A Comparative Study of Deep Learning Models for Skin Cancer Detection: Leveraging Transfer Learning

Skin cancer represents a great concern for global health expressed in a number of more than 3.5 million diagnosed cases per year, which include melanoma, basal cell carcinoma, and squamous cell carcinoma. An astounding factually presented is that this exceeds cancers such as breast, lung, and colon combined and that melanoma kills one person a minute [1]. Thus, the burden of SC underlines the importance of early detection through materialized and efficient treatment. Catching the disease early is crucial for successful treatment. For instance, the ten-year survival rate of detected cases is 94% if diagnosed early; if the cancer progressed, as in case melanoma, into stage four that rate drops to 15% [2].

This fact has inevitable practical implications – numerous lives could be saved, and a substantial amount of expensive advanced care-related costs could be eliminated. However, early SC detection is problematic in terms of it being expensive. Skin lesions are often visually similar, which results in a difficulty to differentiate between benign and malignant sources. Therefore, a comprehensive examination is to be conduct. Traditionally, dermatologist practitioners utilize a medical device called a dermatoscope to record images of the lesion to take a closer look. Unfortunately, the tool is expensive, and not all practitioners have access to it, which stands as an obstacle in the way of cheap and effective early detection. Skin diseases’ appearance is visually difficult, as conditions look similar.

However, there is a way, which was develop due to current advancements in AI. Modern algorithms are capable of identifying visually similar skin conditions without the additional need for a medical tool. Instead, based on this algorithm, usual photos can be analyzed. The implementation of artificial intelligence in the context of SC detection in the healthcare sector was exceptionally promising. SC detection accuracy was increase by DL, which allows the assessment of complex patterns & features of dermatological images [3]. AI algorithms to infer potential signs of SC could carefully analyze skin images. In many cases, these systems demonstrate a sufficiently high ACC and may offer healthcare workers a second opinion, highlight fields of cold interest in the specific pictures, and explain potential diagnoses. Given the rapidly increasing spread of skin cancer around the globe and the urgent need for accurate diagnostics, this application of ACC and AI is highly applicable.

The focus of this research is to implement the thorough process of SC detection through image data analysis. The journey commences with accurate data processing, with the outline including the essential procedures of pre-processing and cleaning. Next is the dataset division into separate train and test sets, which paves the way for sustainable model training and evaluation.

Subsequently, we explore data modeling, using advanced methodologies to identify potentially interesting patterns and characteristics in the images.

Utilizing the above-mentioned algorithms and models, as well as making some critical changes in the hyperparameter tuning process, we aspire to develop a powerful and accurate model for SC detection. This work aims to attain not solely the best performance but also to contribute to other areas of medical image analysis and diagnostic tools.

Considerable research has focused on the classification of skin lesions from dermoscopic images, but limited work been conducted on classification using general images.

A notable research effort led by Esteva and colleagues [2] has emerged as one of the most comprehensive studies skin lesion classifications. They carefully classified lesions into 23 distinct categories, utilizing CNNs for the job. Specifically, they utilized the capabilities of VGG-16 and VGG-19 models [12]. Esteva and the team enhanced the performance of a pre-trained VGG model through transfer learning. Notably, they accomplished an impressive 90%.

ACC rate in binary classification, successfully distinguishing between cancerous and non-cancerous lesions.

In the field of melanoma detection, Moussa and colleagues adopted a unique approach, identifying the condition based on its geometric features. Moussa et al. [13] employed the k-Nearest Neighbors algorithm for classification. Despite the constraint of a limited dataset, their method demonstrated an impressive 89% ACC rate.

Alternative approaches have investigated the use of a Total Dermatoscopy Score (TDS) to distinguish between malignant and benign skin lesions. Azmi et al. [14] explored this method, assigning TDS scores by extracting features following the ABCD rule. The scores varied from 1.0 to 8.9, with a score above 5.45 indicating an elevated likelihood of melanoma.

Masood and collaborators led a research project as conducted by Masood et al. [15]. In this investigation, a set of 135 images underwent classification into cancerous and non-cancerous lesions using an Artificial Neural Network (ANN) system. The methodology involved the application of the Histogram Analysis-based Fuzzy C Mean Algorithm for Level Set Initialization to segment the images. Following the segmentation process, features were extracted, encompassing histogram features and statistical features derived from the Grey Level Co-occurrence Matrix (GLCM).

They employed a two-layer feed-forward neural network for classification, incorporating three distinct training algorithms: Levenberg-Marquardt, Scaled-Conjugate Gradient, and Resilient Back Propagation. Remarkably, the study attained an accuracy rate of 91.9% when utilizing the Scaled Conjugate Gradient training algorithm.

Many researchers have actively explored the computer vision approach for detecting SC. In the segmentation of skin lesions in input images, current systems often use manual, semi-automatic, or fully automatic border detection methods. These segmentation techniques depend on diverse features, such as texture, color, shape and luminance. The literature includes numerous border detection methods, including histogram thresholding [16], global thresholding on optimized color channels followed by morphological operations, and hybrid thresholding [17].

In our study, we applied an automatic thresholding and border detection method, utilizing various image-processing techniques to extract these features. For example, Garnavi [18] introduced an automated method for global border detection in dermoscopy images, relying on color-space analysis and global histogram thresholding. This method demonstrated high performance in detecting the borders of melanoma lesions. As reported by Celebi et al. [19], the approach included segmenting the input image into different clinically significant regions using the Euclidean distance transform to extract color and texture features.

The ABCD rule of dermoscopy emphasizes asymmetry as a fundamental feature, alongside border irregularity, Color, and diameter. Numerous studies have aimed to quantify asymmetry in skin lesions. Some approaches compute symmetry using geometric measurements across the entire lesion, incorporating metrics like symmetric distance and circularity [20]. Conversely, other research suggests the circularity index as a measure of border irregularity in dermoscopy images [21, 22].

Maglogiannis and Doukas [23] presented an overview of the most notable implementations in the literature. They compare the performance of various classifiers for the specific problem of skin lesion diagnosis, providing valuable insights into the field of SC detection.

According to Shorfuzzaman’s study [24], malignant melanoma, a highly fatal skin cancer originating from abnormal melanocyte proliferation, must be detected early for cure. In the present work, an explainable CNN-based stacked ensemble framework is suggested for automated melanoma diagnosis. Multiple CNN sub-models EfficientNetB0, Xception, and DenseNet121 are developed through transfer learning, and their forecasts are used by a separate meta-learner. The ensemble model was assessed using open-access data that included both benign and malignant melanoma images. An explainability strategy with Shapley adaptive explanations is used to generate heat maps that show the melanoma-suggestive zones, rendering the model more interpretable for dermatologists. The ensemble framework accomplishes the highest performance, reaching an accuracy of 95.76%, a sensitivity of 96.67%, and an AUC of 95.7%, surpassing individual models.

According to Tembhurne et al. [25], skin cancer is a prevalent disease, with melanoma being particularly deadly, causing six out of seven skin cancer-related deaths. To improve diagnosis accuracy and reduce false negatives, this paper presents a novel approach combining machine learning and deep learning techniques. The model uses neural networks for feature extraction and processes these features with methods like Contourlet Transform and Local Binary Pattern Histogram. Achieving an accuracy of 93% and recall scores of 99.7% for benign and 86% for malignant cases, the model was teste on a Kaggle dataset from the ISIC Archive. This ensemble method outperforms expert dermatologists and existing techniques, offering a valuable tool for accurate skin cancer detection.

Imran et al. [26] introduced an ensemble of deep learners, specifically VGG, CapsNet, and ResNet, for skin cancer detection. The proposed ensemble model outperforms individual models, achieving superior results in sensitivity, accuracy, specificity, F-score, and precision. The experimental results, with the proposed ensemble achieving an accuracy of 93.5%, sensitivity of 0.87, specificity of 0.84, F-score of 0.92, false-positive rate of 0.06, and precision of 0.94, demonstrate the effectiveness of this approach and suggest its potential application to other disease detection tasks. In Table 1, there are overview description of related works.

Table 1. Overview of related works in skin cancer detection

Vineeth et al. [27] studied a serious public health problem associated with the growing prevalence of common skin cancer around the world. The authors used deep learning techniques to create a model capable of predicting the disease in its early stages. A CNN model, which was designed by applying images of skin to analyze cancer-related attributes, brings desirable results in three hidden layers using the hybrid combination of activation function results in an accuracy of 95%. The system predicted with high accuracy for the test data of unseen data, concluding that the system can be used as an effective platform for the early diagnosis of skin cancer to facilitate newcomers and assist doctors, doctors and clinics.

Guergueb and Akhloufi [28] focused on melanoma, one of the most lethal types of skin cancer, capable of metastases if detected too late. According to the paper, the author sought to investigate the application of state-of-the-art Convolutional Neural Networks to detect melanoma, involving more than 36,000 images extracted from several datasets. The outcomes indicated that the top-performing model based on deep learning delivers extraordinary classification accuracy and a high Area under the Curve exceeding 99%, which illustrates the significant contribution of the model to the advancement of melanoma detection and early examination.

Kadampur and Al Riyaee [29] solved the problem of early detection of skin cancer, through the developed deep learning-based models to classify dermal cell images. The study uses a cloud-based model-driven architecture combined with advanced deep learning-based algorithms to improve the prediction performance. The models are evaluated using standard datasets, and an AUC close to 99.77% was achieved. Besides, InceptionV3 obtains the highest precision of 98.19% with an AUC of 99.23%. The results show that the proposed model-driven architecture can assist the practitioners in developing highly accurate skin cancer prediction models.

In the research methodology, we used Collab pro which is a cloud-based computing platform that helped us in doing our work more efficiently as well as effectively.

The high-performance computational environment, equipped with powerful GPUs, enabled the efficient training of machine learning models, overcoming the limitations commonly associated with standard hardware infrastructures [49]. Collab Pro enabled fast model training and validation due to the considerable amount of local configurations that were not needed [50]. In addition to this, the collaboration features enabled us to distribute code, share data and communicate with other members in our research team [51], creating a conducive work environment. All in all, Collab Pro helped us focus on conducting our SC detection experiment resulting into a seamless research experience. The current methodology is based on the application of cutting-edge convolutional neural network architectures: InceptionV3, NASNetMobile, VGG16, and Xception. These models are internationally recognized for their high efficacy in dealing with complex image data. The multi-model basis of the current study allows the integration of the most effective CNN architectures and ensures that the method is scalable and generalizable. For example, although Esteva et al. [2] used VGG models to achieve high classification accuracy of skin lesions, the current methodology applies more than a single model and enhances their performance by fine-tuning them to better process images of skin lesions. This approach not only improves the array of anomalies that can be detected via the models but also reduces the risk of overfitting, which is common in the single-model method. The methodology also operates a thorough pre-processing routine to ensure that the images fed to the neural networks are of high quality, increasing training and performance accuracy. In the study conducted by Moussa et al. [13], detection of melanoma relied on geometric features, making its analysis less informative than the current approach. The systematic use of advanced evaluations as accuracy, precision, and confusion matrices permits a detailed evaluation of each model’s advantages and alignment with practical use in clinical practice. Thus, not only does the current research follow the recent trends in medical imaging development, but it also presents a novel, versatile, and scalable framework that exceeds many established approaches in the field of skin cancer detection.

5.1 CNN model

We used strict statistical validation applications to assure the robustness and reliability of performance related to the CNN model utilized in skin cancer detection in this study. We used k-fold cross-validation to divide the dataset onto five distinct subsets to ensure that each subset is used as a testing set in one of the five folds, while the other four categories are used to train the model. This method eliminates bias and variance hence giving a reliable measure of how the model can perform across different subsets of the data used.

Our CNN model for skin cancer detection demonstrates exceptional diagnostic capabilities, achieving an overall accuracy of 97%. This high accuracy is crucial in clinical settings, where correct classification significantly impacts patient outcomes. Table 3 views the classification report for the database of CNN model. The model's sensitivity, or true positive rate, is 100%, ensuring nearly all malignant lesions are identified, which is vital for early detection and treatment of melanoma.

The specificity, at 94.44%, indicates the model's effectiveness in identifying benign lesions, although there is a slight inclination towards false positives. Balancing sensitivity and specificity are essential for clinical usability, aiming to reduce false positives without compromising the identification of malignant cases.

Table 3. Classification report of CNN model

Furthermore, the area under the Receiver Operating Characteristic curve of 0.996 is an indicator of a high discriminative ability. The ROC curve is a graphical representation of the diagnostic ability of a binary classifier system. Figure 4 shows that the model has high confidence in distinguishing between benign and malignant skin Furthermore, the area under the Receiver Operating Characteristic curve of 0.996 is an indicator of a high discriminative ability. The ROC curve is a graphical representation of the diagnostic ability of a binary classifier system.

504.png

Figure 4. ROC of CNN

Training and validation loss and accuracy curves provide deeper insights into the model’s learning dynamics. The steady decrease in training loss coupled with consistent improvements in training accuracy demonstrate the model’s effective learning from the training data. However, the observed fluctuations in the validation loss suggest moments of lesser generalization, which might be due to variations in data distribution or overfitting on the training data. These spikes highlight the importance of regularization techniques or potentially exploring more robust or diverse training data to enhance the model’s generalization capabilities (Figure 5).

505.png

506.png

Figure 5. Loss vs. accuracy

In summary, the CNN model is highly accurate and sensitive and specific and should be implemented in clinical settings for skin cancer detection. The ratio of sensitivity to specificity could be further improved, which could enhance the model’s reliability.

5.2 InceptionV3

Our InceptionV3 model in skin cancer detection has shown excellent diagnostic performance which is confirmed from the Table 4. The model achieved an accuracy of 93% which shows high correctness of classification of skin lesions. Its sensitivity was 92.67% which indicates that this model can effectively identify all the malignant lesions which is essential in early diagnosis. The model’s specificity was 93.33% which shows the models capacity to perform True Negative classification and then correctly identify the number of benign cases to help in the reduction of unnecessary intervention. The model’s precision for the malignant class was 92% and that of the benign was 94%. This indicates that there is a high level of correct positive predictions and a low false positive for the benign class. The model’s F1-score for the two classes was 0.93 which shows that the model can well balance the capacity to classify benign and malignant lesions.

Table 4. Classification report of InceptionV3 model

The area under the ROC curve was 0.9818, which indicates the model’s excellent discriminative ability. The high AUC value means that the model is extremely effective at distinguishing between benign and malignant cases. This is a crucial feature to assure the accuracy of diagnostic in clinical practice. Based on the principles of AUC, a value close to 1.0 indicates a high true positive rate with an extremely low false positive scale. Therefore, the model is reliable when it comes to accurately identifying potential skin cancer in the early stages. A_ROC curve visually confirms the model’s performance. When representing the true positive rate against the false positive rate at different threshold settings, the curve explains sensitivity and specificity’s trade-off. When the curve is closer to the upper left corner of the plot, it means that its performance is high. An AUC of 0.98 confirms the high model quality in terms of distinguishing benign and malignant cases (Figure 6).

6666.png

Figure 6. ROC of InceptionV3

The curve of training and validation loss decrease continuously indicating successful learning, and the accuracy curves show constant growth. Yet, the spikes of validation loss are indicators of overfitting, emphasizing the necessity of regularisation and various training data to foster the capacity of generalization. Therefore, the InceptionV3 model shows strong performance in skin cancer detection, with high accuracy, sensitivity, specificity, and great AUC (Figure 7). However, there is still a necessity for improvement aiming to make the balance between sensitivity and specificity more sophisticated to eliminate the number of false positives; thus, the model will become more reliable and practical in clinical practice.

5071.png

7777.png

Figure 7. Loss vs. accuracy

Table 5. Classification report of NasNetMobile

5.3 NASNetMobile model performance

A number of key performance metrics were used to evaluate the diagnostic capacity available in our NASNetMobile model for skin cancer detection The achieved overall accuracy of the model was 69%, reflecting the model overall moderate level of correctness in identifying/ classifying skin lesions.

Our NASNetMobile model performance at skin cancer diagnosis is analysed, through key metrics, and this will give us a clear-cut idea regarding the effectiveness of our model. The model had an overall accuracy of 69% in making a correct identification on whether the image contained a skin lesion or not. Sensitivity of the NASNetMobile algorithm, which equals true positive rate, was 38.33%. Humane Medicine Organization writes that a sensitivity of merely above 98% is acceptable. Consequently, a sensitivity of slightly more than 38% means that the model fails to recognize a significant percent of Malignant images. This is a critical issue because correct early detection is the goal of melanoma diagnosing in clinical practice. On the other hand, specificity of the algorithm was 95.27%. This value is acceptable because the lower the number of false positives, the fewer extra treatment patients go through, and the fewer states of confusion occur. However, it is twice lower than sensitivity; hence, the algorithm’s proclivity for false negatives could be concluded from these numbers. The Precision was 87% for the class malignant and 65% for the class benign. The precision of 87% seems pretty good, meaning that if the model recognizes a case as Malignant, it is correct 8 times out of 9. The F1-score calculated as 53% for malignant and 77% for benign concerning the balance between precision and sensitivity as shown in Table 5. A worse F1-score for malignant implies the model’s low capability of combining correct Malignant predictions and correct Malignant predictions.

The area under the ROC curve (AUC) indicates the model’s moderate discriminative ability, namely 0.8228 An AUC closer to 1.0 would be desirable, because that would reflect increasing ability of the model to discriminate between benign and malignant lesions; an AUC of 0.82 indicates there is still substantial room for improvement prev=regrTest (negative). These output/functions are presented in Figure 8 below: ROC curves are a very useful tool to represent with the help of graphic how well some classifier is capable to discriminate between classes at different threshold settings. Thus, considering the shape of curve and our AUC value being merely 0.82, the curves show that model has some discriminative capability but is far from optimal for clinical use.

508.png

Figure 8. ROC of NasNet

From the loss curves in Figure 9, we can observe that both training and validation set is increasing quite well but a little higher in case of validation set which means there’s clear sign of overfitting. If you look the training accuracy seems to getting better over time but validation accuracy gets worse which further highlights that we are having an overfitting problem. It appears that the model is only memorizing the training data and not generalizing which implies that there is a dire need of much better regularization techniques along with more diverse training data.

The NASNetMobile has a great potential in skin cancer detection with high specificity but suffers from lack of sensitivity and generalization properties. This is further confirmed by the AUC scores that are fairly moderate and there is a decent amount of difference in the Training and Validation performances. Future work on developing the model can concentrate on overcoming overfitting and improving sensitivity and specificity with a better balance between them to enable more reliable application of the model for clinical purposes.

5091.png

5092.png

Figure 9. Loss vs. accuracy

5.4 VGG16

The above mentioned indexes explain in assessment of the VGG16 model used for skin cancer detection, which demonstrate its capability to enable comprehensive diagnosis. The overall accuracy for the model was 77%, which means we have a decent classification in place. Sensitivity(recall) of model, i.e. True Positive rate was 89.67%. The high sensitivity of the model ensures that most malignant lesions are identified, which is very important in order to prompt early action and treatment. The definiteness, or true negative rate was 65.56%, which can be seen as consistent with the performed pathology; after all this AI has more difficulty in saying that a lesion is benign, resulting in higher rates of false positives, but not necessarily negatives the model had a precision of 68% for malignant class and 88% for benign class. This does indeed show that the model is relatively good if it makes a prediction of malignant, however not all those predictions are correct in Table 6. The F1-score, which can be seen as a compromise between precision and recall, was 0.78 in the case of malignant cases, and 0.75 in the case of benign cases These relatively matched F1-scores indicate moderate performance of the model in detection of both serious-prognosed and harmless lesions, but it is still far from ideal. There should be some measures to take, particularly increasing specificity in case of malignancy.

Table 6. Classification report of VGG16

As seen in Figure 10, the area under the ROC curve (AUC) was 0.8058 indicates moderators discriminative power of DT model. Mapping flow chart: Incomplete Block Symbol, an AUC value closer to 1 indicates improved performance in differently weed out between benign and malignant lesions. An AUC of 0.81 implies that the model does have some predictive power, but there is a lot of potential for improvement. ROC curve: In next ROC (Receiver Operating Characteristic) curve and it provide a graphical plot which illustrate the operational ability of a classifier model as its discrimination thresholds varied. The shape of the curve and AUC value of 0.81 indicates that the model shows some discriminative power but is still far from ideal for clinical use. Variability is also shown by the training and validation loss curves (Figure 11) as well, there are several points in which it jumps again due to increasing validation loss, indicating some level of overfit at those moments. While the training accuracy seems to increase over time (though slow), validation values fluctuates horribly, another strong sign of "we are indeed overfitting"! All of these fluctuations suggest the necessity for better regularization methods and a more varied training set to improve generalization.

Figure 10. ROC of VGG16

5111.png

5112.png

Figure 11. Loss vs. accuracy

The results of the study of the VGG16 model in skin cancer detection demonstrate the high sensitivity of the tool but are insufficient in relation to specificity and generalization. The described AUC, as well as the differences in performance between the training and validation models, are with the trend that justifies the need for its additional refinement. Thus, in the context of further research, it is proposed to work in the direction of mitigating the problem of overfitting and significantly increasing specificity, which will make it possible to achieve a better balance between sensitivity and specificity and thus develop a more reliable model for clinical use.

5.5 Xception

In this work various performance metric was examined, which gives the complete and overall idea of Xception model diagnostic result for skin cancer detection. As seen in Table 7, The model was able to achieve an overall accuracy of 95%, that means having good general accuracy in correctly identifying skin lesion. with the model having a sensitivity of 96.33% as seen above, it implies that the minority of malignant lesion will be missed making it to identify majority (most) of them and this is important for treatment also early intervention. The specificity (or true negative rate) was 93.89%, showing the level at which this model can propose that patients with benign lesions do not need to receive unnecessary treatments or feel as anxious concerning their health and well-being. The precision for the malignant class is 93%, and between the benign it goes to 97%. This follows that our model presents high correct positive rate — no matters how much false positives we have. The F1-score, which considers both precision and recall was 0.95 in malignant as well as benign cases. Both these balanced F1-scores indicate that the model is good at identifying both types of lesions — benign as well as malignant.

Table 7. Classification report of Xception

The area under the ROC curve (AUC) was 0.9886 (Figure 12), which also confirms excellent discriminative ability of our final model. An AUC number close to 1.0 suggests a model with high true positive rate and low false positive rate, which is the best diagnostic case you can get The ROC curve which is a plot of true positive rate vs false positive rate gives away an overall idea how well the model is able to distinguish between classes células in multiple threshold settings. With the curve shape together with AUC of 0.99, we can be sure that model is very effective in distinguishing benign from malignant lesions (as expected). Loss curves on both training and validation go down (Figure 13), indicating the learning process is going well. Accuracy curve goes up over Epochs constantly. Nonetheless, the sometimes spikes of validation loss indicate actual points in time when the model had begun to slightly overfit, which underscores the importance of regular monitoring and potential regularization methods to ensure generalization.

512.png

Figure 12. ROC of Xception

5131.png

5132.png

Figure 13. Loss vs. accuracy

5.6 Comparison between models

The performances of various models developed for skin cancer detection were assessed using several key metrics including accuracy, precision, recall, F1-score, sensitivity, specificity, and area under ROC curve. The CNN model performed as the best model with the highest achievable accuracy of 97%, perfect precision for class 0 and good recall for class 1 hence high/uptake of F1-scores and a near-perfect ROC curve with 0.9962 indicating perfect ability to correctly predict skin lesions with minimal false positive and negative cases (Table 8). The Xception model performed second best with an accuracy of 95%, high precision and recall achieved for the two classes and an area under ROC curve of 0.9800 indicating good discriminative ability. The InceptionV3 model performed relatively worse than the first two models albeit with a high accuracy of 93% and an area under ROC curve of 0.9819 from which it exhibits a perfect balance of sensitivity and specificity. This was the same case for NASNetMobile model that performed worst with an accuracy of 69%, and a very low recall for class 1 hence low f1-scores and area under ROC curve of 0.8228, indicating poor analytically and classification performance. The VGG16 model followed suit with an accuracy of 77% and area under the ROC curve of 0.8057, with the lowest sensitivity and relatively low f1-scores. Despite the model having an equal score for precision and recall the model did not perform as well as the other 3 models above. Generally, the CNN and Xception models performed better with equal accuracy, sensitivity, and specificity ratios with the other two models having areas needing improvement hence more sensitive as well as reducing false negative rates.

Table 8. Comparative table