SCC-Net: A lightweight attention-enhanced deep learning model for automated squamous cell carcinoma detection

Abstract

Objectives

The primary objective of this study was to develop and validate SCC-Net, a lightweight GhostNet-CBAM model, for the early and accurate detection of cutaneous squamous cell carcinoma (cSCC) from dermoscopy images, addressing the limitations of resource-constrained clinical settings.

Methods

We developed SCC-Net, a lightweight GhostNet-CBAM model trained on 5150 public dermoscopy images (628 cSCC) formulated as a binary classification task (cSCC vs. non-cSCC). Using five-fold cross-validation and external testing against five CNN baselines, evaluating accuracy, recall, F1, AUC, and Grad-CAM interpretability.

Results

SCC-Net achieved 96.6% accuracy, 88.7% recall, and 0.986 AUC internally, and 94.3% accuracy, 93.8% recall, and 0.986 AUC externally, outperforming all comparables with only 1.59 M parameters. Grad-CAM heatmaps aligned with expert lesion focus.

Conclusion

SCC-Net delivers accurate, explainable, and low-cost cSCC detection, promising for teledermatology and resource-limited settings.

Introduction

Cutaneous squamous cell carcinoma (cSCC) is the second most common non-melanoma skin cancer, originating primarily from the keratinocytes of the epidermis. ¹ Its pathogenesis is typically associated with prolonged ultraviolet radiation exposure, immunosuppression, and chronic inflammation. ² According to the World Health Organization, cSCC accounts for approximately 20–50% of all skin malignancies, with an incidence rate second only to basal cell carcinoma. ³ With the global aging population, the prevalence of cSCC is rising annually. ⁴ Although generally treatable, delayed diagnosis can lead to deep tissue invasion and metastasis to lymph nodes or distant organs, resulting in significant morbidity and mortality (4000–8000 deaths annually).^5,6

Currently, the standard diagnosis of cSCC relies on clinical examination followed by histopathological confirmation. Dermatologists typically inspect lesions for morphological cues—such as irregular boundaries, color variation, and scaling—often utilizing dermoscopy to aid visualization. ⁷ However, this assessment is inherently subjective and dependent on the clinician's expertise, making the accurate differentiation of early-stage cSCC from benign keratosis challenging. While tissue biopsy remains the gold standard, ³ it is invasive, time-consuming, and can delay definitive treatment.

The early recognition of skin cancer is crucial for patient prognosis. Early stage cSCC is usually confined to the epidermis or superficial dermis, grows slowly and does not easily metastasize, and usually presents as nodules, ulcers, or scaly lesions, which are easy to confuse clinically with other skin disorders, such as benign nevi, skin ulcers, and so on.^8,9 Treatment is based on surgical excision, which has a high cure rate and low risk of recurrence. ¹⁰ Studies have shown that patients with early squamous skin cancer have a 5-year survival rate of >95%. ¹¹ In addition, advanced tumors have the potential to metastasize to the lymph nodes or other distant organs, especially in individuals with compromised immune function and chronic exposure to ultraviolet light. This complicates treatment and can lead to a range of complications, such as infection, bleeding, or functional impairment of bones and internal organs. ⁹ Therefore, the early and precise identification of cutaneous SCC (cSCC) is essential for clinical examination and treatment planning.

Deep learning is a branch of artificial intelligence that draws on computational methods of the human brain to extract patterns from raw data through computational networks. Over the past decade, the application of deep learning in image analysis, particularly for classification, segmentation, and recovery, has developed rapidly. It is widely used in tasks such as tumor detection, organ segmentation, and lesion region identification.^12,13 Deep learning can automatically determine the type of lesion by learning a large number of annotated images of skin lesions and can identify tiny features of the lesion, such as the shape of the spot, the change in its color, and the boundary between the lesion and the surrounding normal skin. Recently, with an increase in datasets and computational power, deep learning has made significant progress in the early identification of skin cancers. Some deep-learning-based diagnostic tools have been able to match the diagnostic results of dermatologists and have even demonstrated higher sensitivity and accuracy in some studies. ¹⁴ In addition, deep learning can be used to provide more comprehensive diagnostic advice to doctors by combining it with other clinical data (e.g., patient age, sex, and medical history) for multidimensional analysis. ¹⁵ While deep learning has achieved remarkable success in melanoma detection, dedicated research on cSCC remains relatively sparse compared to other skin malignancies. Previous studies have explored cSCC detection using heavy architectures; for instance, Ramos-Briceño et al. utilized a dermoscopically informed deep learning model to distinguish cSCC from actinic keratosis with high performance, ¹⁶ while Thwin et al. employed an ensemble of InceptionV3, VGG16, and ResNet-50 networks to achieve robust multi-class skin lesion classification (including cSCC). ¹⁷ However, these models typically rely on substantial computational resources and large parameter volumes (e.g., hundreds of millions of parameters in ResNet-50, InceptionV3, and VGG16 backbones), limiting their deployment in resource-constrained settings such as handheld dermoscopes or mobile teledermatology apps. Furthermore, cSCC lesions often present with subtle features—such as irregular vascularity and scale—that can be easily confused with benign keratosis, necessitating a model that can focus on fine-grained morphological details rather than just global texture. Therefore, we developed an algorithm to intelligently diagnose SCC.

The significance of deep learning in clinical medicine is not only reflected in the improvement of diagnostic efficiency and accuracy but also lies in its far-reaching impact on the optimization of medical resources, personalization of patient treatment, and the promotion of medical fairness. First, the deep learning-based intelligence-assisted diagnosis model has cross-region and cross-time consistency, which can reduce the variability of the diagnostic process, and is especially important for patients in remote areas and those with insufficient medical resources. This technology improves the accessibility and stability of diagnosis through referrals, providing convenience and reducing waiting times. In teledermatology applications, deep-learning classifiers have been shown to approach the accuracy of dermatologists (0.63) and outperform primary care physicians (0.45) across a wide range of skin conditions, thereby contributing to earlier assessment and treatment.^18–20 Second, the traditional diagnostic process for SCC includes patient consultation, pathology slides, and histology analysis, which usually take a certain amount of time to arrive at a diagnostic result. With deep-learning-assisted diagnosis, doctors can quickly obtain a preliminary judgment about the lesion, significantly saving diagnostic time. ²¹ Additionally, intelligence-assisted diagnosis can effectively reduce healthcare costs. ²² As this deep learning model can automate the analysis of lesion images, it reduces the workload of manual diagnosis, which in turn can reduce the labor and time costs in the diagnosis process. Finally, the deep learning intelligence-assisted diagnostic model for cSCC is not only important for the diagnosis of individual cases, but also promotes the intelligent transformation of the entire field of dermatology. This technology helps promote the standardization of dermatology diagnostic methods, while the introduction of AI technology allows dermatologists to focus more on the complex diagnosis and treatment of diseases, such as the analysis of etiology and the development of treatment plans, thereby improving the overall level of medical care. ²³

This study highlights the potential of deep learning in transforming the diagnosis of squamous cell carcinoma (SCC) by offering a more efficient, accurate, and accessible diagnostic approach. By integrating a lightweight GhostNet architecture with CBAM attention mechanisms, SCC-Net effectively enhances feature extraction while maintaining computational efficiency. The model achieves superior performance in both internal validation and external testing, outperforming conventional architectures in accuracy, recall, and AUC, demonstrating its robustness and reliability in clinical applications. Furthermore, Grad-CAM visualization confirms that SCC-Net focuses on clinically relevant lesion areas, reinforcing its diagnostic credibility and potential as an explainable AI tool. These findings suggest that SCC-Net could serve as a valuable assistant for dermatologists, facilitating early detection, improving diagnostic consistency, and expanding access to specialized care in resource-limited settings.

Methods

Study design and setting

This retrospective study was jointly conducted at The Second Hospital of Dalian Medical University and The First Affiliated Hospital of Dalian Medical University, Dalian, China. The research project, including data acquisition, model development, and validation, spanned from January 2024 to May 2025.

Data collection

The dataset employed in this study was derived from the International Skin Imaging Collaboration (ISIC) Archive, specifically integrating cohorts from the ISIC 2019 and 2020 challenges.^24,25 The total dataset comprised 5150 dermoscopy images. To align with the clinical objective of targeted screening, the problem was formulated as a binary classification task. The positive class consisted of 628 biopsy-verified cutaneous squamous cell carcinoma (cSCC) images, while the negative control class (n = 4522) aggregated diverse non-cSCC lesions, explicitly encompassing melanoma, basal cell carcinoma, nevi, dermatofibroma, vascular lesions, actinic keratosis, pigmented benign keratosis, and seborrheic keratosis. This binary framework (cSCC vs. non-cSCC) was designed to evaluate the model's specific sensitivity to squamous cell carcinoma against a background of confounding skin conditions (Figure 1).

Figure 1.

Representative dermoscopic comparison between cutaneous squamous cell carcinoma (cSCC) and morphologically similar lesions. (a) A typical cSCC presenting with central keratinization and ulceration. (b) A confounding benign lesion (e.g., seborrheic keratosis) sharing similar visual features. The distinct text labels distinguish the positive class (cSCC) from the negative class. Note: Images represent standard dermoscopic views acquired at fixed magnification; therefore, specific scale bars are not applicable to these digital archives.

The training set and internal validation set use the ISIC skin lesion dataset (https://www.kaggle.com/datasets/riyaelizashaju/isic-skin-disease-image-dataset-labelled), which is provided by the International Skin Imaging Collaboration (ISIC) and contains a variety of skin lesion categories and is annotated by experts. The image quality of this dataset is high, covering dermoscopic images and clinical close-up photos, which can effectively support the feature learning of deep learning models. The external validation set uses the Skin Cancer 9 Classes ISIC dataset (https://www.kaggle.com/datasets/nodoubttome/skin-cancer9-classesisic), which expands the category range and covers 9 distinct types of skin lesions: cutaneous squamous cell carcinoma, basal cell carcinoma, melanoma, nevus, dermatofibroma, vascular lesion, actinic keratosis, pigmented benign keratosis, and seborrheic keratosis. This dataset is used to independently test the performance of the model and evaluate its stability and generalization ability under different data distributions. As this study exclusively utilized de-identified images from public open-access repositories, the requirement for written informed consent was waived by the Institutional Review Board (IRB No. 2023-028), in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments.

To strictly prevent data leakage and ensure the independence of the external validation, a rigorous deduplication protocol was implemented prior to model training. Given that both the training and external test sets originate from the ISIC archive ecosystem, there is a potential risk of overlapping patient cases or identical images appearing in different sub-datasets (e.g., “Skin Cancer 9 Classes” containing images from ISIC 2019/2020). We performed a cross-reference check using unique image identifiers (ISIC ID) and perceptual hashing to identify potential duplicates. Any images in the external test set that appeared in the training or internal validation sets were strictly excluded. Consequently, the final external test set consists solely of samples that the model had never encountered during the optimization phase, ensuring an unbiased evaluation of its generalization capability.

Inclusion and exclusion criteria

Images were included in the study if they met the following criteria: (1) dermoscopic modality; (2) availability of definitive histopathological or expert-consensus ground truth labels; and (3) sufficient image resolution for analysis. The exclusion criteria were: (1) low-quality images with severe artifacts (e.g., blur, bubbles) that obscured identifying features; (2) images lacking diagnostic labels; and (3) duplicate images identified across training and testing sets to prevent data leakage.

Data preprocessing

In order to improve the training stability and generalization ability of the deep learning model, we standardized the dataset and used a series of data augmentation methods to optimize the training effect of the model. First, we normalized the size of all images and adjusted the images to 224 × 224 pixels to match the input requirements of the mainstream deep learning model. For the training set, we further performed data augmentation, including random cropping (RandomResizedCrop, 224) to enhance the model's ability to recognize lesions of different scales, random horizontal flipping (RandomHorizontalFlip, P = .5) to simulate the directional changes of skin images, and using normalization (Normalize) to normalize the images with mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225) to adapt to the input characteristics of the pre-trained neural network.

For the validation set and test set, only Resize (224 × 224) and Normalize processing are performed to ensure the stability and consistency of the data during the evaluation phase. The entire dataset adopts a 5-fold cross-validation strategy to ensure the reliability of the training and validation process and avoid overfitting problems. In addition, to reduce the impact of class imbalance, the Mixup data enhancement technology is used in the training process to improve the generalization ability of the model by weighted mixing of two samples, and the Label Smoothing (ε = 0.1) technology is introduced to prevent the model from overfitting the training samples.

Detection model construction

Proposed model—SCC-Net

This study proposed a deep learning-based intelligent detection model for squamous cell carcinoma (SCC) of the skin, named SCC-Net (squamous cell carcinoma neural network). The model uses a lightweight network structure and combines the attention mechanism to improve classification accuracy and computational efficiency. SCC-Net uses FBNet as the infrastructure and introduces the GhostNet lightweight module and CBAM mechanism. This specific combination was selected to address the distinct morphological heterogeneity of cSCC. The GhostNet module efficiently captures intrinsic feature maps of scales and keratin pearls with reduced redundancy, while the CBAM mechanism dynamically recalibrates these features to emphasize lesion boundaries that are often indistinct in early-stage cSCC.

The core structure of SCC-Net consists of multiple convolutional modules, among which the Ghost module generates redundant feature maps through cheap operations, reducing the amount of calculation while maintaining the ability to express features, effectively improving the adaptability of the model on mobile terminals and resource-constrained environments. At the same time, the CBAM mechanism is introduced in the key layer. First, the global feature weights are calculated through channel attention (CA) to strengthen the model's attention to important lesion features. Secondly, the feature map is further optimized through spatial attention (SA), so that the model can more accurately capture the skin lesion area and reduce the interference of irrelevant background information. In addition, SCC-Net retains the global pooling layer of FBNet in the feature extraction backend to reduce the number of parameters. Finally, the fully connected layer was configured with two output nodes to execute the binary classification task, explicitly distinguishing cutaneous squamous cell carcinoma from all other skin lesion categories.

During model training, the cross-entropy loss function (Cross-Entropy Loss) is used for optimization, and label smoothing (Label Smoothing, ε = 0.1) technology is introduced to reduce the risk of overfitting. The Adam optimizer (learning rate = 0.001) is used for training, combined with the ReduceLROnPlateau learning rate decay strategy to ensure that the model converges robustly during training. The Mixup technology (α = 0.4) is used for data enhancement to mix samples and improve the generalization ability of the model.

Comparison model

In order to comprehensively evaluate the performance of SCC-Net, this study selected five classic deep learning models for comparative analysis, including CSPNet, DenseNet-121, FBNet, RegNet, and ResNet-50. These models cover different architecture designs, including both lightweight networks and deeper residual structure networks, which can evaluate the advantages of SCC-Net in cSCC detection tasks from different angles.

CSPNet (Cross Stage Partial Network) adopts the cross-stage partial connection strategy, which splits the feature map into two parts: one part is directly transmitted, and the other part is calculated through convolution, and finally fused in the deep layer, reducing redundant calculations while enhancing feature expression capabilities. Its architecture can improve computational efficiency and optimize gradient propagation, so that it has high accuracy in complex image classification tasks. In order to adapt to the skin lesion classification task, we use the CSPDarkNet53 variant and use a learning rate of 0.001 for training, and use the SGD optimizer (momentum = 0.9, weight_decay = 5e-4) to maintain training stability.

DenseNet-121 (Densely Connected Convolutional Networks) uses a dense connection strategy so that the features of the previous layer can be directly transferred to all subsequent layers, thereby alleviating the gradient vanishing problem, improving feature reuse, and reducing the number of parameters. In this study, we chose DenseNet-121 as a comparison model, which has 4 Dense Blocks, a total of 121 layers of convolution, and uses pre-trained weights (ImageNet initialization) for fine-tuning to improve classification performance. The optimizer uses Adam (learning rate = 0.0001) and combines ReduceLROnPlateau for learning rate decay to prevent the model from falling into a local optimum.

FBNet (Fast and Accurate Network) is a lightweight CNN structure that uses a search strategy based on MobileNetV3 and automatically optimizes the network architecture through Neural Architecture Search (NAS) to achieve a good balance between computational efficiency and accuracy. FBNet has low computational complexity and parameter count, and fast inference speed on mobile devices. We chose the FBNet-C version, which has a FLOPs as low as 325 M, but still maintains high classification accuracy. During training, FBNet uses a learning rate of 0.002 and the AdamW optimizer for optimization, and combines Mixup data enhancement to enhance the model's generalization ability.

RegNet (Regularized Network) uses a channel allocation strategy based on logarithmic regularization, which can automatically generate the optimal network structure according to computing resources. This study selected RegNetY-4GF as the comparison model, which contains 20 convolution blocks and a total parameter volume of about 20 M. At the same time, the SE (Squeeze-and-Excitation) module is used to optimize channel attention and improve the effectiveness of feature extraction. During the training process, RegNet uses a learning rate of 0.0005, Batch Size = 16, and uses the SGD optimizer (momentum = 0.9) for gradient updates.

ResNet-50 is a classic deep residual network (ResNet, Residual Network). It uses the Skip Connection mechanism, which enables deep networks to effectively alleviate the gradient vanishing problem and perform well in various image classification tasks. We use ResNet-50 as the benchmark model, which contains 50 deep convolutional layers, a total of 23.5 M parameters, and uses ImageNet pre-trained weights for fine-tuning. During training, a learning rate of 0.001 and a batch size of 32 are used, and data enhancement strategies (random flipping, color jitter) are combined to optimize model performance.

Model evaluation

This study used multiple indicators to comprehensively evaluate the performance of the model, including accuracy, recall, F1-score, precision, model parameters and area under the curve (AUC) to ensure a comprehensive analysis of the cutaneous squamous cell carcinoma (SCC) detection model.

Accuracy is used to measure the proportion of correct classifications of the model on the entire test set and is an important indicator of overall performance. Recall focuses on the model's ability to identify positive samples (squamous cell carcinoma), which is particularly suitable for clinical tasks to ensure that missed diagnoses are minimized. F1-score combines precision and recall to provide a balanced indicator, which is particularly suitable for classification tasks with uneven data distribution. Precision reflects the proportion of samples predicted by the model as positive that are truly positive, which is crucial to reducing the risk of misdiagnosis. In terms of computing resources and deployment optimization, we count the number of model parameters to evaluate the complexity of the model and examine its applicability in resource-constrained environments. In addition, we calculate AUC (area under the curve) through the ROC curve (Receiver Operating Characteristic Curve) to measure the comprehensive discrimination ability of the model at different classification thresholds. The closer the AUC value is to 1, the stronger the model's discrimination ability is.

Statistical analysis

Continuous variables were summarized as means and standard deviations, while categorical variables were expressed as frequencies and percentages. To rigorously determine whether the performance improvements of SCC-Net over the baseline architectures (e.g., ResNet-50, CSPNet) on the external test set were statistically significant, McNemar's test was employed to compare the paired proportions of correct and incorrect predictions. A two-sided P-value of less than .05 was considered to indicate a statistically significant difference. All statistical calculations and visualizations were implemented using the SciPy and Scikit-learn libraries in Python.

Results

Experimental setup

To ensure the generalization capability of the model and mitigate the impact of class imbalance, we implemented a stratified five-fold cross-validation strategy. This method ensures that the relative frequency of cSCC and non-cSCC lesions remains consistent across all training and validation folds, thereby preventing sampling bias. Specifically, the internal dataset was randomly partitioned into five subsets; in each iteration, four subsets (80%) were utilized for training and one subset (20%) for internal validation. All experiments were conducted on an NVIDIA RTX 3090 GPU (24GB). The deep learning framework was implemented using PyTorch (v1.12.1), with image preprocessing and augmentation performed via Torchvision (v0.13.1) and Albumentations. Model performance metrics were calculated using Scikit-learn (v1.1.2), and data visualization was handled by Matplotlib and Seaborn. The training process uses the Adam optimizer (learning rate = 0.001, weight_decay = 1e-4), combined with the ReduceLROnPlateau learning rate adjustment strategy (factor = 0.1, patience = 3) to achieve dynamic learning rate adjustment and avoid overfitting problems. The training batch size (Batch Size) is set to 32.

Model training

During the training process of SCC-Net, the number of training rounds (Epochs) for each fold is set to 30, and the loss change (Loss) and validation set accuracy (Validation Accuracy) are recorded in the training and validation stages. From the training curve (Figure 2), it can be seen that the model converges quickly in the first few rounds of training, and the validation set accuracy rises rapidly to more than 95% in the first 5 rounds, indicating that SCC-Net has strong learning ability and good feature extraction ability. As the training progresses, the validation set accuracy increases steadily and stabilizes after the 20th round, eventually reaching a level close to 98%. The loss curve shows that both the training loss (Training Loss) and the validation loss (Validation Loss) show a steady downward trend, indicating that the model is constantly optimizing feature representation while avoiding serious overfitting problems. It is worth noting that in the middle and late stages of training, the loss curve is relatively smooth, further indicating that SCC-Net can effectively learn data distribution during training and maintain a high generalization ability. Overall, SCC-Net exhibits fast convergence, high stability and good generalization ability during training.

Figure 2.

Training iteration curve of SCC-Net, where (a) is the loss curve and (b) is the accuracy change curve.

Model results

Five-fold cross-validation results

This study uses five-fold cross-validation to evaluate the performance of SCC-Net and multiple comparison models in the detection task of cSCC, and calculates accuracy, recall, F1-score, precision, and parameters as evaluation indicators (results are shown in Table 1). The experimental results show that SCC-Net achieves the best performance in all core indicators, especially in recall (0.8867), F1-score (0.9112), and precision (0.9412), which is significantly better than comparison models such as CSPNet, DenseNet-121, FBNet, RegNet, and ResNet-50.

Table 1.

Evaluation comparison of each model in five-fold cross-validation.

Model	Accuracy	Recall	F1-score	Precision	Parameters
CSPNet	0.9425	0.814	0.8325	0.8591	21,616,168
DenseNet-121	0.9495	0.8587	0.8534	0.8552	6,922,906
FBNet	0.9627	0.8634	0.8832	0.9074	3,591,170
RegNet	0.9608	0.8502	0.8659	0.8926	3,904,026
ResNet-50	0.9262	0.7736	0.7926	0.8166	23,512,130
SCC-Net	0.9656	0.8867	0.9112	0.9412	1,591,286

In terms of computational complexity, SCC-Net has only 1.59 M parameters, which is significantly lighter than ResNet-50 (23.5 M) and CSPNet (21.6 M). It also performs well in AUC scores, with the average AUC of five-fold cross-validation reaching 0.9791, showing excellent classification ability. The AUC scores of the comparison models are shown in Figure 3, where the ROC curve of SCC-Net is close to the upper left corner, indicating that it has extremely high discrimination ability in distinguishing squamous cell carcinoma from other skin lesions. Overall, SCC-Net has achieved a good balance between high accuracy, computational efficiency, and generalization ability, providing a reliable solution for intelligent detection of cSCC.

Figure 3.

Receiver Operating Characteristic (ROC) curves illustrating the performance of five-fold cross-validation on the internal dataset. The colored curves represent the performance of individual folds, while the black bold curve indicates the mean ROC across all five folds. To ensure clarity, the specific Mean Area Under the Curve (AUC) values for each model are explicitly listed as follows: (a) CSPNet (Mean AUC = 0.8881), (b) DenseNet-121 (Mean AUC = 0.9617), (c) FBNet (Mean AUC = 0.9694), (d) RegNet (Mean AUC = 0.9673), (e) ResNet-50 (Mean AUC = 0.9268), and (f) SCC-Net (Mean AUC = 0.9791). Note that SCC-Net (f) demonstrates the highest discriminative ability with the curve closest to the upper-left corner.

External test results

In order to further verify the generalization ability of SCC-Net in a real clinical environment, we conducted an independent evaluation on an external test set and compared the performance of models such as CSPNet, DenseNet-121, FBNet, RegNet, and ResNet-50. The experimental results show (Table 2) that SCC-Net outperforms other models in all core indicators, among which accuracy (0.9432), recall (0.9377), F1-score (0.9333), and precision (0.9293) all reach the highest level, indicating that SCC-Net is more robust on external data.

Table 2.

Evaluation comparison of each model in external testing.

Model	Accuracy	Recall	F1-score	Precision
CSPNet	0.9232	0.8903	0.9053	0.9253
DenseNet-121	0.9155	0.8949	0.8989	0.9032
FBNet	0.9339	0.9253	0.9223	0.9194
RegNet	0.9186	0.9215	0.9067	0.8958
ResNet-50	0.9094	0.8819	0.8902	0.9001
SCC-Net	0.9432	0.9377	0.9333	0.9293

In the ROC curve analysis (Figure 4), the AUC (0.9864) of SCC-Net is significantly higher than that of other models, and the curve is close to the upper left corner, indicating that the model has a very high ability to distinguish squamous cell carcinoma from other skin lesions. In addition, SCC-Net's advantage in recall rate shows that it can effectively reduce the risk of missed diagnosis, while the higher precision ensures a lower misdiagnosis rate, making it more valuable in clinical applications. Overall, SCC-Net performed stably on external test data, further verifying the generalization ability and clinical applicability of the model.

Figure 4.

ROC curves evaluating the generalization performance on the independent external test set. The blue solid line represents the model's performance, and the red dashed line indicates the reference line for random chance (AUC = 0.5). The corresponding AUC values for each architecture are listed below to supplement the graphical data: (a) CSPNet (AUC = 0.9311), (b) DenseNet-121 (AUC = 0.9583), (c) FBNet (AUC = 0.9714), (d) RegNet (AUC = 0.9722), (e) ResNet-50 (AUC = 0.9583), and (f) SCC-Net (AUC = 0.9864). SCC-Net achieves the highest AUC, confirming its robustness in distinguishing cSCC from other lesions in an independent cohort.

To rigorously validate the classification reliability, we generated confusion matrices for all models using the external test set (Figure 5). As illustrated in Figure 5(a), SCC-Net demonstrated superior discriminative capability, correctly identifying 182 cSCC cases (True Positives) and 432 non-cSCC cases (True Negatives). In the context of skin cancer screening, minimizing false negatives (missed diagnoses) is clinically paramount. SCC-Net achieved a remarkably low false negative count of 15, significantly outperforming heavier architectures such as ResNet-50 (37 missed cases, Figure 5(f)) and CSPNet (38 missed cases, Figure 5(d)). Furthermore, statistical analysis using McNemar's test confirmed that the predictive accuracy of SCC-Net was statistically superior to ResNet-50 (P < .05) and CSPNet (P < .05). These results suggest that the attention-enhanced lightweight architecture effectively captures subtle morphological features of cSCC, thereby reducing the risk of diagnostic omission compared to conventional deep learning baselines.

Figure 5.

Confusion matrices comparing the classification performance of SCC-Net and state-of-the-art models on the external independent test set. (a) SCC-Net, (b) FBNet, (c) DenseNet-121, (d) CSPNet, (e) RegNet, and (f) ResNet-50. The rows represent the true diagnostic labels (0: Non-cSCC, 1: cSCC), while columns represent predicted labels. The diagonal elements (dark blue) indicate correct predictions. Notably, SCC-Net (a) demonstrates the lowest number of false negatives (15) among all architectures, highlighting its sensitivity in detecting malignant lesions.

Grad-CAM visualization analysis

To interpret the model's decision-making process, we systematically analyzed the correspondence between high-attention regions (indicated by peak saliency in the heatmaps) and specific dermoscopic structures. As illustrated in the representative examples in Figure 6, SCC-Net demonstrates a highly selective focus mechanism. Quantitatively, the high-attention tiles consistently align with clinically predictive features of cSCC, specifically central keratinization (hyperkeratosis), ulceration, and irregular vascular patterns (Figure 6(a)). In contrast, low-attention areas predominantly correspond to healthy surrounding skin, hair artifacts, or nonspecific background textures.

Figure 6.

Grad-CAM visualization of decision saliency in cSCC detection. Columns represent the original dermoscopic images and their corresponding class activation heatmaps. (a) The model's attention (warm colors) is strictly localized to the central hyperkeratosis and ulceration. (b) The heatmap highlights the irregular erythema and scaling while effectively suppressing background artifacts, confirming the model's reliance on clinically relevant morphological features.

This focal attention is distinct from the patterns typically observed in benign lesions, where models often rely on global pigment network distribution. For instance, in Figure 6(b), the network effectively suppresses the surrounding clear skin and explicitly targets the erythematous base with irregular boundaries and scaling. This suggests that SCC-Net has learned to identify “scaly and ulcerated” morphological traits as key discriminators for cSCC diagnosis, rather than relying on confounding contextual bias. This alignment with expert dermatological criteria—prioritizing structural disruption over color intensity alone—validates the model's potential as an explainable, clinically reliable assistant.

Discussion

This study proposed a deep learning-based cSCC intelligent detection model SCC-Net and evaluated its performance through five-fold cross-validation and external testing. SCC-Net combines the lightweight structure of GhostNet and the CBAM attention mechanism to enhance the ability to focus on key lesion areas while maintaining computational efficiency. Experimental results show that SCC-Net outperforms the comparison models in terms of accuracy, recall, F1-score, and AUC, especially in the recall (0.9377) and AUC (0.9864) indicators, indicating its high reliability in the detection task of cSCC. In addition, we visualized the decision-making mechanism of SCC-Net through Grad-CAM. The results showed that the model can accurately focus on the squamous cell carcinoma lesion area and effectively avoid background interference, which is highly consistent with the clinical diagnostic characteristics. Overall, this study demonstrated the feasibility and clinical application potential of SCC-Net in intelligent auxiliary diagnosis of cSCC.

Our study advances beyond previous attempts at automated cSCC detection by addressing the trade-off between diagnostic accuracy and computational efficiency. Unlike prior works that prioritized accuracy through computationally intensive backbones,^26,27 SCC-Net demonstrates that a lightweight architecture enhanced with attention mechanisms can achieve comparable or superior performance. The integration of CBAM is particularly pertinent for cSCC, as the channel and spatial attention modules compel the network to prioritize clinically relevant regions—specifically the central ulceration and peripheral erythema characteristic of invasive SCC—while suppressing background noise from healthy skin or hair. This architectural choice effectively bridges the gap between the high performance required for clinical screening and the low-latency requirements of point-of-care devices.

The superior performance of SCC-Net in the detection task of cutaneous squamous cell carcinoma (SCC) is mainly attributed to its innovative architecture design and optimization strategy. First, the model adopts the GhostNet lightweight feature extraction module, which generates redundant features through cheap operations, reduces the amount of calculation while maintaining efficient feature expression capabilities, making SCC-Net have higher computational efficiency and lower parameter volume, only 1.59 M, which is much lower than the traditional ResNet-50 (23.5 M) and CSPNet (21.6 M), but still achieves better classification performance. Secondly, the model integrates the CBAM (Convolutional Block Attention Module) mechanism, strengthens lesion-related features through channel attention (CA), and uses spatial attention (SA) to further optimize the model's focus area, so that it can accurately locate the SCC lesion area and reduce the misjudgment of background information. In addition, this study used Mixup data enhancement technology to improve the model's adaptability to different data distributions, introduced label smoothing (Label Smoothing, ε = 0.1) to reduce the risk of overfitting, and combined with adaptive learning rate adjustment (ReduceLROnPlateau) to optimize the training process, so that SCC-Net can converge quickly and maintain stable performance in fewer training rounds. Compared with the traditional CNN structure, SCC-Net also uses five-fold cross-validation for training to minimize data segmentation bias, and further verifies the generalization ability on an external test set. The experimental results show that SCC-Net is superior to the comparison model in terms of recall rate, F1-score and AUC, indicating that it has a lower misdiagnosis and missed diagnosis rate while accurately detecting squamous cell carcinoma, ensuring its clinical feasibility. The combination of these optimization strategies enables SCC-Net to achieve better classification results while ensuring efficient calculation.

The superior performance of SCC-Net is of great significance in clinical applications, especially in improving the accuracy of early screening, optimizing the doctor's diagnostic process, and assisting primary medical institutions. Squamous cell carcinoma (SCC) is one of the most common non-melanoma skin cancers. Although its mortality rate is lower than that of melanoma, if it is not accurately diagnosed and intervened in time at an early stage, it may lead to deep tissue invasion and even distal metastasis. Traditional SCC diagnosis relies on the experience of dermatologists, and a preliminary judgment is made through visual examination combined with dermoscopy, and then confirmed by biopsy and pathological analysis. However, this process is not only time-consuming, but also depends on the professional level of the doctor. Doctors with different experience levels have certain differences in the interpretation of dermoscopic images, which may lead to misdiagnosis or missed diagnosis. Therefore, developing an efficient, accurate, and scalable SCC intelligent detection model will help optimize the current clinical diagnosis system.

SCC-Net performed well in terms of recall (0.9377) and AUC (0.9864), indicating that it can effectively reduce the risk of missed diagnosis in clinical applications, which is crucial for the early screening of squamous cell carcinoma. Especially in primary medical institutions or inexperienced doctor environments, SCC-Net can be used as an auxiliary tool to provide preliminary imaging analysis results to help doctors quickly identify high-risk lesions, thereby deciding whether further biopsy is needed and improving screening efficiency. At the same time, in the field of telemedicine, SCC-Net can also be used as an artificial intelligence-assisted diagnosis system, combined with mobile devices or cloud platforms, to help patients complete preliminary self-examination at home, reduce unnecessary hospital visits, and improve the utilization of medical resources. In addition, SCC-Net adopts a lightweight architecture (only 1.59 M parameters), which greatly reduces computing costs while ensuring high performance, making it feasible for deployment on mobile terminals or portable medical devices, further broadening its clinical application scenarios. It is worth noting that the Grad-CAM interpretability analysis of SCC-Net shows that the model focuses on key areas of squamous cell carcinoma lesions during the classification process, such as ulcers, scales, abnormal vascular proliferation, and other features, which are highly consistent with the diagnostic basis of clinicians. This feature not only increases the explainability of the model, allowing doctors to trust AI prediction results, but can also be used as a medical teaching and training tool to help young doctors better understand the imaging manifestations of SCC and shorten the learning curve. Therefore, SCC-Net is not only an efficient auxiliary screening tool but also can play a role in multiple fields such as medical education, telemedicine, and intelligent image analysis, providing strong support for early detection, accurate diagnosis, and personalized treatment of squamous cell carcinoma. Beyond the initial detection, it is crucial to contextualize the role of SCC-Net within the broader landscape of cSCC management. Since approximately 95% of cSCC cases can be effectively cured by surgical excision, the paramount clinical challenge lies in the early identification and risk stratification of lesions prone to recurrence or metastasis. Recent advancements in computational histopathology have provided valuable insights into this process. For instance, Knuutila et al. demonstrated that combining AI analysis of hematoxylin and eosin-stained slides with clinical risk factors significantly enhances metastasis prediction. ²⁸ Similarly, recent deep learning frameworks developed by Peleva et al. and Pisula et al. utilize whole slide images (WSI) to identify high-risk prognostic features, such as tumor boundary heterogeneity and specific differentiation patterns.^29,30 While SCC-Net is primarily designed for dermoscopic screening rather than histological prognostication, it serves as the essential gatekeeper in this clinical workflow. By accurately flagging suspicious lesions at the screening stage—particularly in primary care or teledermatology settings—SCC-Net ensures that patients are promptly referred for biopsy, thereby facilitating the subsequent application of these advanced pathology-based risk stratification models.

Although SCC-Net performs well in the detection task of cutaneous squamous cell carcinoma (SCC), it still has certain limitations. First, while we implemented strict de-duplication protocols to prevent data leakage between training and testing subsets, the current model development relies primarily on the ISIC archive. Consequently, the shared dermoscopic acquisition protocols across ISIC subsets may introduce a degree of image homogeneity that does not fully reflect the variability of raw images captured in non-standardized clinical settings. To rigorously evaluate the model's robustness across distinct imaging environments, future work will prioritize validation on completely independent datasets, such as the MILK10K collection, as well as proprietary multi-center clinical cohorts. This will help verify the model's adaptability to different device manufacturers and lighting conditions. Secondly, the current study relies only on imaging data and does not integrate multimodal data such as clinical history and pathological information, which may be limited in the identification of early or atypical cases. Although SCC-Net combined with Grad-CAM enhances interpretability, more fine-grained feature-level interpretability methods are still needed to improve doctors’ trust in AI predictions. In addition, this study has not been deployed and verified in a real clinical environment. In the future, it needs to be combined with mobile or clinical system testing to ensure the stability and practicality of the model in practical applications. In summary, SCC-Net shows high clinical potential in the intelligent detection of SCC, but it still needs to further optimize the data scale, multimodal fusion, and clinical testing to promote its development into an applicable intelligent auxiliary diagnosis system.

Conclusion

This study proposed a deep learning-based cSCC intelligent detection model SCC-Net, which combines the GhostNet lightweight structure and CBAM attention mechanism to enhance the ability to focus on the lesion area while maintaining computational efficiency. Experimental results show that SCC-Net achieved excellent classification performance in both five-fold cross-validation and external testing, especially in key indicators such as recall rate (0.9377), F1-score (0.9333), and AUC (0.9864), which is superior to traditional models and shows good generalization ability. In addition, Grad-CAM visualization analysis further verifies the accuracy of SCC-Net in lesion area identification, which is highly consistent with the clinical judgment of dermatologists. The results of this study show that SCC-Net has important potential in early screening, clinical auxiliary diagnosis, and telemedicine applications of SCC, especially for primary medical institutions and resource-constrained environments. It can be used as an intelligent auxiliary tool for doctors to improve diagnostic efficiency and reduce the risk of missed diagnosis and misdiagnosis.