A Study on the Development of a Web Platform for Scalp Diagnosis Using EfficientNet

Abstract

Along with their physical health, modern people also need to manage the health of their scalp and hair due to changes in lifestyle habits, job stress, and environmental pollution. In this study, a machine learning model was developed to diagnose scalp conditions such as fine dandruff and perifollicular erythema. Then, transfer learning was conducted using EfficientNet-B0. A web platform that allows users to easily diagnose the condition of their scalp was also proposed. The results showed that the accuracy of the diagnosis model for fine dandruff and perifollicular erythema was 75% and 82%, respectively. It showed good performance in classifying normal, mild, moderate, and severe cases compared to previous studies. Finally, a fast and convenient web platform was developed where users can upload an image and immediately visualize their scalp condition, receive diagnostic results, and see similar cases and solutions. The analysis of user satisfaction indicates that this web application has achieved exceptional outcomes in terms of user satisfaction, garnering high evaluations for its usability, design effectiveness, and overall user experience. This setup enables users to easily check their scalp condition and is accessible to everyone, which is a significant advantage. This is expected to play a crucial role in contributing to global scalp health by advocating the benefits of the early detection and treatment of scalp-related conditions.

1. Introduction

In modern society, everyone pursues beauty, and there is a tendency to judge people’s abilities in interpersonal and social interactions based on their appearance. Therefore, people invest significantly in improving how they look [1,2,3]. The scalp and hair, along with the face, are the first features visible to others, shaping first impressions. Therefore, their importance is being increasingly emphasized in the contemporary society [1,3].

Consequently, along with their health, modern people also need to manage the health of their scalp and hair due to changes in lifestyle habits, job stress, and worsening environmental pollution [1,2,3]. The factors that impact the health of the scalp and hair are various: External damage and the accumulation of waste can inhibit hair’s health and growth and can cause hair loss [2,4,5]. The causes of an oily scalp include excessive sebum due to overactive sebaceous glands and excess sweat from overactive sweat glands. This makes the scalp and hair oily, which causes it to become sticky and oxidizes dandruff, leading to foul odors [3,4,5]. The causes of a dry scalp include insufficient sebum production, which causes itchiness; damage to the acid mantle; and scalp tension, which leads to itching, inflammation, skin peeling, and scalp damage [3,4,5]. The symptoms of seborrheic scalp include follicular erythema, itching, dandruff, and, when severe, oozing from the skin and the formation of thick crusts [2,3,4,5]. Thus, it is important to recognize and prevent conditions such as fine dandruff and perifollicular erythema, which are the early signs of deteriorating scalp health, before the condition worsens.

Traditional scalp diagnostic methods have relied mainly on patient visits to the doctor or specialist for a consultation depending on subjective opinions and memories, limiting accuracy. Additionally, many products and services that are promoted as improving scalp health disseminate inaccurate or unreliable information through misleading advertising and information [6,7].

Although there has been previous research on providing scalp diagnosis services using deep learning models such as EfficientNet-B0, VGGNet based on support vector machine (VGG-SVM), convolutional neural network (CNN), and deep neural network (DNN) [7,8,9,10,11,12,13], the development of websites and platforms has been lacking. The platform introduced a convenient scalp diagnosis system utilizing a smartphone and a portable scalp microscope [6,7,14], integrating artificial intelligence over the Internet of Things (AIoT) to enable the real-time management of diagnosis results and seamless integration with various services [15]. Nevertheless, the system’s validation approach predominantly emphasizes deep learning accuracy [6,7,14,15], underscoring the necessity for a more comprehensive evaluation that includes factors such as user satisfaction and ease of use. Moreover, the diagnostic accuracy of previous models, ranging from 65% to 77%, has not been significantly meaningful and, while diagnostic results through the classification of alopecia have been made available, the provision of solutions has often been insufficient [6,7,8,9,10,11,12,13,14,15,16,17]. That is, the system provides diagnostic results but does not offer personalized solutions tailored to the specific conditions of the user’s scalp.

Therefore, in this study, we aimed to overcome the limitations of the existing scalp diagnostic systems by developing a fast and accurate web platform service which can provide personalized solutions and diagnose conditions such as fine dandruff and perifollicular erythema using the EfficientNet model. Here, the term “diagnosis” refers to the detection of the presence or absence of fine dandruff or erythema perifollicular. The web platform will allow users to easily take photos of their scalp and receive online diagnoses at any time. This will enable users to immediately check their scalp health and take necessary actions quickly. Regular scalp checks to monitor changes and take preventive measures will be possible, allowing for the service to be used without the inconvenience of appointment scheduling and waiting times. This approach will make scalp health management much more efficient compared to traditional methods. Additionally, the web platform will provide customized solutions, suitable for users’ scalp conditions. By presenting the prevention and treatment methods according to the scalp condition and its various stages, the platform will facilitate the accurate and personalized management of scalp health. The web platform that was developed in this study combines efficient deep learning technology with a user-friendly interface, providing an innovative solution in the field of scalp diagnosis and contributing to improving users’ scalp health. Ultimately, this study focused on developing a platform that will help to maintain healthy scalp conditions through early detection and accurate diagnosis.

Another aim of this study was to provide everyone with the opportunity to easily receive scalp diagnoses, allowing them to check and address scalp health. It also aimed to raise people’s awareness of scalp health and to provide innovative solutions in the field of scalp diagnosis, contributing to improving scalp health worldwide. The contributions of this study can be summarized as follows:

• A pre-trained EfficientNet-B0 model was fine-tuned on a scalp image dataset to achieve high performance even in a limited data environment.
• By utilizing the EfficientNet-B0 model, a diagnostic model for fine dandruff and perifollicular erythema was developed, achieving high diagnostic accuracy for each condition. This demonstrates that machine learning can be a practical and reliable tool in the field of medical diagnostics.
• A user-friendly web platform was developed to allow users to upload images of their scalp and instantly receive diagnostic results and personalized solutions.

2. Research Backgrounds

2.1. Related Works

Various studies on diagnosing scalp conditions have utilized image processing and deep learning technologies. Kim et al. [18] used a microscopy image analysis technique, which quantitatively measures four scalp characteristics for use in scalp diagnoses: hair density, hair thickness, the number of pores, and erythema. The image processing techniques include contrast stretching, morphological operations, and edge detection to accurately identify hair and scalp features. The experimental results demonstrated the system’s accuracy and reliability, providing a practical tool for scalp self-diagnosis and management.

Sim et al. [19] proposed a method for assessing scalp conditions using the impedance method. This approach is useful for distinguishing between dry, oily, and normal scalp conditions and can be applied in a non-invasive and convenient manner. The research findings indicated that changes in impedance values have a significant correlation with scalp conditions, demonstrating that this method can be an effective tool for diagnosing and managing scalp health.

Kim et al. [20] suggested a scalp diagnosis system comprising an Android-based device and server. Initially, scalp images are taken via a dermoscopic camera, then hair thickness, density, and the presence of scalp inflammation are assessed using image processing, and deep learning is then used to analyze the number and distribution of hair strands. The system also explored the possibility of allowing users to share and consult on scalp conditions with experts in real time via video calling using Web real-time communication (WebRTC), enabling self-diagnosis with an affordable endoscopic camera and smart device, and providing a comprehensive analysis of the hair and scalp.

Lee et al. [14] developed a mobile scalp diagnosis system which uses a smartphone and macro lens to capture images of the scalp, diagnosing scalp conditions using fuzzy theory and data mining techniques. The research results demonstrated the system’s accuracy and usefulness, showing its potential to contribute to the smart healthcare field.

Chang et al. [15] proposed an intelligent scalp examination and diagnosis system called “ScalpEye” using Faster Regions with Convolutional Neural Networks (R-CNN) for deep learning-based detection. The ScalpEye system consists of a portable microscope, a mobile device app, a cloud-based AI training server, and a cloud-based management platform. It is capable of detecting and diagnosing four common scalp conditions: dandruff, folliculitis, hair loss, and oily hair. The experimental results showed that the ScalpEye system operated with high precision, ranging from 97.41% to 99.09%.

Baek et al. [6] developed a mobile web service for the early diagnosis and prevention of scalp diseases by applying CNN based on EfficientNet-B2 techniques. The system operates through data communication between the web, AI server, and database. The web application is developed based on React, with the user interface structured using the model, view, and controller (MVC) pattern powered by “Spring Boot”. When a user uploads an image, it is sent to a Flask server for analysis. This system is also available in a mobile environment, allowing users to take scalp photos using their smartphone cameras and upload them. The server analyzes the images in real time and returns the results in java-script object notation (JSON) format, displaying the diagnostic outcomes to the user.

Cho et al. [21] proposed a method to visually show the presence and location of keratin for the early diagnosis of seborrheic dermatitis. The Faster R-CNN model with Atrous convolution can detect keratin with high accuracy, suggesting the potential for the easy and precise self-diagnosis of scalp health conditions.

Lee et al. [7] developed the EfficientNet-B0 model for diagnosing scalp conditions using AI Hub’s “type-specific scalp image” data. They implemented it as a smartphone application. Scalp conditions are classified into six categories, with each condition evaluated on a scale from mild to severe across four stages. The model’s average accuracy was 65%, with the highest accuracy reaching up to 77% for categories with more data. The EfficientNet-B0 model maintained high diagnostic accuracy despite its lightweight structure, performing with similar success to expensive scalp diagnostic devices.

Lee [22] proposed a model to classify seven types of scalp conditions using EfficientNet and transfer learning. The data imbalance problem was improved by oversampling using image augmentation for each state. The experimental results showed that the image augmentation and down-sampling techniques were very effective in increasing accuracy.

Ha et al. [23] proposed deep learning models, such as EfficientNet, ResNet, and ViT, to accurately classify the severity of six common scalp conditions. Explainable AI (XAI) techniques, such as Grad-CAM and attention rollout, were applied to visualize the model’s prediction process. The experimental results showed that the ViT-B/16 model achieved the highest performance with an accuracy of 78.31%. Based on these results, the “Scalp Checker (ver.1.0)” software was developed, enabling users to easily monitor and track their scalp condition.

Table 1. Summary of recent research studies.

Ref.	Database	Classifier	Accuracy
[15]	Self-correction	Faster R-CNN Inception_ResNet_v2_Atrous model	Fine dandruff: 97.41% Perifollicular erythema: 97.61%
[21]	Self-correction	Faster R-CNN Inception_ResNet_v2_Atrous model	Dandruff: 97.2%
[7]	AI-Hub Scalp Images	EfficientNet-B0	Fine dandruff: 62.4% Perifollicular erythema: 77.9%
[22]	AI-Hub Scalp Images	EfficientNet-B0	Fine dandruff: 77.5% Perifollicular erythema: 84.8%
[23]	AI-Hub Scalp Images	EfficientNet-B6, ResNet-152, and ViT/16	Fine dandruff: 77.09% Perifollicular erythema: 81.38%

summarizes the results of the last five studies, and these various research studies provide solutions to the existing technical challenges in assessing scalp conditions and disease detection, offering valuable information for future research directions. Calibration methods related to accuracy vary across studies, and techniques such as cross-validation, confusion matrix, receiver operating characteristic (ROC) curve, and area under the curve (AUC) analysis have been used.

2.2. EfficientNet

The EfficientNet used in this study is an innovative architecture that was proposed by researchers to maximize the efficiency of CNN models. It was introduced by Quoc V. Le in 2019 [22,24,25]. CNNs are a form of deep learning technology that specialize in image analysis. They learn patterns directly from data, effectively recognizing specific objects, such as faces, and classifying images. CNNs use the principle of convolution to detect complex patterns and features within images. Due to these characteristics, CNNs are utilized in various fields, including object detection, face recognition, and natural language processing [10,24,26,27]. CNNs have historically evolved by increasing the size of the model to enhance performance, which has resulted in higher computational costs and a larger number of parameters as performance improved [20,21,26,27]. EfficientNet emerged as an alternative, which addressed these issues with the traditional CNNs [7,22,23,24].

EfficientNet provides an efficient method for improving model performance by simultaneously adjusting the network’s depth, width, and resolution for optimization, as shown in Figure 1. This approach successfully maintains model performance while effectively managing model size [7,22,23,24]. This feature has garnered attention in various applications, aiming to develop lightweight models that operate effectively in environments with limited computational resources.

One of the key features of EfficientNet is optimizing computational resources to enhance the efficiency of the network. In scalp image analysis where large amounts of data and high resolutions have been required, EfficientNet has proved to be extremely helpful. Moreover, despite its smaller model size, EfficientNet provides high accuracy, making it suitable for scalp diagnoses using mobile devices or in environments with limited resources [7,22,23,24].

2.3. Transfer Learning

Deep learning models based on CNNs show high performance in image classification but require significant time and resources for training [10,24,26,27,28,29,30,31]. Meaningful performance is difficult to achieve if there is a lack of data, thus necessitating substantial training data [6,7,32,33]. Transfer learning involves retraining a pre-trained model for a different domain. Utilizing an existing model can significantly reduce the design time, thus addressing the aforementioned challenges [22,24,28,29,30,31].

Transfer learning aims to extract knowledge from one or more source tasks and apply it to a target task [34,35,36,37]. Unlike multi-task learning, which focuses on learning all the source and target tasks simultaneously, transfer learning prioritizes the target task. In transfer learning, the roles of the source and target tasks are no longer symmetric. Figure 2 illustrates the difference between the traditional learning processes and transfer learning techniques. The traditional machine learning methods attempt to learn each task from scratch. In contrast, transfer learning techniques aim to transfer knowledge from previous tasks to a target task, particularly when the target task has limited high-quality training data.

In this study, the amount of data was adjusted due to the given development environment. Transfer learning is suitable for creating scalp diagnosis models with limited data because it allows the use of weights from models that were trained in areas with abundant data [32,33]. Therefore, this study applied the method of retraining the EfficientNet model to achieve high performance even with limited input data.

3. Materials and Methods

3.1. Database

This study was a secondary analysis research study which used the “Scalp Image by Type” data from the public research data platform, AI-Hub, Republic of Korea, to develop a model for diagnosing scalp conditions, focusing on fine dandruff and perifollicular erythema [38].

The type-specific scalp images from AI-Hub are publicly available data, which are essential for diagnosing and treating scalp disorders. These images were compiled into a dataset of over 100,000 entries in 2020 to enhance the competitiveness of scalp care products and services and to create job opportunities that were accessible to socio-economically disadvantaged groups. The details of the data provided by the institution are shown in

Table 2. Construction status of scalp images by type.

Classification		Total Number of Images
Data construction	Collection	126,948
	Processing	101,027
	Inspection	101,027
Number of constructions by symptom	Fine dandruff	17,434
	Excessive sebum	80,416
	Perifollicular erythema	67,414
	Follicular erythema/pustule	4592
	Dandruff	40,482
	Hair loss	25,682
	Good condition	811

. The type-specific scalp images are categorized into seven types: fine dandruff, excessive sebum, perifollicular erythema, follicular erythema/pustule, dandruff, hair loss, and good condition. They are further classified into three severity stages: 1 (mild), 2 (moderate), and 3 (severe). This study only utilized the fine dandruff and perifollicular erythema classes. The image data for fine dandruff and perifollicular erythema provided by AI-Hub are shown in

Table 3. The number of images for fine dandruff and perifollicular erythema.

		Normal	Mild	Moderate	Severe
Fine dandruff	Train	534	4435	5486	2284
	Val	152	1267	1568	652
Perifollicular erythema	Train	534	29,960	12,957	332
	Val	152	8560	3702	1221

.

The workflow to develop a machine learning model based on EfficientNet using transfer learning with scalp image data is illustrated in Figure 3. Initially, the scalp image data were preprocessed to enhance the efficiency of the model training, and the datasets for training, validation, and testing were then organized. Subsequently, a machine learning model was developed using transfer learning based on EfficientNet-B0, incorporating additional layers to effectively extract and classify various features of the images.

3.2. Image Data Preprocessing

The original dataset for fine dandruff provided by AI-Hub consisted of a total of 17,434 images, as shown in Table 3. Due to limitations in the given development environment, the quantity of data used in this study was reduced. The number of training images for the fine dandruff diagnosis model was reduced to 5034, and the number of validation images was reduced to 1259. The test dataset was composed of 1007 images, extracted at a 5:1 ratio from the training data. The preprocessing results are shown in

Table 4. Preprocessing of image data for fine dandruff.

		Normal	Mild	Moderate	Severe
Fine dandruff	Train	534	1500	1500	1500
	Val	134	375	375	375
	Test	107	300	300	300

.

The original dataset for perifollicular erythema provided by AI-Hub consisted of a total of 67,414 images, as shown in Table 3. Due to limitations in the given development environment, the number of training images for the perifollicular erythema diagnosis model was maintained at 4692 and the number of validation images was reduced to 1174, extracted at an 8:2 ratio from the training data. The test dataset comprised 652 images, extracted at a 9:1 ratio from the training data. The preprocessing results are shown in

Table 5. Preprocessing of image data for perifollicular erythema type.

		Normal	Mild	Moderate	Severe
Perifollicular erythema	Train	427	2000	2000	265
	Val	107	500	500	67
	Test	59	278	278	37

.

The dataset exhibited a significant imbalance among the different class types, which can degrade the performance of machine learning classification models [32,33]. To address the class imbalance, two measures were implemented. First, data augmentation techniques were used to increase data diversity by transforming the data. Initially, the normalization of the image data was performed by converting pixel values from 0 to 255 to a range of 0 to 1. Subsequently, to augment the data, the images were rotated by 20 degrees, shifted horizontally and vertically by 10%, sheared by 20%, and horizontally flipped, with the fill mode set to “nearest” to increase data diversity, as shown in Figure 4. Second, weights were assigned to address the data imbalance. These calculated class weights served as a means to assign appropriate importance to each class during the model training, mitigating the effects of the unbalanced class distribution. In the case of fine dandruff, the assigned weights for each class were 2.36 for good conditions and 0.84 for mild, moderate, and severe conditions. For perifollicular erythema, the assigned weights for each class were 2.75 for good conditions, 0.59 for mild, 0.59 for moderate, and 4.43 for severe conditions.

3.3. Model Configuration

The total number of fine dandruff and perifollicular erythema images used in this study were 7300 and 6518, respectively. The original size of the images provided by AI-Hub was 640 × 480, but they were resized to 224 × 224 to fit the EfficientNet-B0 model.

The model parameters for the fine dandruff and perifollicular erythema modeling were set to batch sizes of 16 and 24, respectively, and the class mode was set to “sparse”. In the test data, the shuffle was set to “false” to allow for comparison between the predicted results and the actual labels.

The training model used the pre-trained EfficientNet-B0. After adjusting the network weights, new layers were created so that the existing layers of EfficientNet-B0 could be re-used, and only the directly added layers were trained. EfficientNet ranges from B0 to B7 depending on the model size; however, considering the resources available, B0 was used for training in this study.

The models for the diagnosis of fine dandruff and perifollicular erythema are shown in Figure 5. A GlobalAveragePooling2D layer was added to convert the 3D tensor to 1D and to average the feature maps. A BatchNormalization layer was then added to normalize the features, with a mean of 0 and a variance of 1. A fully connected layer, which was dense, was added to decide the number of neurons and the activation function based on user customization. To prevent model overfitting, a dropout layer was added to deactivate neurons. The loss function used for image classification was sparse categorical cross entropy, and the optimizer used was Adam. For the training of the fine dandruff data, the initial learning rate was set at 0.001, the step count for decreasing the learning rate was set at 300, and the reduction rate was set at 0.97 to set up a learning rate schedule. For the training of the perifollicular erythema data, the initial learning rate was set at 0.0001, the step count for decreasing the learning rate was set at 1500, and the reduction rate was set at 0.98 to set up a learning rate schedule.

The training parameters of the model were set at 30 and 50 epochs, and an EarlyStopping callback was set up to stop the training if there was no improvement in the model’s validation loss for 10 and 5 consecutive evaluations for the training of fine dandruff and perifollicular erythema, respectively.

4. Results

4.1. Severity Classification of Two Hairy Scalp Disorders

To measure the performance of this diagnostic model, we used the confusion matrix-based accuracy, recall, precision, F1-score, and support. Given that this model was for multi-class classification, accuracy was considered to be the most critical metric [39,40].

Based on the results presented in

Table 6. Performance evaluation metrics for the fine dandruff diagnosis model.

	Precision	Recall	F1-Score	Support
Normal	0.73	0.98	0.84	107
Mild	0.84	0.62	0.71	300
Moderate	0.61	0.76	0.68	300
Severe	0.88	0.79	0.83	300

, the accuracy of the fine dandruff diagnosis model was approximately 75%. This table highlights the type-specific learning results for the fine dandruff model, providing detailed performance metrics for different categories of dandruff severity: normal, mild, moderate, and severe. The model exhibited relatively high performance in identifying normal, mild, and severe cases, as indicated by their F1-scores, which were all above 70%. Specifically, the normal category achieved an F1-score of 0.84, reflecting the model’s strong ability to correctly identify scalps in good condition. Similarly, the mild category had an F1-score of 0.71, and the severe category had an F1-score of 0.83, both indicating reliable performance in these severity levels. However, the model’s performance in the moderate category was comparatively lower, with an F1-score of 0.68. This suggests that while the model can reasonably identify the moderate cases of dandruff, its performance is less accurate than in the other categories. The lower F1-score for the moderate category highlights an area where the model could be improved to achieve more balanced and accurate diagnostic capabilities across all the levels of dandruff severity. Overall, while the fine dandruff diagnosis model demonstrated robust performance in most categories, the moderate category’s lower F1-score indicates a need for further refinement and additional training data to enhance its diagnostic precision. The comprehensive performance metrics provided in Table 6 offer valuable insights into the model’s strengths and weaknesses, guiding future efforts to optimize and balance its diagnostic capabilities.

Table 7. Performance evaluation metrics for the perifollicular erythema diagnosis model.

	Precision	Recall	F1-Score	Support
Normal	0.84	0.96	0.89	59
Mild	0.83	0.73	0.78	278
Moderate	0.81	0.84	0.82	278
Severe	0.75	0.97	0.85	37

provides a comprehensive set of performance evaluation metrics for the perifollicular erythema diagnosis model. The metrics include precision, recall, F1-score, and support for different severity levels: normal, mild, moderate, and severe. Based on these results, the overall accuracy of the perifollicular erythema model was found to be approximately 82%. In more detail, the normal category achieved a precision of 0.84, a recall of 0.96, and an F1-score of 0.89, with a support value of 112, indicating the number of actual normal instances in the dataset. This high F1-score reflects the model’s strong ability to correctly identify the normal conditions of perifollicular erythema. The mild category, on the other hand, showed comparatively lower performance metrics. It achieved a precision of 0.83, a recall of 0.73, and an F1-score of 0.78, with a support value of 278. Although the precision was fairly high, the recall and F1-scores were somewhat lower, indicating that the model has more difficulty accurately identifying the mild cases of perifollicular erythema. The model performed well in the moderate category, with a precision of 0.81, a recall of 0.84, and an F1-score of 0.82, with a support value of 278. These metrics demonstrate a balanced performance in identifying the moderate cases, with both precision and recall contributing to a strong F1-score. The severe category had the highest recall, at 0.97, but a slightly lower precision, at 0.75, resulting in an F1-score of 0.85 and a support value of 37. This indicates that while the model is very effective at identifying severe cases (high recall), it sometimes misclassifies other severity levels as severe (lower precision).

Overall, the normal, moderate, and severe categories all displayed relatively high F1-scores (above 80%), signifying the model’s robust performance in these areas. In contrast, the mild category, with an F1-score below 80%, showed a comparatively low performance, suggesting that the model has room for improvement in accurately diagnosing mild perifollicular erythema.

Building on these findings, the next section proposes enhancements to the perifollicular erythema model to address the identified weaknesses. Specifically, the focus is on increasing the amount of training data for the mild category and refining the model’s parameters to improve precision and recall. Additionally, the development of an integrated platform with a web interface is discussed, aiming to provide users with accessible diagnostic tools for perifollicular erythema. This platform will enable users to upload images, receive real-time analysis, and obtain personalized care recommendations, leveraging the strengths of the current model while addressing its limitations. Through these improvements, the proposed system aims to enhance the user experience and diagnostic accuracy in managing perifollicular erythema.

Compared to the study conducted by Lee [22], the performance of the model presented in this paper in diagnosing fine dandruff and perifollicular erythema was found to be approximately 2% lower. This slight performance discrepancy can be primarily attributed to the difference in the amount of training data available. Although the current study achieved a level of accuracy similar to that reported in Lee’s research [22], the smaller dataset used in this study most likely impacted the final results. Despite this minor difference, the model developed still demonstrates a high degree of accuracy and reliability. Building on these findings, the next section proposes the development of a comprehensive platform, integrated with a web interface that was specifically designed for scalp diagnosis. The aim of this platform is to provide users with easy access to diagnostic tools, enabling them to upload scalp images, receive an immediate analysis, and obtain personalized recommendations for scalp care. By leveraging the robust model developed in this study, the proposed platform seeks to enhance user experience and accessibility in managing scalp health, offering a practical application for the research outcomes.

4.2. Scalp Diagnosis Web Platform

The development of the Scalp Diagnosis Web Platform marks a significant innovation in the field of scalp health management. This platform utilizes the EfficientNet-B0 model to diagnose scalp conditions such as fine dandruff and perifollicular erythema, achieving diagnostic accuracy rates of 75% and 82%, respectively. The primary objective of this platform is to provide users with an accessible, efficient, and reliable method to assess and manage their scalp health in real time.

The Scalp Diagnosis Web Platform allows users to upload images of their scalp directly through a web interface. Once the image is uploaded, the platform processes it using the trained EfficientNet-B0 model. This involves converting the image into a suitable format, normalizing it, and then passing it through the neural network to obtain a diagnostic result. The platform is designed to be user-friendly, ensuring that even users with minimal technical knowledge can easily navigate and utilize the service. Figure 6 illustrates the overall flow of the platform. When a user uploads an image, it is sent to the server and processed through PIL. The processed image is then input into the EfficientNet-B0 model to generate diagnostic results. These results are returned to the user, visualized, and accompanied by similar cases and solutions. The back-end of the platform was built using Python ver.3.11, which handles image processing and model inference. The images uploaded by users are converted into PIL Image objects, transformed into arrays, and normalized to match the input requirements of the EfficientNet-B0 model. The model, which was trained and stored in an H5 format, is then used to generate diagnostic results. This setup allows for swift and accurate diagnostics, providing users with immediate feedback on their scalp condition. The front-end was developed using JavaScript, with the Chart.js library employed to visualize diagnostic results. The users receive a graphical representation of their scalp condition, categorized into stages (normal, mild, moderate, and severe), and they are also provided with suggested solutions and similar case comparisons. This visual and informative feedback is crucial in enabling users to understand their scalp health status and take appropriate actions.

Solutions are features that inform users about appropriate treatment and prevention methods based on their scalp condition. For mild symptoms, the solution content includes internal health management—such as diet, care habits, and nutritional management—as well as external care—such as using shampoos and essences that are suitable for the individual’s scalp type and massages. For moderate to severe symptoms, it includes the aforementioned advice with an added recommendation to consult a doctor as a priority [41].

Similar cases have used the existing type-specific scalp image data from AI-Hub, providing two images each that match the respective conditions. Figure 7 and Figure 8 depict the final GUI that is presented to the users. Figure 7 shows the interface for uploading images. Users can easily upload images through this interface, and the uploaded images are processed in real time, with the diagnostic results visualized for the user.

Figure 8 depicts the results screen of the scalp diagnosis system. Similar cases for the scalp’s microscaling condition can be viewed in the “Microscaling step1 Similar Cases” section, which includes two images showing the side of the head and the top of the head. Additionally, similar cases for follicular erythema are provided in the “Follicular Erythema step0 Similar Cases” section, which also shows images of the side and top of the head. The user’s scalp condition is visually displayed, with an image and a graph illustrating the step-by-step scalp condition for microscaling and follicular erythema. The text description states, “Your Scalp is that ‘Microscaling’ is Mild, ‘Follicular Erythema’ is Good”, indicating that the user’s scalp condition is mild for microscaling and good for follicular erythema. The “solution section” offers specific advice for each condition. For mild microscaling, the advice is to “Use exfoliation products regularly to remove dead skin cells. Maintain your scalp properly by using moisturizing shampoo and conditioner”. For good follicular erythema, the recommendation is, “Regular cleansing twice a week is important. Keep your scalp healthy with gentle shampoo and conditioner and promote blood circulation through scalp massages.” This system aims to help users easily understand their scalp condition and apply appropriate management methods.

The Scalp Diagnosis Web Platform offers several benefits over traditional scalp health assessment methods. First and foremost, it eliminates the need for time-consuming visits to specialists by providing instant online diagnostics. This convenience is especially beneficial for users with busy schedules or those living in remote areas with limited access to dermatological services. Furthermore, the platform enhances diagnostic consistency by leveraging a trained machine learning model, reducing the variability associated with human assessments. The provision of customized solutions based on the diagnosed condition promotes proactive scalp health management, allowing users to address issues before they escalate. Despite its advantages, the platform currently faces certain limitations. The diagnostic accuracy, although promising, could be further improved by expanding the dataset and incorporating more diverse scalp conditions. The moderate category, in particular, exhibited lower performance, indicating the need for more granular data to enhance the model’s differentiation capabilities. Future enhancements should focus on augmenting the dataset with more varied and balanced samples, employing advanced data augmentation techniques, and exploring additional deep learning architectures to boost performance. Incorporating user feedback will also be vital in refining the platform’s usability and effectiveness.

4.3. User Evaluation of Scalp Diagnosis Web Platform

The questionnaire in

Table 8. Questionnaire and analysis results for Scalp Diagnosis Web Platform.

Questionnaire	Average Scores
Using this web application is simple and intuitive.	4.5
2. I found the design of the web application to be user-friendly.	4.7
3. It did not require much effort to understand the functions of the application.	4.4
4. I did not feel confused while using the web application.	4.6
5. This web application helps me easily accomplish the tasks I want to perform.	4.8
6. The response speed of the web application is fast and smooth.	4.5
7. When errors occurred during use, I could easily resolve them.	4.6
8. I am willing to use this web application again.	4.7
9. I find using it on a computer convenient and accessible.	4.6
10. I would recommend this web application to others.	4.8

was used to evaluate the user-friendliness, user experience, and accessibility of the Scalp Diagnosis Web Platform. This survey is based on the system usability scale (SUS) and is designed to efficiently and effectively assess user experience. Each item is rated on a 5-point Likert scale where users can select from “Strongly Disagree” (1 point) to “Strongly Agree” (5 points) for each statement. Additionally, the analysis of each item based on the survey results from 10 male and 10 female users, aged between 20 and 40, is presented in Table 8.

The web application received very positive feedback overall, with all the scores ranging between 4.4 and 4.8. This indicates a high level of user satisfaction across the various aspects of the application. In terms of ease of use and intuitiveness, the statement “Using this web application is simple and intuitive” scored 4.5, showing that users found the application easy to navigate and use without significant difficulty. The design aspect received a particularly high score of 4.7, indicating that users were very satisfied with the layout, esthetics, and overall usability of the web application. Notably, the statements “This web application helps me easily accomplish the tasks I want to perform” and “I would recommend this web application to others” both scored 4.8, highlighting the application’s effectiveness in helping users achieve their goals and the confidence users have in recommending it to others. Additionally, the statement “I am willing to use this web application again” scored 4.7, demonstrating that users found significant value in the application and are highly inclined to reuse it. On the other hand, the statement “It did not require much effort to understand the functions of the application” scored slightly lower at 4.4. This suggests there may be room for improvement in making the functions more intuitive or providing better onboarding and help features. Similarly, the statements “The response speed of the web application is fast and smooth” and “When errors occurred during use, I could easily resolve them” scored 4.5 and 4.6, respectively. While these are strong scores, they indicate that further optimization in speed and more user-friendly error handling could enhance the overall user experience.

5. Discussion

This study represents a noteworthy advancement in the field of scalp health diagnosis, presenting both a robust machine learning model and a practical application through a web-based platform. The utilization of EfficientNet-B0 in diagnosing scalp conditions such as fine dandruff and perifollicular erythema demonstrates promising results, with the model achieving accuracy rates of 75% and 82%, respectively. When comparing the results of this study with the most recent study [22], it is observed that [22] achieved an accuracy of 77.5% in diagnosing fine dandruff, while this study achieved an accuracy of 75%. Although there is a difference of approximately 2.5%, the performance levels are generally similar. This discrepancy may be attributed to the differences in the dataset size, preprocessing methods, or specific aspects of the training and validation processes. In diagnosing perifollicular erythema, [22] reported an accuracy of 84.8%, whereas this study recorded an accuracy of 82%. Despite a 2.8% difference, both studies demonstrate comparable performance. In conclusion, the results of this study and those of [22] exhibit a similar level of achievement overall. The minor differences are likely due to the variations in the dataset size or specific settings during model training. This consistency in the results between the two studies, both utilizing the AI-Hub Scalp Images database and EfficientNet-B0 model, highlights the reliability of the findings. These results underscore the potential of machine learning in medical diagnostics, particularly in providing accessible and reliable health assessments.

One of the significant strengths of this research study was the integration of EfficientNet-B0. EfficientNet-B0 is renowned for its efficient scaling of CNNs, balancing the trade-offs between model depth, width, and resolution [22,24,25]. This balance allows for high performance with relatively low computational requirements, making it an excellent choice for applications requiring real-time processing and deployment on devices with limited resources [7,22,23,24]. By leveraging this architecture, the study addresses a critical need for efficient and scalable solutions in medical imaging.

The implementation of transfer learning is another crucial aspect of this study. Transfer learning, which involves retraining a pre-existing model on a new, often smaller, dataset, proved to be effective in this study [6,7,32,33]. The pre-trained EfficientNet-B0 model, which was initially trained on a large dataset, was fine-tuned using specific scalp condition images. This method is particularly advantageous when dealing with limited data as it reduces the need for extensive data collection and annotation. The study effectively demonstrated that transfer learning can enhance model performance and accuracy, even with a relatively small amount of domain-specific data [34,37].

The developed web platform can significantly enhance the accessibility of scalp health diagnostics. Traditional diagnostic methods often involve time-consuming visits to specialists, relying heavily on subjective visual assessments [6,7,42]. These methods can be inconsistent and inconvenient for users [6,7]. The web platform developed in this study allows users to upload images of their scalp and receive instant diagnostic feedback. This immediate response capability is a considerable improvement on traditional methods, offering users a convenient and efficient way to monitor and manage their scalp health. The platform’s ability to provide diagnostic results, similar case comparisons, and suggested solutions further adds to its value as it promotes proactive health management. Overall, this web application has achieved excellent results in terms of user satisfaction, receiving high ratings for ease of use, design effectiveness, and overall user experience. Although there is potential for enhancement in the areas of functionality comprehension, response speed, and error resolution, the overall user satisfaction with the application remains high.

Despite the promising results, this study had several limitations that warrant discussion. One significant limitation was the reduced dataset used for the model training, which was constrained by the development environment. The small dataset most likely impacted the model’s overall accuracy, particularly in diagnosing the moderate cases of fine dandruff. In this study, the small dataset resulted in a lack of diversity, which, in turn, led to inadequate model training. Consequently, the model was unable to effectively learn the necessary patterns. This limitation highlights the importance of a comprehensive and balanced dataset to train machine learning models capable of handling nuanced variations in medical conditions. Another limitation was the moderate category’s lower performance in fine dandruff diagnosis. Moderate conditions are often more challenging to diagnose accurately due to their subtle and varied presentations. This issue underscores the necessity for more granular and detailed data to improve the model’s ability to differentiate between the severity levels of scalp conditions.

To build upon the current study’s success, future research should focus on several key areas: Expanding the dataset to include a more diverse range of scalp conditions and ensuring balanced class distributions would significantly enhance the model’s robustness and accuracy. Accordingly, we will integrate existing public datasets related to scalp conditions to enhance data diversity. Furthermore, we plan to collaborate with various clinics, hospitals, and research institutions to gather a more comprehensive dataset. Employing advanced data augmentation techniques could also help mitigate data scarcity issues and improve model generalization. For example, employing synthetic data generation techniques, such as generative adversarial networks (GANs), can produce realistic images of scalp conditions. This method can potentially address the limitations associated with a small dataset.

Exploring other deep learning architectures and hybrid models could also provide further improvements in diagnostic performance. Additionally, integrating more personalized treatment recommendations based on users’ specific scalp conditions and medical history could enhance the platform’s utility and user satisfaction.

Further studies should also consider the incorporation of user feedback to continually refine and improve the platform. Moreover, real-world application and user interaction would provide valuable insights into usability and effectiveness, guiding iterative improvements.

6. Conclusions

The Scalp Diagnosis Web Platform represents a pioneering step in the integration of machine learning with practical health applications. By providing an accessible, efficient, and reliable method for scalp health assessment, with diagnostic accuracy rates of 75% for fine dandruff and 82% for perifollicular erythema, it has the potential to revolutionize personal health management. Here, the term “diagnosis” refers to the detection of the presence or absence of fine dandruff or erythema perifollicular. The analysis of user satisfaction indicates that this web application has achieved outstanding outcomes in terms of user satisfaction, garnering high evaluations for its usability, design effectiveness, and overall user experience with all the scores ranging between 4.4 and 4.8. The continued development and enhancement of this platform will contribute significantly to global scalp health, offering a robust tool for the early detection and treatment of scalp-related conditions.

The potential future research directions based on this study include several key areas. First, expanding the dataset to encompass a more diverse range of populations and scalp conditions could enhance the generalizability of the diagnostic algorithms, thereby enabling more reliable diagnostics for a broader user base. Second, integrating advanced machine learning algorithms, such as deep learning or reinforcement learning, could improve the accuracy and efficiency of the diagnosis process. Third, the development of real-time diagnostic capabilities could significantly enhance user experience, allowing for immediate feedback and facilitating prompt intervention for early treatment. Fourth, expanding the diagnostic scope of the platform to include a wider array of scalp-related conditions beyond those currently addressed would provide a more comprehensive solution for scalp health management. Finally, leveraging user data to offer personalized scalp health management solutions could result in tailored recommendations, thereby contributing to more effective individual health management.