Advanced Classification of Poxvirus-Based Skin Diseases Using Deep Learning Techniques

For the development of our classification model, the dataset was partitioned using a 70:10:20 ratio for training, validation, and testing sets, respectively, as detailed in Table 1. Figure 1 provides most samples of diverse pox diseases. The picture preprocessing pipeline blanketed annotation in a folder-primarily based total structure, accompanied by means of data augmentation strategies by each horizontal and vertical transformation with a total of 9,914 images. For data, horizontal and vertical flip processes were used as augmentation techniques, and although those technics were applied to increase dataset size and improve generalization and robustness, their potential effects on classification prediction were also examined. These augmentation methods may introduce fine-tuning of the model at a level that could affect its ability to discriminate between patterns in predictions.

Table 1. Train, test, and validation sets

1.jpg

Figure 1. Sample images from the augmented images were applied with a 50% probability of horizontal flipped and a 50% probability of vertical flipped

To compare models without any differences, all models were trained under the same conditions, including the same dataset split, learning rate, batch size, and number of epochs. The methodological framework hired for this has a look at applied Python-primarily based totally libraries, in particular Pandas and NumPy, for data manipulation and analysis. The implementation was accomplished at the Google Colab platform, leveraging the PyTorch and TorchVision libraries for our model. On the other hand, we used high-performance virtual machines provided by GoogleColabPro+ for data analysis and modeling. Thus, it provides NVIDIA A100 and up to 32 GB RAM, which provides a suitable environment for large-scale data processing and training deep learning models. Therefore, a small change in performance may be due primarily to differences in model structure rather than degradation in the training set.

2.1 Data preprocessing and collecting

The dataset utilized in this study was carefully curated from two publicly available sources: Google Image Search and Kaggle. These platforms offer a wide range of visual resources; however, they also present inherent challenges, including variations in image quality and labeling accuracy. To address these issues, a meticulous selection and preprocessing process was implemented. All images were evaluated for clarity, relevance, and resolution, with efforts made to anonymize any residual identifiers through cropping and resolution adjustments. No personally identifiable information (PII) was associated with the images. Since the dataset comprised only publicly sourced and anonymized images, formal ethical approval was not required. Nonetheless, recognizing the sensitive nature of medical imagery, we affirm our strong commitment to maintaining high ethical standards and adhering to data privacy principles in the use and dissemination of clinical image data.

• To tackle these limitations, we established a detailed screening process. Our team came up with a comprehensive selection method that went beyond merely collecting images. Each image was thoroughly assessed based on a few key criteria:

• Visual clarity and sharpness to ensure the image is clear,

• Sufficient resolution so we can conduct detailed analysis,

• Display of clear and important skin-related features to ensure relevance in dermatological studies.

Moreover, we were committed to creating a representative and inclusive dataset. This meant intentionally seeking out images that showcased the rich diversity of skin types and appearances. We carefully selected images captured under various lighting conditions and representing different skin tones and textures. By taking these meticulous steps, we aimed to minimize potential biases and create a robust, representative dataset that could provide meaningful insights into dermatological research.

This dataset focuses on the classification of five distinct diseases: Monkeypox, Measles, Chickenpox, Cowpox, and Smallpox, along with samples representing healthy human skin. The creation of the dataset involved a meticulous selection process to ensure that it accurately reflects the diversity of the targeted diseases. The distribution of the dataset and representative image samples are detailed in Table 1.

Data augmentation, the process of generating new and synthetic data by applying certain operations to existing datasets, is an important technique in machine learning [33]. In machine learning applications, it is inevitable to use large datasets to improve model performance. However, working with limited data can lead to overfitting models. This will negatively affect the performance of the model. As a solution to this problem, data augmentation methods are an alternative. The most common data augmentation method is to create transformed copies of images that belong to the same class as the existing images in the training dataset. Commonly used image augmentation methods include translation, rotation, scaling, zooming, and cropping. One of the main goals is to expand the training set by providing alternative, current, and relevant examples that reflect changes the model may encounter in real-world conditions. As shown in Figure 1, rotating an image horizontally and zooming in from different angles are among the options that can be used in this context. This is because images can be taken from different angles, such as left or right.

Conversely, applying a vertical flip to an image may be inappropriate, as it is unlikely that the model will encounter upside-down representations of the subject in real-world scenarios. This underscores the importance of judiciously selecting data augmentation techniques that are tailored to the specific context of the training dataset and informed by domain knowledge. Furthermore, when working with a limited prototype dataset, it can be beneficial to evaluate data augmentation techniques both individually and in combination to assess their impact on model performance. Such an approach allows for a systematic investigation of which techniques yield measurable improvements in the model's effectiveness [34].

Preprocessing was performed by applying each of the following steps to every image: auto-orienting pixel facts, including EXIF-orientation stripping, resizing to 224 ´ 224 pixels and stretching. Further, augmentation was made on each source image to produce three versions, with a 50% chance of horizontal flip and a 50% chance of vertical flip. Other augmentation strategies include rotation and brightness adjustment; these will also be taken into consideration for version robustness.

All pores and skin lesion photos have been shown using Google's Reverse Image Search and then cross-referenced with other sources. The ones that were no longer recognizable, of low resolution, or terrible quality went through a two-step screening process for discarding. Unique photos that met the high standards were picked, then cropped to identification at the area of interest and resized to 224 × 224 pixels while maintaining the aspect ratio. Normalization techniques were followed afterward to ensure that the pixel values were correctly scaled for training the model with the aim of facilitating improved convergence at the time of training.

The dataset is partitioned into three subsets: training, validation, and test sets. The training set, comprising 6,351 images (70% of the total data), is employed for model training and fitting. The validation set includes 921 images (10% of the data), playing a critical role in fine-tuning the model and mitigating overfitting during the training phase. The test set, containing 1,803 images (20% of the data), is reserved for evaluating the model's performance on unseen data, ensuring robust generalization. During the preprocessing phase, 43 data instances were identified as invalid and excluded from the dataset. This balanced distribution enables efficient training, validation, and testing of the model.

Table 2. Key hyperparameters for deep learning model training

As shown in Table 2, hyperparameters selection was guided by preliminary experiments and empirical optimization. Learning rates were selected based on a grid search strategy to balance convergence speed and stability. Batch size was determined by considering both model performance and GPU memory constraints.

The table above presents the major hyperparameters for developing this deep-learning model to classify images. Firstly, the --task is a parameter that defines the task; in this context, it was just image classification. Secondly, mode is "train", indicating that the model should be in training mode. Thirdly, the model should train up to 100 epochs, as defined with epochs, while efficient training will be done on a batch size of 16. All images are resized to 128 ´ 128 pixels normalizing the input size. Grayscale images are loaded parallel through 8 workers, further speeding up the training. More importantly, the model makes use of pre-trained weights. The optimizer is specified as "auto" which means automatic selection. The initial learning rate, lr0 stands at 0.01.

Additionally, the parameters include a momentum value of 0.937 and a weight decay factor of 0.0005 to regulate training dynamics and prevent overfitting. A warmup phase of 3 epochs is incorporated to stabilize the learning process initially. The inclusion of label smoothing (set to 0.0) and augmentation parameters, such as hue, saturation, and value adjustments, further aims to enhance model robustness and generalization capabilities. The probability of horizontal flipping is set at 0.5, and the mosaic augmentation is employed with a probability of 1.0, contributing to the diversity of training samples [35].

2.2 ResNet32

Residual Networks represent architecture, developed to cope with performance degradation along with the increase in depth in deep learning. As shown in Figure 2, the ResNet32 architecture was made to overcome the problem of gradient loss at each step, encountered during deep network training. It has a 32-layer network, and the "skip connections" [36] are directly available between layers. This architecture enables the training of deep networks and alleviates the problem of gradient loss. It focuses on attention with its high accuracy rates, especially in tasks such as image classification and object recognition. Hayward is another advantage of this model having 32 layers; the remaining connections let the model go deeper. The learning process then will be more stable, and effective, and reach high accuracy in complex image processing. Its success also comes to reinforcing the importance of ResNet architecture due to winning deep learning research.

2.png

Figure 2. Traditional RESNET32 architecture

2.3 YOLOv5

You Only Look Once (YOLOv5) is a deep-learning architecture designed for real-time object detection. As shown in Figure 3, each forward pass processes objects within images, allowing for precise and rapid object detection simultaneously [37]. With the use of convolutional layers along with a certain design of the network, the model is particularly efficient for real-time applications. Because of the multilayered architecture of this model, it is capable of quite effective object detection for different scales [38]. YOLOv5 model is well known among learners performing object detection in complex environments because of its straightforward implementation and great community support. It is, however, more known for precise and accurate object detection as posed by moderate difficulties given by overlapping objects and variations in object size and pose. Due to this quality, it has been used widely across multiple industries such as automation, security, and self-driving cars [39]. Due to this quality, it has been widely used across different industries such as automation, security, and autonomous vehicles. Also, it is user-friendly and has great community support which allows developers to use it for many applications.

3.png

2.4 YOLOv8

The newest member of the YOLO family, YOLOv8, brought several enhancements for object detection, as shown in Figure 4, with important gains in both accuracy and speed compared to the previous ones. Using state-of-the-art layers and modern optimization techniques for better feature extraction, YOLOv8 can detect a wide range of objects very fast [40]. Its upgraded configurations allow it to perform effectively across various applications. According to Sohan et al. [41] this latest version is designed to enhance object detection performance through innovative features. YOLOv8's refined architecture not only boosts detection capabilities but also supports diverse applications, from surveillance systems to autonomous driving, positioning it as a leading choice in deep learning for computer vision tasks.

4.png

2.5 Model performance metrics

The process of choosing the optimal classifier is one of the most important issues in machine/deep learning classification methods model development. It’s essential to choose the training dataset for building the model and the test section for testing the model when choosing the best classifier.

While the confusion matrices provide an overview of classification performance, a closer examination reveals misclassifications, particularly between diseases with visually similar symptoms, such as monkeypox and smallpox. These errors may arise due to overlapping dermatological features, image resolution limitations, or insufficient distinguishing characteristics in certain cases. Understanding these misclassifications is crucial for improving model robustness.

True Positive (TP) and True Negative (TN) metrics indicate the number of times the algorithm correctly predicts positive and negative samples, respectively, and are utilized to measure the algorithm's accuracy. The False Positive (FP) value reflects the number of instances where the algorithm incorrectly predicts a negative sample as positive, serving as a measure of the algorithm's precision. Conversely, the False Negative (FN) metric denotes the number of times the algorithm predicts a positive sample as negative, which is used to measure the algorithm's Recall. The confusion matrix is illustrated in Figure 5.