[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Next Article in Journal
Exploring the Role of Symptom Diversity in Facial Basal Cell Carcinoma: Key Insights into Preoperative Quality of Life and Disease Progression
Previous Article in Journal
Prioritizing Context-Dependent Cancer Gene Signatures in Networks
Previous Article in Special Issue
Cutaneous Squamous Cell Carcinoma: An Updated Review
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Cancers
Volume 17
Issue 1
10.3390/cancers17010137
cancers-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

AI-Driven Enhancement of Skin Cancer Diagnosis: A Two-Stage Voting Ensemble Approach Using Dermoscopic Data

by
Tsu-Man Chiu
1,2,
Yun-Chang Li
2,†,
I-Chun Chi
3,† and
Ming-Hseng Tseng
3,4,*
1
School of Medicine, Chung Shan Medical University, Taichung 402, Taiwan
2
Department of Dermatology, Chung Shan Medical University Hospital, Taichung 402, Taiwan
3
Department of Biomedical Sciences and Engineering, National Central University, Taoyuan 320, Taiwan
4
Information Technology Office, Chung Shan Medical University Hospital, Taichung 402, Taiwan
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Cancers 2025, 17(1), 137; https://doi.org/10.3390/cancers17010137
Submission received: 24 November 2024 / Revised: 30 December 2024 / Accepted: 1 January 2025 / Published: 3 January 2025
(This article belongs to the Special Issue Recent Advances in Skin Cancers)
Browse Figures
Figure 1
<p>Research flow chart.</p> "> Figure 2
<p>Model architecture.</p> "> Figure 3
<p>Confusion matrix of the CSMUH dataset: (<b>a</b>) three-class training set; (<b>b</b>) three-class test set.</p> "> Figure 4
<p>Confusion matrix of the ISIC dataset: (<b>a</b>) three-class training set; (<b>b</b>) three-class test set.</p> "> Figure 5
<p>Confusion matrix of the CSMUH test set: (<b>a</b>) the three-class model; (<b>b</b>) the three-class model converted to binary classification; (<b>c</b>) binary classification after identifying benign cases in the three-class model; (<b>d</b>) binary classification of the two-stage model.</p> "> Figure 6
<p>Confusion matrix of the ISIC test set: (<b>a</b>) the three-class model; (<b>b</b>) the three-class model converted to binary classification; (<b>c</b>) binary classification after identifying benign cases in the three-class model; (<b>d</b>) binary classification of the two-stage model.</p> ">
Versions Notes

Simple Summary
This study utilized datasets from two ethnic groups to develop an AI diagnostic model. This model was trained using transfer learning, leveraging eight pre-trained models, including convolutional neural networks and vision transformers. The three-class AI model assists doctors in distinguishing between patients with melanoma who require urgent treatment, those with non-melanoma skin cancers who can be treated later, and benign cases that do not require intervention. The proposed two-stage classification strategy significantly improved diagnostic accuracy and reduced false negatives. This research demonstrates the success of the proposed method in both datasets. These findings highlight the potential of AI technology in skin cancer diagnosis, particularly in resource-limited medical settings, where it could become a valuable clinical tool to improve diagnostic accuracy, reduce skin cancer mortality, and decrease healthcare costs.
Abstract
Background: Skin cancer is the most common cancer worldwide, with melanoma being the deadliest type, though it accounts for less than 5% of cases. Traditional skin cancer detection methods are effective but are often costly and time-consuming. Recent advances in artificial intelligence have improved skin cancer diagnosis by helping dermatologists identify suspicious lesions. Methods: The study used datasets from two ethnic groups, sourced from the ISIC platform and CSMU Hospital, to develop an AI diagnostic model. Eight pre-trained models, including convolutional neural networks and vision transformers, were fine-tuned. The three best-performing models were combined into an ensemble model, which underwent multiple random experiments to ensure stability. To improve diagnostic accuracy and reduce false negatives, a two-stage classification strategy was employed: a three-class model for initial classification, followed by a binary model for secondary prediction of benign cases. Results: In the ISIC dataset, the false negative rate for malignant lesions was significantly reduced, and the number of malignant cases misclassified as benign dropped from 124 to 45. In the CSMUH dataset, false negatives for malignant cases were completely eliminated, reducing the number of misclassified malignant cases to zero, resulting in a notable improvement in diagnostic precision and a reduction in the false negative rate. Conclusions: Through the proposed method, the study demonstrated clear success in both datasets. First, a three-class AI model can assist doctors in distinguishing between melanoma patients who require urgent treatment, non-melanoma skin cancer patients who can be treated later, and benign cases that do not require intervention. Subsequently, a two-stage classification strategy effectively reduces false negatives in malignant lesions. These findings highlight the potential of AI technology in skin cancer diagnosis, particularly in resource-limited medical settings, where it could become a valuable clinical tool to improve diagnostic accuracy, reduce skin cancer mortality, and reduce healthcare costs.

1. Introduction

Skin cancer is widespread around the world and is one of the most common cancers in humans. On the basis of histopathological classification, it can mainly be divided into melanoma and non-melanoma skin cancers. Although melanoma has a lower incidence, it is the most lethal type of skin cancer. Non-melanoma skin cancers, which include basal cell carcinoma and squamous cell carcinoma (collectively known as keratinocyte carcinomas), have a higher incidence but relatively lower mortality rates. Melanoma is the deadliest form of skin cancer [1,2]. Due to the highly malignant nature of melanoma, early diagnosis is crucial to improve treatment outcomes. The prognosis and burden of skin cancer are highly dependent on the type of cancer and the stage at which it is diagnosed. Early diagnosis can greatly increase the chances of successful treatment and potentially reduce the need for surgery, radiation therapy, chemotherapy, targeted therapy, or immunotherapy, thus improving patient care and reducing healthcare care costs. The dermoscopic examination has become an essential tool in diagnosing melanoma and other pigmented skin lesions, with computer-aided skin lesion classification methods based on dermoscopic images receiving significant attention [3], especially in distinguishing melanoma from other benign skin lesions. However, due to the high visual similarity between different skin lesions, relying solely on visual observation and experience for diagnosis presents certain challenges and limitations. Beyond dermoscopic imaging, novel approaches, such as hyperspectral imaging (HSI), are gaining traction in skin cancer diagnostics. Lin et al. [4] proposed the spectrum-aided visual enhancer (SAVE), which converts RGB images into hyperspectral images and has demonstrated remarkable performance using the YOLO framework. With the advancement of artificial intelligence technologies, computer-aided diagnostic systems based on dermoscopic images have emerged as a potential solution. This study proposes an ensemble model, refined through various attempts, to improve the accuracy of predicting benign and malignant skin lesions, particularly melanoma.
To assess the effectiveness of AI models in dermoscopy images, numerous researchers rely on public datasets such as PH2, MED-NODE, and ISIC. This research examined 30 recent studies focused on computer-aided diagnosis (CAD) of melanoma for binary classification tasks, published from 2016 to 2024, and provided the minimum and maximum values for four performance evaluation metrics, as illustrated in Table 1.
The aim of this study is to enhance the accuracy of diagnosing malignant melanoma and other skin lesions from dermoscopic images by leveraging multiple pre-trained deep-learning models combined with an ensemble model strategy. Specifically, using a two-stage classification approach, this study seeks to reduce the false negative rate in the diagnostic process, which is of significant clinical importance. By conducting precise model training and validation, this research aims to provide clinicians with an effective auxiliary diagnostic tool that allows accurate identification of lesions that require immediate attention at an early stage.

2. Materials

2.1. CSMUH Dataset

This study used the CSMUH dataset, which was compiled by dermatologists at Chung Shan Medical University Hospital and comprises dermoscopic images of skin lesions collected from 2019 to 2023. Photos were obtained from individuals who had skin lesions excised in the dermatology department, with all lesions verified by pathologists at the Chung Shan Medical University Hospital. The inclusion criteria encompassed patients aged 18 years or older who consented to clinical imaging of skin lesions without regard to gender while excluding children and special groups. Exclusion criteria included photos that did not meet diagnostic classification, were unclear or indeterminate, or had dubious pathological diagnoses. This study used a Canon EOS 500D camera and a HEINE DELTA 20 dermatoscope. Photographs may or may not include the built-in scale of the dermatoscope. The lesions were cropped to the center of the image and must be clearly visible. Natural exposure settings were used, with no post-processing adjustments made to aspects such as color tone, sharpness, or saturation. Throughout the data collection phase, the dataset was classified into seven categories according to the ISIC framework, with subsequent categorization and case counts. There were 111 cases of melanoma, 199 cases of melanocytic nevus, 45 cases of basal cell carcinoma, 82 cases of actinic keratosis (and Bowen’s disease, keratoacanthoma, and squamous cell carcinoma), 134 cases of benign keratosis, 42 cases of dermatofibroma and 53 cases of vascular lesions, which is a total of 666 cases.
According to the objectives of this study, our objective was to identify key malignant melanomas by reclassifying the seven categories in the ISIC database into three distinct groups as follows: melanoma was classified as category 2, basal cell carcinoma and actinic keratosis (and Bowen’s disease, keratoacanthoma, and squamous cell carcinoma) were combined into category 1, and all other categories were grouped into category 0. In this classification system, a higher category number indicates a more severe condition. The machine learning model was then trained using this three-category classification. The number of instances in each category after reclassification is presented in Table 2.
Furthermore, to improve accuracy and reduce false negatives, we designed a two-stage voting process. In this approach, instances classified as category 0 by the first AI model were further screened by a second AI model, which was trained to reclassify the seven ISIC categories into two groups: required treatment (malignant and premalignant) and no treatment required (benign). Possible treatment-required cases were classified as category 1, and benign cases as category 0. The number of cases in each category after this reclassification is shown in Table 3.

2.2. ISIC Dataset

This study performed data integration based on the ISIC2018 Task 3 dataset [28]. Due to the problem of data imbalance in this dataset, additional data from ISIC2017 [29], ISIC2018 Task 1 [30], ISIC2019 [28,31], and ISIC2020 [32] to include the less represented skin disease categories. After removing duplicate images, these datasets were combined into a comprehensive ISIC 2017–2020 dataset for further research.
The number of instances for each category in the ISIC 2017–2020 dataset is as follows: 5008 cases of melanoma, 6705 cases of melanocytic nevus, 3261 cases of basal cell carcinoma, 1043 cases of actinic keratosis, 2657 cases of benign keratosis, 236 cases of dermatofibroma, and 253 cases of vascular lesions, totaling 19,163 cases. These categories were reclassified as follows: melanoma was classified as category 2, basal cell carcinoma and actinic keratosis were combined into category 1, and all other categories were combined into category 0. In this classification, a higher category number indicates a more severe condition. The number of cases in each category after reclassification is shown in Table 4. Furthermore, the categories were divided into benign and malignant groups, with malignant cases classified as category 1 and benign cases as category 0. The number of cases in each category after this reclassification is shown in Table 5.

3. Methods

3.1. Research Structure

In this study, the research flow chart is shown in Figure 1. Both the ISIC dataset and the CSMUH dataset were used for model training and testing, with 19,163 and 666 samples, respectively. The dermoscopic images were first preprocessed by resizing all images to 224 × 224 pixels and normalizing them to facilitate subsequent deep-learning model computations. The training set was augmented by randomly flipping, rotating, translating, and scaling the dermoscopic images. This increases the amount of data from a small number of categories and brings the training set closer to a balanced distribution of categories. After completion of image preprocessing, transfer learning was applied to integrate an ensemble model for training. The model was validated using five random experiments, with the primary evaluation metric being accuracy (ACC), calculated as shown in Equation (1). Here, TP represents true positives where the model correctly identifies positive samples, FN represents false negatives where the model incorrectly predicts positive samples as negative, FP represents false positives where the model incorrectly predicts negative samples as positive, and TN represents true negatives where the model correctly identifies negative samples.
ACC = (TP + TN)/(TP + TN + FP + FN)
To further improve the classification performance of the AI model, this study conducted separate model training for categories of three classes and two classes on the same dataset. Subsequently, images predicted as category 0 in the three-class test set were re-evaluated using the two-class model. This two-stage classification strategy effectively reduces the false-negative rate of the model.

3.2. Data Preprocessing

The dataset was divided into the training set and the test set, with all images resized to 224 × 224 pixels and normalized to facilitate computations in the deep learning model. To enhance the volume of image data during preprocessing, data augmentation techniques [33], such as random flipping, rotation, translation, and scaling, were applied to the training dataset.

3.3. Transfer Learning

Transfer learning involves applying a model originally trained in one domain to a new domain, transferring its learned knowledge to the target task without the need to train the model from scratch. The advantage of this approach is the reduced requirement for large amounts of training data. In this study, fine-tuning was performed on pre-trained models by fixing the weights of specific layers in the earlier parts of the model while training the later layers to obtain new weight values. Eight different pre-trained models were selected for this study, including EfficientNetB2, EfficientNetB3, EfficientNetB4, EfficientNetB5 [34], EfficientNetV2B2, EfficientNetV2B3 [35], ViT [36], and Swin [37].

3.4. Ensemble Learning

This study employed ensemble learning techniques to enhance the predictive performance of models by combining different models to achieve better overall results. Specifically, the voting ensemble method [38,39] was used, where the outputs of multiple independent models were combined as inputs for a new model, which was then trained as a new model. In this technique, multiple models are trained independently, and their predictions are combined through a voting process. Each model makes a prediction for each input sample, and the final output prediction is determined by a majority vote among all models. This approach is commonly used to take advantage of the strengths of multiple learning algorithms, improve predictive performance, and reduce the likelihood of errors from any single model.

3.5. Model Architecture

In this study, an ensemble deep learning model was constructed based on three pre-trained models using the voting method. The top three pre-trained models with the best training performance were selected, and a fully connected layer with a ReLU activation function and a Global Average Pooling (GAP) layer were added to each. To avoid overfitting, Batch Normalization (BN) and Dropout layers were incorporated. After obtaining the prediction results for the skin lesions of each model, a voting process was applied to generate the final prediction result. The model architecture is illustrated in Figure 2.

4. Results

4.1. Step 1: Three-Class Classification

4.1.1. CSMUH Dataset

This study explored the performance of various deep learning models in the classification of skin lesions, including the Swin Transformer, vision transformer (ViT), ResNet50, VGG16, and multiple versions of EfficientNet, and combined these models into an ensemble model. The optimal ensemble model consisted of the Swin Transformer, vision transformer (ViT), and EfficientNetB5, with its best performance displayed in the confusion matrix shown in Figure 3.
As shown in Table 6, the ensemble model achieved an accuracy of 99.77% during the training phase and 97.31% during the testing phase. This outperformed individual models, such as the Swin Transformer, which had a training accuracy of 98.46% and a testing accuracy of 89.85%; the vision transformer (ViT) model, which reached 99.47% accuracy during training and 90% during testing; the ResNet50 model, which had a training accuracy of 98.72% and a testing accuracy of 86.12%; the VGG16 model, which reached 75.71% accuracy during training and 68.36% during testing; and the EfficientNetB5 model, which achieved 99.36% accuracy in training and 89.7% in testing. The ensemble model demonstrated superior performance compared to these single models.

4.1.2. ISIC Dataset

In this study, the performance of various deep learning architectures was evaluated in the skin lesion classification task, including the Swin Transformer, the vision transformer (ViT), and multiple versions of EfficientNet. These models were combined into an ensemble model. The optimal ensemble model consisted of the Swin Transformer, EfficientNetB5, and EfficientNetV2B2, with its best performance illustrated by the confusion matrix shown in Figure 4.
As shown in Table 7, the ensemble model achieved an accuracy of 95.86% during the training phase and 85.38% during the testing phase. This outperformed individual models, such as the Swin Transformer, which had a training accuracy of 87.05% and a testing accuracy of 77.43%; the EfficientNetV2B2 model, which achieved 90.62% accuracy during training and 81.27% during testing; and the EfficientNetB5 model, which achieved 95.78% accuracy during training and 85.09% during testing. The ensemble model demonstrated superior performance compared to these single models.

4.2. Step 2: Two-Stage Strategy

4.2.1. CSMUH Dataset

This study used a two-stage model classification strategy to reduce the false negative rate and improve the diagnostic accuracy of skin lesions. As shown in Figure 5, the initial three-class classification results (Figure 5a) correctly identified 84 benign labels as benign; however, one malignant label was incorrectly classified as benign. In the second stage of testing, all cases initially classified as benign were re-evaluated using a binary classification to distinguish between benign and malignant. The results of this stage (Figure 5c) further confirmed the accuracy of the 84 benign cases and successfully corrected the previous misclassification, reducing the number of false negatives to zero. This validation method effectively minimized the occurrence of false negatives, not only in diagnostic accuracy but also in the applicability and reliability of the model in clinical practice.

4.2.2. ISIC Dataset

In this study, a two-stage model classification approach was employed to reduce the false negative rate and thus improve the accuracy of skin lesion diagnosis. As shown in Figure 6, the confusion matrix on the left shows the results of the three-class classification, where 1903 benign labels were correctly predicted as benign, but 124 malignant cases were misclassified as benign. In the second stage of testing, all images initially classified as benign were re-evaluated using a binary classification (benign versus malignant). The confusion matrix on the right shows that at this stage, 1898 benign labels were correctly identified, and the number of originally misclassified malignant cases was reduced from 124 to 45. This demonstrates that further validation can reduce the occurrence of false negatives. This method not only improved diagnostic accuracy but also enhanced the reliability of the model’s application in clinical practice.

4.2.3. Comparison of Performance Improvement Using the Two-Stage Strategy

To further illustrate the effectiveness of the two-stage classification strategy proposed in this study, Table 8 presents the values of five binary classification evaluation metrics: accuracy, Sensitivity, Specificity, false negative rate, and false positive rate. These metrics are evaluated for three AI classification models: the individual three-class model, the ensemble three-class model, and the two-stage model, all based on the CSMUH test set. Table 9 displays the performance comparison results for the ISIC test set. Whether considering the CSMUH dataset for the Eastern population or the ISIC dataset for the Western population, the experimental results in Table 8 and Table 9 clearly indicate that the two-stage model exhibits the best classification performance. It is followed by the ensemble three-class model in second place and the individual three-class model, which demonstrates the worst classification performance. Using the critical false negative rate (FNR) indicator in medical diagnosis as an example, in the CSMUH test set, the FNR indicator is 0.128 in the individual three-class model. In the ensemble three-class model, it is 0.021, and in the two-stage model, it drops to 0. In the ISIC test set, the FNR indicator is 0.151 in the individual three-class model, 0.021 in the ensemble three-class model, and drops to 0.024 in the two-stage model. It is worth noting that Table 6 and Table 7 show the average accuracy of five experiments, while Table 8 and Table 9 display the best accuracy from one of the five experiments.

5. Discussion

This study demonstrates the potential of ensemble models in dermoscopic images, particularly in significantly reducing the false negative rate in the diagnosis of malignant melanoma. The study utilized two datasets and trained eight different pre-trained model architectures, selecting the top three models to fine-tune and create a highly stable and reliable ensemble model. Compared to traditional single-model approaches, the ensemble model combines the strengths of multiple models, thus improving diagnostic accuracy and stability.
The study used a two-stage model classification strategy: First, a three-class classification was performed to identify benign cases, and then these benign cases were subtracted into a second-stage binary classification to differentiate between benign and malignant. In the CSMUH dataset, this two-stage classification strategy reduced the number of misclassified malignant cases to zero; in the ISIC dataset, it significantly reduced the number of malignant cases misclassified as benign from 124 to 45. Although some malignant lesions were initially misclassified as benign in the three-class classification, subsequent binary verification significantly reduced this number, demonstrating good performance. This approach not only improved diagnostic accuracy but also improved the model’s applicability and reliability in clinical practice.
Through this study, a highly stable and reliable ensemble model was established. Using a two-stage classification method, starting with a three-class classification followed by binary verification, the model performance was significantly improved. This approach achieved the best accuracy in predicting malignant melanoma from dermoscopic images and effectively reduced the false negative rate of malignant skin tumors. Clinically, this could provide substantial benefits compared to traditional diagnostic methods. The AI model developed in this study not only improved diagnostic accuracy but also significantly reduced the risk of misdiagnosis, particularly in controlling false negatives, which is crucial in clinical applications, as false negatives could lead to delayed treatment with severe consequences for patients. This study offers clinicians a tool to help identify lesions that require further examination and treatment at an early stage (malignant melanoma or malignant skin lesions), potentially reducing mortality rates and healthcare costs.
This research presents an effective method to accurately predict benign and malignant skin lesions, helping patients and dermatologists identify skin lesions early, thus reducing misdiagnosis rates and improving treatment efficiency.
However, there are some limitations to this study. For example, although the ensemble model showed good diagnostic performance, its computational demands are relatively high, which may limit its application in resource-constrained medical environments. Additionally, the CSMUH dataset used in this study is relatively small, which may affect the model’s generalizability across different populations. Future research should focus on expanding the dataset and incorporating more data from diverse sources to validate the model’s applicability across broader populations.
Table 10 provides a comparative summary of recent techniques for the binary classification of melanoma from 2021 to the present. It includes information on the authors, year, method, validation, dataset, classification, and overall accuracy of the test set. Establishing valid comparisons is challenging due to the use of different datasets, varying sizes, and various performance metrics across various studies. However, the two-stage voting ensemble approach proposed in this study demonstrates outstanding performance.
In general, this study suggests that the use of AI technology to aid in the diagnosis of skin lesions is a promising direction, particularly in the early diagnosis of melanoma and in the identification of benign and malignant skin lesions. The application of ensemble models helps improve diagnostic precision and reduce misdiagnosis rates, thus improving patient outcomes.

6. Conclusions

This study effectively improved the accuracy of diagnosing benign and malignant skin lesions in dermoscopic images using ensemble learning and a two-stage classification strategy, particularly achieving significant success in predicting malignant melanoma. First, the results demonstrated that the three-class AI model can assist doctors in distinguishing between melanoma patients who require urgent treatment, non-melanoma skin cancer patients who can be treated later, and benign cases that do not require intervention. Subsequently, the two-stage classification strategy effectively reduced the false negative rate during the diagnostic process. In particular, in the CSMUH dataset, the model achieved 100% accuracy in detecting malignant lesions, completely eliminating false negatives. This outcome not only underscores the potential of AI technology in the diagnosis of skin lesions but also provides clinicians with a reliable auxiliary diagnostic tool to help them more accurately identify lesions that require immediate attention at an early stage. This could potentially reduce skin cancer-related mortality rates and save healthcare costs.
Future research should focus on expanding the dataset and further optimizing the model, particularly improving its computational efficiency and adaptability in resource-limited environments. In general, this study provides strong evidence for the clinical value of AI-assisted skin lesion diagnosis, with significant potential for practical application.

7. Future Work

Future research should focus on improving the extraction of features and the annotation of dermoscopic images to further improve the diagnostic accuracy of AI models. Furthermore, the incorporation of clinical data such as age, sex, and location of the lesion into the models could improve diagnostic performance, particularly in primary care settings, where it would improve the precision of early diagnoses and reduce unnecessary procedures. Expanding the size of the dataset to include more pathologically confirmed skin lesion images would also help improve the generalizability of the model. Future studies should explore the cost-effectiveness of AI technology in the early detection of skin cancer and its acceptance in clinical practice, as these will be key factors in determining the widespread adoption of AI technology.

Author Contributions

M.-H.T. and T.-M.C. designed the study. M.-H.T. and I.-C.C. trained the models. I.-C.C. and Y.-C.L. collected the data. M.-H.T. and T.-M.C. analyzed and interpreted the data. All authors prepared the manuscript. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by grants from Chung Shan Medical University Hospital (CSH-2024-C-047).

Institutional Review Board Statement

This study was approved by the Institutional Review Board (IRB) of Chung Shan Medical University Hospital (CSMUH No: CS2-23084).

Informed Consent Statement

Patient consent was waived due to deidentified patients.

Data Availability Statement

The ISIC data that support the findings of this study are available in [ISIC Archive] at [https://www.isic-archive.com/ (accessed on 1 January 2024)], reference numbers [27,28,29,30,31]. The CSMUH data that support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Garbe, C.; Keim, U.; Gandini, S.; Amaral, T.; Katalinic, A.; Hollezcek, B.; Martus, P.; Flatz, L.; Leiter, U.; Whiteman, D. Epidemiology of Cutaneous Melanoma and Keratinocyte Cancer in White Populations 1943–2036. Eur. J. Cancer 2021, 152, 18–25. [Google Scholar] [CrossRef] [PubMed]
  2. Karimkhani, C.; Boyers, L.N.; Dellavalle, R.P.; Weinstock, M.A. It’s time for “keratinocyte carcinoma” to replace the term “nonmelanoma skin cancer”. J. Am. Acad. Dermatol. 2015, 72, 186–187. [Google Scholar] [CrossRef] [PubMed]
  3. Mahbod, A.; Schaefer, G.; Wang, C.; Dorffner, G.; Ecker, R.; Ellinger, I. Transfer learning using a multi-scale and multi-network ensemble for skin lesion classification. Comput. Methods Programs Biomed. 2020, 193, 105475. [Google Scholar] [CrossRef] [PubMed]
  4. Lin, T.-L.; Lu, C.-T.; Karmakar, R.; Nampalley, K.; Mukundan, A.; Hsiao, Y.-P.; Hsieh, S.-C.; Wang, H.-C. Assessing the efficacy of the spectrum-aided vision enhancer (SAVE) to detect acral lentiginous melanoma, melanoma in situ, nodular melanoma, and superficial spreading melanoma. Diagnostics 2024, 14, 1672. [Google Scholar] [CrossRef]
  5. Adjed, F.; Gardezi, S.J.S.; Ababsa, F.; Faye, I.; Dass, S.C. Fusion of structural and textural features for melanoma recognition. IET Comput. Vis. 2018, 12, 185–195. [Google Scholar] [CrossRef]
  6. Salido JAA, J.C. Using deep learning for melanoma detection in dermoscopy images. Int. J. Mach. Learn. Comput. 2018, 8, 61–68. [Google Scholar] [CrossRef]
  7. Warsi, F.; Khanam, R.; Kamya, S.; Suárez-Araujo, C.P. An efficient 3D color-texture feature and neural network technique for melanoma detection. Inform. Med. 2019, 17, 100176. [Google Scholar] [CrossRef]
  8. El-Khatib, H.; Popescu, D.; Ichim, L. Deep learning-based methods for automatic diagnosis of skin lesions. Sensors 2020, 20, 1753. [Google Scholar] [CrossRef]
  9. Al-Masni, M.A.; Kim, D.H.; Kim, T.S. Multiple skin lesions diagnostics via integrated deep convolutional networks for segmentation and classification. Comput. Methods Programs Biomed. 2020, 190, 105351. [Google Scholar] [CrossRef]
  10. Iqbal, I.; Younus, M.; Walayat, K.; Kakar, M.U.; Ma, J. Automated multi-class classification of skin lesions through deep convolutional neural network with dermoscopic images. Comput. Med. Imaging Graph. 2021, 88, 101843. [Google Scholar] [CrossRef]
  11. Li, Y.; Shen, L. Skin lesion analysis towards melanoma detection using deep learning network. Sensors 2018, 18, 556. [Google Scholar] [CrossRef] [PubMed]
  12. Almaraz-Damian, J.A.; Ponomaryov, V.; Sadovnychiy, S.; Castillejos-Fernandez, H. Melanoma and nevus skin lesion classification using handcraft and deep learning feature fusion via mutual information measures. Entropy 2020, 22, 484. [Google Scholar] [CrossRef] [PubMed]
  13. Bissoto, A.; Perez, F.; Ribeiro, V.; Fornaciali, M.; Avila, S.; Valle, E. Deep-learning ensembles for skin-lesion segmentation, analysis, classification: RECOD titans at ISIC challenge 2018. arXiv 2018, arXiv:1808.08480. [Google Scholar]
  14. Gessert, N.; Sentker, T.; Madesta, F.; Schmitz, R.; Kniep, H.; Baltruschat, I.; Werner, R.; Schlaefer, A. Skin lesion diagnosis using ensembles, unscaled multi-crop evaluation and loss weighting. arXiv 2018, arXiv:1808.01694. [Google Scholar]
  15. Gong, A.; Yao, X.; Lin, W. Classification for dermoscopy images using convolutional neural networks based on the ensemble of individual advantage and group decision. IEEE Access 2020, 8, 155337–155351. [Google Scholar] [CrossRef]
  16. Li, X.; Wu, J.; Jiang, H.; Chen, E.Z.; Dong, X.; Rong, R. Skin lesion classification via combining deep learning features and clinical criteria representations. bioRxiv 2018. [Google Scholar] [CrossRef]
  17. Lucius, M.; De All, J.; De All, J.A.; Belvisi, M.; Radizza, L.; Lanfranconi, M.; Lorenzatti, V.; Galmarini, C.M. Deep Neural Frameworks Improve the Accuracy of General Practitioners in the Classification of Pigmented Skin Lesions. Diagnostics 2020, 10, 969. [Google Scholar] [CrossRef]
  18. Wu, H.C.; Tu, Y.C.; Chen, P.H.; Tseng, M.H. An interpretable hierarchical semantic convolutional neural network to diagnose melanoma in skin lesions. Electron. Res. Arch. 2023, 31, 1822–1839. [Google Scholar] [CrossRef]
  19. Zhuang, J.; Li, W.; Manivannan, S.; Wang, R.; Zhang, J.J.G.; Pan, J.; Jiang, G.; Yin, Z. Skin lesion analysis towards melanoma detection using deep neural network ensemble. ISIC Chall. 2018, 2, 1–6. [Google Scholar]
  20. Adegun, A.; Viriri, S. Deep learning model for skin lesion segmentation: Fully convolutional network. In Image Analysis and Recognition, Proceedings of the 2019 International Conference on Image Analysis and Recognition, Waterloo, ON, Canada, 27–29 August 2019; Karray, F., Campilho, A., Yu, A., Eds.; Springer: Cham, Switzerland, 2019; pp. 232–242. [Google Scholar]
  21. Alfi, I.A.; Rahman, M.M.; Shorfuzzaman, M.; Nazir, A. A non-invasive interpretable diagnosis of melanoma skin cancer using deep learning and ensemble stacking of machine learning models. Diagnostics 2022, 12, 726. [Google Scholar] [CrossRef]
  22. Collenne, J.; Monnier, J.; Iguernaissi, R.; Nawaf, M.; Richard, M.A.; Grob, J.J.; Gaudy-Marqueste, C.; Dubuisson, S.; Merad, D. Fusion between an algorithm based on the characterization of melanocytic lesions’ asymmetry with an ensemble of convolutional neural networks for melanoma detection. J. Investig. Dermatol. 2024, 144, 1600–1607.e2. [Google Scholar] [CrossRef] [PubMed]
  23. Abbas, Q.; Celebi, M.E. DermoDeep-A classification of melanoma-nevus skin lesions using multi-feature fusion of visual features and deep neural network. Multimed. Tools Appl. 2019, 78, 23559–23580. [Google Scholar] [CrossRef]
  24. Abbes, W.; Sellami, D. Deep neural network for fuzzy automatic melanoma diagnosis. In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), Prague, Czech Republic, 25–27 February 2019; pp. 47–56. [Google Scholar]
  25. Chang, C.C.; Li, Y.Z.; Wu, H.C.; Tseng, M.H. Melanoma detection using XGB classifier combined with feature extraction and K-means SMOTE techniques. Diagnostics 2022, 12, 1747. [Google Scholar] [CrossRef] [PubMed]
  26. Nasr-Esfahani, E.; Samavi, S.; Karimi, N.; Soroushmehr, S.M.R.; Jafari, M.H.; Ward, K.; Najarian, K. Melanoma detection by analysis of clinical images using convolutional neural network. In Proceedings of the 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Orlando, FL, USA, 16–20 August 2016; IEEE: Orlando, FL, USA, 2016; pp. 1373–1376. [Google Scholar]
  27. Harangi, B. Skin lesion classification with ensembles of deep convolutional neural networks. J. Biomed. Inform. 2018, 86, 25–32. [Google Scholar] [CrossRef]
  28. Tschandl, P.; Rosendahl, C.; Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data 2018, 5, 180161. [Google Scholar] [CrossRef]
  29. Berseth, M. ISIC 2017—Skin lesion analysis towards melanoma detection. arXiv 2017, arXiv:1703.00523. [Google Scholar]
  30. Codella, N.C.F.; Gutman, D.; Celebi, M.E.; Helba, B.; Marchetti, M.A.; Dusza, S.W.; Kalloo, A.; Liopyris, K.; Mishra, N.; Kittler, H.; et al. Skin lesion analysis toward melanoma detection A challenge at the 2017 International symposium on biomedical imaging (ISBI), hosted by the international skin imaging collaboration (ISIC). In Proceedings of the 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), Washington, DC, USA, 4–7 April 2018; pp. 168–172. [Google Scholar]
  31. Combalia, M.; Codella, N.C.; Rotemberg, V.; Helba, B.; Vilaplana, V.; Reiter, O.; Carrera, C.; Barreiro, A.; Halpern, A.C.; Puig, S.; et al. BCN20000: Dermoscopic Lesions in the Wild. arXiv 2019, arXiv:1908.02288. [Google Scholar]
  32. Rotemberg, V.; Kurtansky, N.; Betz-Stablein, B.; Caffery, L.; Chousakos, E.; Codella, N.; Combalia, M.; Dusza, S.; Guitera, P.; Gutman, D.; et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Sci. Data 2021, 8, 34. [Google Scholar] [CrossRef]
  33. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet classification with deep convolutional neural networks. Commun. ACM 2017, 60, 84–90. [Google Scholar] [CrossRef]
  34. Mingxing Tan, Q.V.L. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. In Proceedings of the 36th International Conference on Machine Learning (ICML), Long Beach, CA, USA, 9–15 June 2019. [Google Scholar]
  35. Mingxing Tan, Q.V.L. EfficientNetV2: Smaller Models and Faster Training. In Proceedings of the 38th International Conference on Machine Learning (ICML), Virtual, 18–24 July 2021. [Google Scholar]
  36. Alexey Dosovitskiy, L.B.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; Uszkoreit, J.; et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations (ICLR), Vienna, Austria, 4 May 2021. [Google Scholar]
  37. Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; Guo, B. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, BC, Canada, 11–17 October 2021. [Google Scholar]
  38. Tseng, M.H. GA-based weighted ensemble learning for multi-label aerial image classification using convolutional neural networks and vision transformers. Mach. Learn. Sci. Technol. 2023, 4, 045045. [Google Scholar] [CrossRef]
  39. Zhou, Z.-H. Ensemble Methods: Foundations and Algorithms; CRC Press: Boca Raton, FL, USA, 2012. [Google Scholar]
  40. Raza, R.; Zulfiqar, F.; Tariq, S.; Anwar, G.B.; Sargano, A.B.; Habib, Z. Melanoma classification from dermoscopy images using ensemble of convolutional neural networks. Mathematics 2021, 10, 26. [Google Scholar] [CrossRef]
  41. Roshni Thanka, M.; Bijolin Edwin, E.; Ebenezer, V.; Martin Sagayam, K.; Jayakeshav Reddy, B.; Günerhan, H.; Emadifar, H. A hybrid approach for melanoma classification using ensemble machine learning techniques with deep transfer learning. Comput. Methods Programs Biomed. Update 2023, 3, 100103. [Google Scholar] [CrossRef]
  42. Azeem, M.; Kiani, K.; Mansouri, T.; Topping, N. SkinLesNet: Classification of Skin Lesions and Detection of Melanoma Cancer Using a Novel Multi-Layer Deep Convolutional Neural Network. Cancers 2023, 16, 108. [Google Scholar] [CrossRef] [PubMed]
  43. Qasim Gilani, S.; Syed, T.; Umair, M.; Marques, O. Skin Cancer Classification Using Deep Spiking Neural Network. J Digit Imaging 2023, 36, 1137–1147. [Google Scholar] [CrossRef]
  44. Hossain, M.M.; Hossain, M.M.; Arefin, M.B.; Akhtar, F.; Blake, J. Combining State-of-the-Art Pre-Trained Deep Learning Models: A Noble Approach for Skin Cancer Detection Using Max Voting Ensemble. Diagnostics 2023, 14, 89. [Google Scholar] [CrossRef]
  45. Thwin, S.M.; Park, H.-S. Skin Lesion Classification Using a Deep Ensemble Model. Appl. Sci. 2024, 14, 5599. [Google Scholar] [CrossRef]
  46. Faghihi, A.; Fathollahi, M.; Rajabi, R. Diagnosis of skin cancer using VGG16 and VGG19 based transfer learning models. Multimed. Tools Appl. 2024, 83, 57495–57510. [Google Scholar] [CrossRef]
Cancers 17 00137 g001
Figure 1. Research flow chart.
Figure 1. Research flow chart.
Cancers 17 00137 g001
Cancers 17 00137 g002
Figure 2. Model architecture.
Figure 2. Model architecture.
Cancers 17 00137 g002
Cancers 17 00137 g003
Figure 3. Confusion matrix of the CSMUH dataset: (a) three-class training set; (b) three-class test set.
Figure 3. Confusion matrix of the CSMUH dataset: (a) three-class training set; (b) three-class test set.
Cancers 17 00137 g003
Cancers 17 00137 g004
Figure 4. Confusion matrix of the ISIC dataset: (a) three-class training set; (b) three-class test set.
Figure 4. Confusion matrix of the ISIC dataset: (a) three-class training set; (b) three-class test set.
Cancers 17 00137 g004
Cancers 17 00137 g005
Figure 5. Confusion matrix of the CSMUH test set: (a) the three-class model; (b) the three-class model converted to binary classification; (c) binary classification after identifying benign cases in the three-class model; (d) binary classification of the two-stage model.
Figure 5. Confusion matrix of the CSMUH test set: (a) the three-class model; (b) the three-class model converted to binary classification; (c) binary classification after identifying benign cases in the three-class model; (d) binary classification of the two-stage model.
Cancers 17 00137 g005
Cancers 17 00137 g006
Figure 6. Confusion matrix of the ISIC test set: (a) the three-class model; (b) the three-class model converted to binary classification; (c) binary classification after identifying benign cases in the three-class model; (d) binary classification of the two-stage model.
Figure 6. Confusion matrix of the ISIC test set: (a) the three-class model; (b) the three-class model converted to binary classification; (c) binary classification after identifying benign cases in the three-class model; (d) binary classification of the two-stage model.
Cancers 17 00137 g006
Table 1. Results of the performance evaluation test on the models’ melanoma binary classification of melanoma.
Table 1. Results of the performance evaluation test on the models’ melanoma binary classification of melanoma.
AuthorsDatasetAUCAccuracySensitivitySpecificity
[5,6,7]PH2NA0.861~0.9750.790~0.9810.925~0.938
[8]Subset of PH2NA0.9500.9250.966
[9]ISIC 20160.7660.8180.8180.714
[9,10,11]ISIC 20170.870~0.9640.857~0.9330.490~0.9330.872~0.961
[10,12,13,14,15,16,17,18,19]ISIC 20180.847~0.9890.803~0.9380.484~0.8880.957~0.978
[20,21]Subset of ISIC 20180.9700.880~0.9100.920~0.960NA
[10,15]ISIC 20190.919~0.9910.896~0.9240.483~0.8960.976~0.977
[8,22]Subset of ISIC 20190.9420.870~0.9300.920~0.9250.820~0.933
[13,14,23,24]Combined0.880~0.9600.803~0.9500.851~0.9300.844~0.950
[25]Subset of combining
ISIC 2018 and ISIC 2019		0.9810.9650.8780.993
[26]MED-NODE0.810NA0.8100.800
[27]Subset of ISBI 20170.8910.8660.5560.785
Table 2. The number of cases in each category for the three-class classification of CSMUH.
Table 2. The number of cases in each category for the three-class classification of CSMUH.
ClassNumLabelImage
Melanoma (mel)1112Cancers 17 00137 i001
Basal cell carcinoma (bcc) and
actinic keratosis (ak) (and Bowen’s disease, keratoacanthoma, and squamous cell carcinoma)		1271Cancers 17 00137 i002
Melanocytic Nevus (nv), benign keratosis (bkl), dermatofibroma (df), and vascular (vasc)4280Cancers 17 00137 i003
Table 3. The number of cases in each category for the two-class classification of CSMUH.
Table 3. The number of cases in each category for the two-class classification of CSMUH.
ClassNumLabelImage
Melanoma (mel), basal cell carcinoma (bcc), and actinic keratosis (ak) (and Bowen’s disease, keratoacanthoma, and squamous cell carcinoma)2381Cancers 17 00137 i004
Melanocytic nevus (nv), benign keratosis (bkl), dermatofibroma (df), and vascular (vasc)4280Cancers 17 00137 i005
Table 4. The number of cases in each category for the three-class ISIC classification.
Table 4. The number of cases in each category for the three-class ISIC classification.
ClassNumLabelImage
Melanoma (mel)50082Cancers 17 00137 i006
Basal cell carcinoma (bcc) and
Actinic keratosis/intraepithelial carcionma (akiec)		43041Cancers 17 00137 i007
Melanocytic nevus (nv), benign keratosis (bkl), dermatofibroma (df), and vascular (vasc)98510Cancers 17 00137 i008
Table 5. The number of cases in each category for the two-class classification of ISIC.
Table 5. The number of cases in each category for the two-class classification of ISIC.
ClassNumLabelImage
Melanoma (mel), basal cell carcinoma (bcc), and actinic keratosis/intraepithelial carcinoma (akiec)93121Cancers 17 00137 i009
Melanocytic nevus (nv), benign keratosis (bkl), dermatofibroma (df), and vascular (vasc)98510Cancers 17 00137 i010
Table 6. Results of training and prediction of the three-class model for skin disease of CSMUH.
Table 6. Results of training and prediction of the three-class model for skin disease of CSMUH.
ModelTrain ACCTest ACC
Swin Transformer0.9846 ± 0.01080.8985 ± 0.0224
Vision Transformer0.9947 ± 0.00300.9000 ± 0.0304
EfficientNetB20.9805 ± 0.00400.8776 ± 0.0421
EfficientNetB30.9850 ± 0.01060.8776 ± 0.0121
EfficientNetB40.9914 ± 0.00680.8687 ± 0.0180
EfficientNetB50.9936 ± 0.00190.8970 ± 0.0256
EfficientNetV2B20.9820 ± 0.00890.8701 ± 0.0342
EfficientNetV2B30.9594 ± 0.03070.8463 ± 0.0516
ResNet500.9872 ± 0.00250.8612 ± 0.0312
VGG160.7571 ± 0.27630.6836 ± 0.2280
Ensemble 0.9977 ± 0.00220.9731 ± 0.0060
Table 7. Results of the training and test of the three-class model for ISIC skin disease.
Table 7. Results of the training and test of the three-class model for ISIC skin disease.
ModelTrain ACCTest ACC
Swin Transformer0.8705 ± 0.03920.7743 ± 0.0310
Vision Transformer0.7957 ± 0.28670.7103 ± 0.2442
EfficientNetB20.8329 ± 0.10790.7459 ± 0.0979
EfficientNetB30.6368 ± 0.23380.5845 ± 0.2033
EfficientNetB40.8563 ± 0.08100.7703 ± 0.0708
EfficientNetB50.9578 ± 0.00890.8509 ± 0.0072
EfficientNetV2B20.9062 ± 0.05290.8127 ± 0.0426
EfficientNetV2B30.8333 ± 0.28500.7539 ± 0.2452
Ensemble 0.9586 ± 0.00790.8538 ± 0.0095
Table 8. Comparison of performance improvement using the two-stage strategy at CSMUH test set.
Table 8. Comparison of performance improvement using the two-stage strategy at CSMUH test set.
CSMUH ModelAccuracy
(ACC)		Sensitivity SpecificityFalse
Negative Rate		False
Positive Rate
Three-classIndividual0.9400.8720.9770.1280.023
Ensemble0.9780.9790.9770.0210.023
Two-stageEnsemble0.9851.0000.9770.0000.023
Table 9. Comparison of performance improvement using the two-stage strategy at ISIC test set.
Table 9. Comparison of performance improvement using the two-stage strategy at ISIC test set.
ISIC ModelAccuracy
(ACC)		Sensitivity SpecificityFalse
Negative Rate		False
Positive Rate
Three-classIndividual0.8620.8490.8730.1510.127
Ensemble0.9000.9260.8800.0740.120
Two-stageEnsemble0.9690.9760.9630.0240.037
Table 10. A comparative summary of recent techniques for binary classification of melanoma.
Table 10. A comparative summary of recent techniques for binary classification of melanoma.
Author, YearMethodValidationDatasetClassTest ACC
Raza, R., et al. [40], 2021Ensemble with Xception,
Inception-ResNet-V2,
DenseNet121,
DenseNet201		Holdout (7:1:2)
full: 724		Dongsan Clinic in KeiMyung University Daegu, Korea20.979
Alfi IA, R.M., Shorfuzzaman M, Nazir A. [21], 2022Ensemble with MobileNet, Xception, ResNet50,
ResNet50V2, and DenseNet121		Holdout (8:2)
full: 3297		Subset of ISIC 201820.910
Chang C-C, L.Y.-Z., Wu H-C, Tseng M-H. [25], 2022InceptionResNetV2 + XGB + K-means SMOTEHoldout (8:2)
full: 2299		Subset of combining
ISIC 2018 and ISIC 2019		20.965
Wu H-C, T.Y.-C., Chen P-H, Tseng M-H. [18], 2023MEL-HSNetHoldout (9:1)
full: 4331		ISIC 201820.938
Roshni Thanka, M., et al. [41], 2023VGG16 + XGBoost 5-StratifiedKFold
train: 1000
test: 416		ISIC20.991
VGG16 + LightGBM20.972
Azeem, M., et al. [42], 2023SkinLesNet
VGG16
ResNet50		Holdout (8:2)
full: 1314		PAD-UFES-20-Modified dataset30.790
0.820
0.960
Qasim Gilani, S., et al. [43], 2023Spiking VGG-13Holdout (70:15:15)
full: 6993		ISIC 201920.896
Hossain, M.M., et al. [44], 2023Ensemble (max voting) with MobileNetV2, AlexNet, vgg16, ResNet50, DenseNet121, DenseNet201, InceptionV3, ResNet50V2, Inception, ResNetV2, XceptionHoldout
train: 2597
validation: 100
test: 1000		ISIC 201820.932
Thwin, S.M. and H.-S. Park [45], 2024Ensemble with VGG, ResNet-50, and Inception-V3.Holdout (75:25)
full: 995		ISIC30.910
Faghihi, A.F., M.; Rajabi, R. [46], 2024VGG1910-fold Cross Validation
full: 2541		ISIC20.987
Our approach, 2024Voting Ensemble with
Swin, EfficientNetB5, and EfficientNetV2B2		Holdout (8:2)
full:19,163		ISIC 2017~202030.900
Two-Stage Voting Ensemble with
Swin, EfficientNetB5, and EfficientNetV2B2		Holdout (8:2)
full:19,163		ISIC 2017~202020.969
Voting Ensemble with
Swin, ViT, and EfficientNetB5		Holdout (8:2)
full: 666		CSMUH30.978
Two-Stage Voting Ensemble with
Swin, ViT, and EfficientNetB5		Holdout (8:2)
full: 666		CSMUH20.985
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Chiu, T.-M.; Li, Y.-C.; Chi, I.-C.; Tseng, M.-H. AI-Driven Enhancement of Skin Cancer Diagnosis: A Two-Stage Voting Ensemble Approach Using Dermoscopic Data. Cancers 2025, 17, 137. https://doi.org/10.3390/cancers17010137

AMA Style

Chiu T-M, Li Y-C, Chi I-C, Tseng M-H. AI-Driven Enhancement of Skin Cancer Diagnosis: A Two-Stage Voting Ensemble Approach Using Dermoscopic Data. Cancers. 2025; 17(1):137. https://doi.org/10.3390/cancers17010137

Chicago/Turabian Style

Chiu, Tsu-Man, Yun-Chang Li, I-Chun Chi, and Ming-Hseng Tseng. 2025. "AI-Driven Enhancement of Skin Cancer Diagnosis: A Two-Stage Voting Ensemble Approach Using Dermoscopic Data" Cancers 17, no. 1: 137. https://doi.org/10.3390/cancers17010137

APA Style

Chiu, T.-M., Li, Y.-C., Chi, I.-C., & Tseng, M.-H. (2025). AI-Driven Enhancement of Skin Cancer Diagnosis: A Two-Stage Voting Ensemble Approach Using Dermoscopic Data. Cancers, 17(1), 137. https://doi.org/10.3390/cancers17010137

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop