labels of respective images. The 11th line initializes the vector for holding the labels after applying the proposed technique. The line numbers 12th to 21st set the label of test image to be classified to 1(pharyngitis) if the count in any of potential feature is greater than zero or else to 0(non-pharyngitis). The line number 22nd initializes the correctly classified images count to 0. The line numbers 23rd to 27th lines compare the true label with classified label and if it is equal they are equal the count for correctly classified label is incremented and line number 28th calculates the percentage of correctly classified images out of 1600 images taken in the experiment. The line number 29th prints the accuracy.

4. Data set collection

The selection of particular width and height of image is not taken into consideration inside the experiment. We simply assumed it to be 256 $\times$ 256 which may appear comparatively large. But for our model as it requires low computational resources, we are justifying the selection of width and height to be 256 $\times$ 256. Through unsupervised learning between two domains, CycleGAN, a ground-breaking development in the field of generative adversarial networks, transforms image-to-image translation. It offers unparalleled versatility in picture manipulation and style transfer thanks to its novel cyclic consistency loss, which facilitates smooth conversion across various visual domains without the requirement for paired training data. The basic dataset used is taken from previous works [38]. Next, we applied basic augmentation operations such as zooming from 0% to 20%, width and height translation from $-$ 5% to $+$ 5%, left and right flip, random rotation from $-$ 10^∘ to 10^∘ and random brightness change from 10% to 10%. Next, we applied Cycle GAN technique to increase number of samples of each class [39, 47]. Table 1 shows the details of accumulation of data set with regard to different augmentation techniques.

5. Experimental results

The experiment was conducted in Google Colab environment with CPU resource and implemented in python open source framework and details are given in Table 2. For our proposed work there is no need of GPU resources and CPU resources are enough. As the potential feature templates involve the combination of Red, Green and Blue components, we have taken the choice of considering RGB information compared to grey scale.

Table 1
Pharyngitis dataset accumulation

Data	Class	Number of test set(%)
Raw Data	Pharyngitis	104(80)
	Non-Pharyngitis	160(80)
Basic Augmentation	Pharyngitis	416(80)
	Non-Pharyngitis	640(80)
CycleGAN-based Augmentation	Pharyngitis	800(80)
	Non-Pharyngitis	800(80)

Table 2

Simulation parameters of experiment

S.No	Simulation environment	Client machine parameters
1	Google Colab-CPU	Windows-10,
		4GB RAM,
		Good Internet Connectivity

Table 3

Model applicability on scarcity of dataset

S.No	Author [Citation]	Applicable for scarcity of data set or not
1	Askarian B et al. [35]	No
2	Yoo et al. [38]	No
3	PFDP	Yes

Table 4 shows the values of accuracy over size of template matching area. The accuracy was calculated on performing test on 1600 images of 800 per each of two classes. The accuracy was projected with regard to varying sizes of width and height in Algorithm 2. The highest accuracy of 72.56% was achieved when width and height are set to 2. The standard deviation of size of template matching area is taken as 2. If this size is lesser, then we have deeper scanning of the image for detection of potential features. On the other hand, if this size is larger, then the happening of relatively weak features may affect the accuracy. Due to this reason, we limited the maximum size of template matching area to 10 $\times$ 10. The more the number of feature templates taken, the more the focus of the potential features inside the experiment. But in some non-pharyngitis images, some of the potential feature templates may exist in low frequency so that their contribution in some of the images may not lead to fruitful classification. The proper threshold on contribution of some of potential feature templates will only bring good accuracy of the model with regard to increasing number of potential feature templates.

Table 4

Accuracies of classification over width and height of template matching area (PFDP)

S.No	Size of template matching area	Accuracy
1	2 $\times$ 2	72.56
2	4 $\times$ 4	71.14
3	6 $\times$ 6	70.28
4	8 $\times$ 8	69.12
5	10 $\times$ 10	68.62

Table 5

Accuracies of classification over starting row number of test Image (PFDP)

S.No	Size of starting row number of test image	Accuracy
1	20	72.56
2	30	71.34
3	40	70.28
4	50	68.12
5	60	67.52

Figure 6.

Accuracy over width or height of template matching area (PFDP).

Figure 7.

Accuracy over starting row of portion of test image (PFDP).

Table 6

Accuracies of classification over feature combination (PFDP)

S.No	Feature combination	Accuracy
1	1	69.23
2	2	70.34
3	3	71.12
4	4	68.12
5	5	69.42
6	1 and 2	71.23
7	3 and 4	72.11
8	3, 4 and 5	71.24
9	1, 2, 3, 4, 5	72.56

The introduction of weights to feature templates may lead to the sensitive nature of model’s performance. The model’s accuracy may vary as the weights are changed meticulously. The exhaustive study of features of image is required to employ weights and currently in our work weights are not given to feature templates.

The width and height is varied starting from 2 $\times$ 2 portion of image with a standard deviation of 2 and ending with 10 $\times$ 10 and projected the accuracies in graph in Fig. 6. In Algorithm 2 we have taken as [20:160] for both width and height in line numbers 7 to 9. We have tested the accuracy with width varying from [20:160] to [60:200]. The resulting graph is projected in Fig. 7. Table 5 shows the values of accuracy over size of starting row number of test image. The starting row number for test image is taken out to be 20 due to the reason that teeth of the patient should not be in potential portion for feature detection. Also teeth colour interferes with that of white patches that indicate infection of pharyngitis. Based on the teeth structures of various patients this gap is taken to be varied in range of 20 to 60 with standard deviation of 10. Though the accuracy is low compared to the accuracy that can be achieved by deep learning models, our proposed approach PFDP takes 0 minutes for training and can be applied for domains where scarcity of data set exists. We are hopeful that efforts to be developed in this direction in future will be more promising in all experimental parameters.

Table 6 shows the accuracies of classification over feature combination. We also plotted the each feature contribution towards accuracy in Fig. 8. The data set for pharyngitis is so typical or the very nature of pharyngitis detection is tough so that the contribution of combination of features plays vital role in showing different accuracies. The intuition behind this makes the author to test the combinations for computing accuracy. As we have limited feature templates and so is the limited number of feature combinations. We have taken 9 combinations of features. The first 5 combinations are five feature templates taken alone and 6th combination is contribution of feature 1 and feature 2 combined and 7th combination is contribution of feature 3 and 4 combined and 8th combination is contribution of features 3, 4 and 5. The 9th combination is contribution of all 5 features taken as combined.

Table 7

Time taken for training in various works

S.No	Author [Citation]	Training time (Hours)
1	Askarian et al. [35]	1
2	Yoo et al. [38]	12
3	PFDP	0

Figure 8.

Accuracy over feature combination (PFDP).

Table 8

Comparitive table for accuracy, precision and recall of various works

S.NO	Author [Citation]	Accuracy (%)	Precision (specificity) (%)	Recall (sensitivity) (%)
1	Askarian et al. [35]	93.75	88	87.50
2	Yoo et al. [38]	92.50	94.20	89.80
3	PFDP	72.56	72.51	71.49

Figure 9.

Training time versus model graph.

Figure 10.

Accuracy versus model.

The time taken for training the model on the data sets in various works is given in Table 7. It has taken around 1 hour for training of the model in work [35]. It has taken around 12 h for training for 1000 epochs in work [38]. Our proposed approach took 2 to 3 minutes which is very minute in proportion when compared to other works. The very nature of image classification based on potential features does not require training of the model. It is all about detecting potential features on test images. So our proposed model PFDP does not require any training of the data. In our proposed work, there exists no training of model on images. Instead, we directly apply classification model on test data set. Also, the proposed method PFDP can be applied to other domains where scarcity of data sets exists. The graph showing the time taken for training versus model is given in Fig. 9.

The accuracy, precision and recall of various works with regard to pharyngitis detection are projected in Table 8. In the work of Askarian [35], they achieved 93.75% accuracy and 88% precision and 87.5% recall. In another work of Yoo [38], they achieved 92.5% accuracy, 94.2% precision and 89.8% recall. In the proposed work, we achieved 72.56% accuracy, 72.51% precision and 71.49% recall.

The graph plotted for accuracy versus model is given in Fig. 10. Though accuracy, precision, recall are lesser than that of what other works achieved, we can confidently say out proposed approach can save lots of training time and achieved reasonably good accuracy. The future works in this direction can undoubtedly improve accuracy and other parameters with less training time and can be applied even in the cases where scarcity of data sets exist. The extracted potential features can also be given to CNN instead of custom made if-then-else rules. But in this scenario, as CNN takes comparatively more time than custom-made if-then-else rules, custom made rules will suffice the achievement of good performance of model.

6. Conclusion and future work

In this Paper, we have devised a new approach of classifying pharyngitis based on matching of potential feature templates in the test images. We have taken varying sizes of portions of image and varying portions of feature template matching areas in our experimentation. The matching is performed so that occurrence of a match is counted when pixel values of test image fall below the limits of min and max pixel values of feature templates of all three color channels named blue, green and red. We projected the graphs with regard to varying sizes of template matching area and portions of test image that is being considered. The highest accuracy of 72.56 is achieved with size of 2 $\times$ 2 size of template matching area. In this case precision is 72.51 and recall is 71.49. This is the accuracy when starting row of test image is taken with 20 and that of all feature combinations together. Our proposed approach takes 0 hours for training. Our proposed approach will take less amount of time at max 5 minutes to extract feature counts and 1 minute to classify using simple if-then-else rules. Also our approach does not require any training of data sets and can be applicable where scarcity of data sets exists. The future work may include strengthening the accuracy of the work by taking more number of features and rigorous analysis on feature contribution.

References

Niska

Bhuiya

. National hospital ambulatory medical care survey: 2007 emergency department summary. Natl Health Stat Rep. 2010; (26): 358.

Kalra

Higgins

Perez

. Common questions about streptococcal pharyngitis. American Family Physician. 2016 Jul 1; 94(1): 24-31.

Choby

. Diagnosis and treatment of streptococcal pharyngitis. American Family Physician. 2009 Mar 1; 79(5): 383-90.

Hing

Cherry

Woodwell

. National Ambulatory Medical Care Survey: 2004 summary. Advance Data. 2006; (374): 1-33.

Dajani

Taubert

Ferrieri

Peter

Shulman

. Committee on Rheumatic Fever, Endocarditis, and Kawasaki Disease of the Council on Cardiovascular Disease in the Young, American Heart Association. Treatment of acute streptococcal pharyngitis and prevention of rheumatic fever: a statement for health professionals. Pediatrics. 1995 Oct 1; 96(4): 758-64.

Watkins

Johnson

Colquhoun

Karthikeyan

Beaton

Bukhman

Forouzanfar

Longenecker

Mayosi

Mensah

Nascimento

. Global, regional, and national burden of rheumatic heart disease, 1990–2015. New England Journal of Medicine. 2017 Aug 24; 377(8): 713-22.

Carapetis

Steer

Mulholland

Weber

. The global burden of group A streptococcal diseases. The Lancet infectious diseases. 2005 Nov 1; 5(11): 685-94.

Klepser

Dering-Anderson

Morse

Smith

Klepser

. Community pharmacist – physician collaborative streptococcal pharyngitis management program. Journal of the American Pharmacists Association. 2016 May 1; 56(3): 323-9.

Spellerberg

Brandt

. Streptococcus. In: Manual of Clinical Microbiology. 11th ed. Washington, DC: American Society of Microbiology; 2015. pp. 383-402.

10.

Fine

Nizet

Mandl

. Large-scale validation of the Centor and McIsaac scores to predict group A streptococcal pharyngitis. Archives of Internal Medicine. 2012 Jun 11; 172(11): 847-52.

11.

Aalbers

O’Brien

Chan

Falk

Teljeur

Dimitrov

Fahey

. Predicting streptococcal pharyngitis in adults in primary care: a systematic review of the diagnostic accuracy of symptoms and signs and validation of the Centor score. BMC Medicine. 2011 Dec; 9: 1-1.

12.

Bisno

. Diagnosing strep throat in the adult patient: do clinical criteria really suffice? Annals of internal medicine. 2003 Jul 15; 139(2): 150-1.

13.

Ebell

. Strep throat: Point of Care Guides. Am Fam Physician. 2003; 68: 937-8.

14.

Jarmusch

Pirro

Kerian

Cooks

. Detection of strep throat causing bacterium directly from medical swabs by touch spray-mass spectrometry. Analyst. 2014; 139(19): 4785-9.

15.

Kellogg

. Suitability of throat culture procedures for detection of group A streptococci and as reference standards for evaluation of streptococcal antigen detection kits. Journal of clinical microbiology. 1990 Feb; 28(2): 165-9.

16.

Ebell

Smith

Barry

Ives

Carey

. Does this patient have strep throat? Jama. 2000 Dec 13; 284(22): 2912-8.

17.

Zhang

. Tongue Image Analysis. 2017. doi: 10.1007/978-981-10-2167-1.

18.

Seo

Tabei

Park

Askarian

Kim

Moallem

Chong

Kwon

. Smartphone with optical, physical, and electrochemical nanobiosensors. Journal of Industrial and Engineering Chemistry. 2019 Sep 25; 77: 1-1.

19.

Gong

Lian

Chen

. Research and analysis of relationship between colour of tongue fix quantity, disease and syndrome. Chin J Inf Tcm. 2005; 7: 45-52.

20.

Yuen

. Tongue image matching using color content. Pattern Recognition. 2002 Feb 1; 35(2): 407-19.

21.

Liu

. Tongue color analysis and discrimination based on hyperspectral images. Computerized Medical Imaging and Graphics. 2009 Apr 1; 33(3): 217-21.

22.

Tang

Liu

. Traditional chinese medicine. The Lancet. 2008 Dec 6; 372(9654): 1938-40.

23.

Chen

Cheng

Chiang

. The study on the agreement between automatic tongue diagnosis system and traditional Chinese medicine practitioners. Evidence-Based Complementary and Alternative Medicine. 2012 Jan 1; 2012.

24.

Kim

Cobbin

Zaslawski

. Traditional Chinese medicine tongue inspection: an examination of the inter-and intrapractitioner reliability for specific tongue characteristics. The Journal of Alternative and Complementary Medicine. 2008 Jun 1; 14(5): 527-36.

25.

Askarian

Tabei

Askarian

Chong

. An affordable and easy-to-use diagnostic method for keratoconus detection using a smartphone. In Medical imaging 2018 computer-aided diagnosis. 2018 Feb 27; 10575: 238-243. SPIE.

26.

Chong

Cho

Tabei

Le-Anh

Esa

McManus

Chon

. Motion and noise artifact-resilient atrial fibrillation detection using a smartphone. IEEE Journal on Emerging and Selected Topics in Circuits and Systems. 2018 Mar 22; 8(2): 230-9.

27.

Tabei

Kumar

Phan

McManus

Chong

. A Novel Personalized Motion and Noise Artifact (MNA) Detection Method for Smartphone Photoplethysmograph (PPG) Signals. IEEE Access. 2018; 6: 60498-60512. [CrossRef] [PubMed].

28.

Tabei

Zaman

Foysal

Kumar

Kim

Chong

. A novel diversity method for smartphone camera-based heart rhythm signals in the presence of motion and noise artifacts. PloS One. 2019 Jun 19; 14(6): e0218248.

29.

Askarian

Jung

Chong

. Monitoring of heart rate from photoplethysmographic signals using a Samsung Galaxy Note8 in underwater environments. Sensors. 2019 Jun 26; 19(13): 2846.

30.

Hui

Thach

. Machine learning for tongue diagnosis. In 2007 6th International Conference on Information, Communications & Signal Processing. 2007 Dec 10 (pp. 1-5). IEEE.

31.

Pang

Zhang

Wang

. Computerized tongue diagnosis based on Bayesian networks. IEEE Transactions on biomedical engineering. 2004 Sep 27; 51(10): 1803-10.

32.

Wang

Zhang

Pang

. Tongue diagnosis based on biometric pattern recognition technology. In Pattern Recognition: From Classical to Modern Approaches 2001; (pp. 575-598).

33.

Zhang

Wang

Jin

Zhang

. SVR based color calibration for tongue image. In 2005 International conference on machine learning and cybernetics. 2005 Aug 18; 8: 5065-5070. IEEE.

34.

Zhang

Wang

You

Zhang

. Tongue color analysis for medical application. Evidence-Based Complementary and Alternative Medicine. 2013 Mar; 2013.

35.

Askarian

Yoo

Chong

. Novel image processing method for detecting strep throat (streptococcal pharyngitis) using smartphone. Sensors. 2019 Jul 27; 19(15): 3307.

36.

Rao

Berg

Quezada

Fader

Walker

Tang

Cowen

Duncan

Sickler

. Diagnosis and antibiotic treatment of group a streptococcal pharyngitis in children in a primary care setting: impact of point-of-care polymerase chain reaction. BMC Pediatrics. 2019 Dec; 19: 1-8.

37.

Sankaranarayanan

Ramadas

Thomas

Muwonge

Thara

Mathew

Rajan

. Effect of screening on oral cancer mortality in Kerala, India: a cluster-randomised controlled trial. The Lancet. 2005 Jun 4; 365(9475): 1927-33.

38.

Yoo

Choi

Jang

Ryu

. Toward automated severe pharyngitis detection with smartphone camera using deep learning networks. Computers in Biology and Medicine. 2020 Oct 1; 125: 103980.

39.

Carneiro

Da Nóbrega

Nepomuceno

Bian

De Albuquerque

Reboucas Filho

. Performance analysis of google colaboratory as a tool for accelerating deep learning applications. Ieee Access. 2018 Oct 7; 6: 61677-85.

40.

Sitaula

Aryal

. New bag of deep visual words based features to classify chest x-ray images for COVID-19 diagnosis. Health Information Science and Systems. 2021 Jun 18; 9(1): 24.

41.

Mansoori

Nejati

Razzaghi

Samavi

. Bag of visual words approach for image retrieval using color information. In 2013 21st Iranian Conference on Electrical Engineering (ICEE) 2013 May 14 (pp. 1-6). IEEE.

42.

Mishra

Kapoor

Tripathi

. Human action recognition using modified bag of visual word based on spectral perception. International Journal of Image, Graphics and Signal Processing. 2019 Sep 1; 11(9): 34-43.

43.

Sultani

Dhannoon

. Modified Bag of Visual Words Model for Image Classification. Al-Nahrain Journal of Science. 2021 Jun 27; 24(2): 78-86.

44.

Nakasi

Mwebaze

Zawedde

Tusubira

Akera

Maiga

. A new approach for microscopic diagnosis of malaria parasites in thick blood smears using pre-trained deep learning models. SN Applied Sciences. 2020 Jul; 2: 1-7.

45.

Minaee

Kafieh

Sonka

Yazdani

Soufi

. Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning. Medical Image Analysis. 2020 Oct 1; 65: 101794.

46.

Choi

Yoo

Seo

Kwak

Rim

. Multi-categorical deep learning neural network to classify retinal images: A pilot study employing small database. PloS One. 2017 Nov 2; 12(11): e0187336.

47.

Swathi

Regunathan

. A novel feature fusion based approach for detecting pharyngitis, oral cancer, and tonsillitis using improved bag of visual words. Soft Computing. 2023 Jul 12: 1-3.

A novel approach to classify pharyngitis using potential features based on doctor’s perspective

Abstract

Keywords

1. Introduction

5. Experimental results

Table 1 Pharyngitis dataset accumulation

References

Table 1
Pharyngitis dataset accumulation