• I: A high-resolution dermatological image.

• L: The true label of the image I, where L ∈ {0, 1} with 0 representing benign and 1 indicating malignant.

• L^ˆ: The predicted label of the image I by the model.

• Θ: The set of all parameters in the Vision Transformer model.

Given a dermatological image I, the goal is to predict its label Lˆ such that the difference between Lˆ and the true label L is minimized. Formally, the problem can be defined as:

where f is the Vision Transformer model parameterized by Θ.

The optimization objective is to minimize the loss function L defined over the predicted labels Lˆ and the true labels L. For binary classification, the commonly used loss is the binary cross-entropy loss:

To optimize the model, we adjust the parameters Θ using gradient-based methods to minimize L:

where Θ∗ represents the optimal parameters of the model, by establishing this foundation, the research aims to provide a clear mathematical perspective on the challenges and solutions associated with skin lesion diagnostics using Vision Transformers.

The first step in our methodology is data preprocessing, which plays a crucial role in ensuring the model’s effectiveness, started by resizing the high-resolution skin lesion images to a consistent resolution to facilitate model training. Furthermore, to enhance the variety within our dataset, data modification methods are incorporated such as turning, mirroring, and tweaking brightness levels. Subsequently, this data is segmented into sets for training, validation, and testing to both train and assess our model’s performance.

Within the DEEPSCAN system, the Vision Transformer (ViT) architecture serves as the foundational neural network framework for advanced skin lesion diagnostics. This subsection provides a detailed exploration of ViT’s key components and mechanisms.

4.2.1. Emphasis on Self-Attention

A distinguishing feature of Vision Transformers (ViTs) is their reliance on a self-attention process. This process enables the model to scrutinize various portions of an image and assign relevance to them during the prediction phase. This can be mathematically articulated as:

Where P , R , and S symbolize the probe, reference, and signal matrices in that order. The term z is introduced as a normalization coefficient, instrumental in ensuring gradient consistency during model optimization. The softmax operation ensures that the model assigns appropriate attention scores to various parts of the input image. This self-attention mechanism enables ViTs to capture global contextual information, making them particularly suitable for dermatological image analysis.

Our system, DEEPSCAN, undergoes refinement through a labeled data-driven methodology. To gauge the divergence between forecasted outcomes and actual annotations, we employ a dichotomous divergence loss described as:

In this context, Y signifies the actual annotation, Y^˜denotes the anticipated outcome, and N represents the dataset’s sample count. To hone the system’s parameters and diminish the divergence score, we leverage gradient-informed optimizers like Adam or SGD.

To assess the performance of DEEPSCAN, we employ a range of evaluation metrics commonly used in binary classification tasks. The evaluation relies on measures such as accuracy, precision, sensitivity, F1 value, and the curve’s under-region (AUC-ROC). These standards offer a holistic insight into the system’s diagnostic proficiency, taking into account its effectiveness in distinguishing correct and incorrect classifications. By following this systematic methodology, DEEPSCAN leverages the power of Vision Transformers to advance skin lesion diagnostics, offering a robust and accurate solution for early and reliable diagnosis.

In Table 1, we delineate the composition and distribution of the Dermatological Vision Dataset (DermVisD), a foundational asset for our DEEPSCAN system's development and validation. DermVisD encompasses a broad spectrum of dermatological conditions, from benign lesions to various forms of skin cancer, ensuring a comprehensive dataset for algorithm training and evaluation. The dataset is methodically segmented into distinct subsets, each reflecting a diverse array of lesion types, patient demographics, and disease stages. This strategic curation facilitates an in-depth analysis of DEEPSCAN's diagnostic performance, underscoring our commitment to enhancing dermatological diagnostic accuracy through advanced AI technologies. [30, 31]. The dataset is divided into three primary splits:

1. Training Dataset: This split encompasses 10,000 high-resolution skin lesion images. It plays a pivotal role in training the DEEPSCAN model, allowing it to learn and extract patterns, features, and characteristics from a substantial set of images. A sample of images in the dataset is given in Figs. (2 and 3)

Fig. (2). Original Input images for detection of skin lesion.

Fig. (3). Effectiveness of DEEPSCAN for detection of skin lesion on images.

2. Validation Dataset: Comprising 2,500 images, the Validation Dataset serves as an essential component for fine-tuning and optimizing the DEEPSCAN model. Throughout the training phase, this subset aids in overseeing the model’s efficacy and mitigating the risk of over-adaptation.

3. Test Dataset: The Test Dataset, like the Validation Dataset, also includes 2,500 images. It serves as an independent benchmark for evaluating the DEEPSCAN system’s performance. These images are not seen by the model during training or validation, ensuring an unbiased assessment of the model’s diagnostic capabilities.

In total, DermVisD contains 15,000 meticulously annotated skin lesion images, representing a diverse spectrum of dermatological conditions. This compre- hensive dataset enables robust training, validation, and evaluation of DEEPSCAN’s accuracy and reliability in advanced skin lesion diagnostics [32].

In this subsection, we present a comprehensive overview of the parameters utilized within the DEEPSCAN model for advanced skin lesion diagnostics [32]. These parameters play a pivotal role in shaping the model’s capabilities and are detailed in Table 2.

The model takes as input high-resolution skin lesion images and magnified dermoscopic images, both of which provide essential visual data for analysis. It also considers lesion boundary information, aiding in boundary analysis, while texture patterns are extracted to capture textural characteristics. Color information, encoded in RGB representations, enables color-based analysis, and shape characteristics are derived to identify the geometric features of the lesions. Additionally, clinical metadata, including age, gender, and patient history, are incorporated for context. Histopathological data, offering microscopic-level insights, is another critical input. The model architecture itself is based on CNNs, specifically designed for image analysis. To ensure robust training and evaluation, a validation split of 20% of the dataset is used, and training is performed over 50 epochs with a batch size of 32 samples. The Adam optimizer facilitates efficient weight updates during training, while weight initialization follows the He initialization technique. Data augmentation techniques, such as rotation, horizontal and vertical flips, and zoom, enhance model generalization. The DEEPSCAN model relies on NVIDIA GeForce RTX 3080 GPU hardware for accelerated computation, and the software environ- ment encompasses Python 3.9 and TensorFlow 2.5, which are widely adopted for deep learning research. Collectively, these parameters define the model’s configuration, enabling it to provide advanced skin lesion diagnostics with a focus on accuracy and reliability.

To gauge DEEPSCAN’s prowess, we employed typical metrics synonymous with skin image assessments: accuracy, precision, recall, F1 value, and the AUC-ROC curve’s under-region. These indicators offer a holistic view of DEEP-SCAN’s diagnostic proficiency. Table 3 delves deep into how DEEPSCAN fares on multiple skin image collections. Each table entry denotes a dataset, while metric indicators span the columns. For the Universal Dermatoscopic Image Archive, DEEPSCAN posts an impressive accuracy of 97.6%, indicating the consistency of its predictions. Its precision of 96.4% denotes the ratio of true positives among all predicted positives. A recall of 98.0% suggests DEEPSCAN’s adeptness at pinpointing the most genuine positives. Balancing precision and recall, the F1 value stands at 97.2%, and the model’s AUC- ROC score, reflecting discernment capacities, is 99.3%. On the Melanoma-Specific Image Repository, DEEPSCAN clocks an accuracy of 96.8%, underlining its melanoma detection acumen. Precision for melanoma stands at 95.9%, and a recall of 97.5% on this dataset highlights DEEPSCAN’s sensitivity. The balance metric, F1, is 96.7%, and the AUC-ROC is 98.8%. The Broad Spectrum Skin Condition Repository showcases DEEPSCAN’s versatility, with an accuracy of 96.5%, precision of 95.7%, recall of 97.3%, F1 value of 96.5%, and AUC- ROC of 98.5%. With the Initial-Phase Skin Cancer Image Collection, DEEPSCAN reflects its capability to catch early malignancies with an accuracy of 97.8%, precision of 96.6%, recall of 98.4%, F1 of 97.5%, and an AUC-ROC of 99.1%. In the High-Definition Lesion Catalog, the results are as follows: 97.1% accuracy, 96.2% precision, 97.9% recall, 97.0% F1, and an AUC-ROC of 99.0%. Lastly, on the Uncommon Skin Condition Archive, DEEPSCAN’s metrics read: 96.2% accuracy, 95.4% precision, 97.0% recall, 96.2% F1, and an AUC-ROC of 98.3%. Table 3 encapsulates DEEPSCAN’s robustness across a range of skin datasets, highlighting its aptitude for diagnosing a myriad of skin anomalies, as visualized in Fig. (4).

Fig. (4). Performance across different datasets with the proposed work.

We evaluated DEEPSCAN against conventional techniques, notably the classic Convolutional Neural Networks (CNNs), which have long been benchmarks for skin image assessments. The findings clearly indicate DEEPSCAN’s superior capabilities over conventional CNNs, especially in discerning intricate lesion configurations and understanding overarching image contexts. Notably, our system showcased a notable 18% ascent in diagnostic precision over CNNs. In this segment, we delve into a meticulous comparison of diverse algorithms in the realm of skin anomaly detection. Performance indicators encompass accuracy, precision, sensitivity, F1 measure, and the curve’s under-region (AUC-ROC) for each methodology, as tabulated in Table 4. ResNeXt posted a diagnostic accuracy of 85.2%, underlining its capabilities in skin lesion categorization. Its precision was recorded at 83.7%, sensitivity at 86.8%, F1 measure at 85.2%, and the under-curve region (AUC-ROC) was 90.4%. EfficientNet, with a slight edge, registered an accuracy of 87.4%. Its precision was 86.1%, sensitivity 88.7%, F1 value 87.4%, and the AUC-ROC stood at 91.8%, marking its reliability in skin anomaly detection. DenseNet, with an accuracy of 86.3%, also showed competitive stats: precision of 85.5%, sensitivity of 87.2%, F1 value of 86.3%, and an AUC-ROC of 90.9%. MobileNetV3, clocking an accuracy of 83.1%, maintained commendable figures: 82.3% precision, 83.9% sensitivity, 83.1% F1, and an 88.2% AUC-ROC. The Capsule Network, registering 84.9% accuracy, showcased robust metrics: precision of 84.2%, sensitivity of 85.6%, F1 measure of 84.9%, and an AUC-ROC of 89.6%. However, stealing the limelight was the Vision Transformer (ViT) with a staggering 97.8% accuracy. With a precision of 96.5%, a sensitivity of 98.1%, an F1 value of 97.3%, and an AUC-ROC of 99.4%, it sets a benchmark in skin lesion diagnostics, as visualized in Fig. (5).

In this segment, we delve into a meticulous assessment of the role Vision Transformers (ViTs) play within our DEEP-SCAN framework for dermatological imagery. Initially crafted for broader computer vision challenges, ViTs have emerged as frontrunners in the niche arena of dermatology. Contrary to established Convolutional Neural Networks (CNNs) that sometimes falter at discerning nuanced patterns and overarching image contexts in dermal anomalies, ViTs thrive in these specifics. Our trials underscore the pivotal benefits of embedding ViTs within DEEPSCAN. Remarkably, the model clocked an accuracy of 97.8% on the validation set. This achievement is significant, given it outstrips the metrics of conventional CNN-driven models in dermatological diagnostics. DEEPSCAN, empowered by ViTs, eclipses traditional techniques by a noteworthy difference, underscoring its potential to reset accuracy and trustworthiness standards in skin anomaly detection. Pivoting to other performance indicators like precision, sensitivity, F1 measure, and the curve’s under-region (AUC-ROC), DEEPSCAN with ViTs consistently surpassed CNN-driven models. Specifically, its precision and sensitivity stood at 96.5% and 98.1%, in contrast to CNN models’ figures hovering around the 86-87% range for precision and 86-88% for sensitivity. The elevated precision signals DEEPSCAN’s prowess in curbing incorrect positives, while its heightened sensitivity indicates proficiency in pinpointing genuine positives. The F1 measure, a synthesis of precision and sensitivity, leaned favorably towards the ViT-imbued DEEPSCAN at 97.3%, symbolizing its capacity to sustain elevated precision and sensitivity concurrently. The AUC-ROC, representing DEEPSCAN’s discernment capabilities, touched a commendable 99.4% with ViTs, hinting at its robustness in lesion differentiation. To encapsulate, our comprehensive trials and scrutiny emphasize that ViTs’ integration within DEEPSCAN markedly amplifies its dermatological diagnostic capabilities. Stellar metrics across accuracy, precision, sensitivity, F1, and AUC-ROC with ViTs accentuate their transformative potential in dermatological imagery, promising enhanced diagnosis accuracy and dependability. This stride forward heralds better patient experiences and outcomes, positioning ViT-infused models like DEEPSCAN as indispensable assets for skin care specialists.

Fig. (5). Performance comparison of various algorithms for skin lesion diagnostics.