Skin cancer classification using novel fairness based federated learning algorithm

Problem statement

Federated learning is considered a very effective privacy-preserving and knowledge sharing problem solving approach; however, its performance is affected by the imbalanced and non-Independent and Identically Distributed (non-IID) data distribution across contributing clients, such as in real-world skin disease datasets. This distribution leads to bias in model performance, showing high performance for the majority classes and low performance for the minority classes in the datasets. This disparity is especially problematic in applications like skin disease detection, where underrepresented lesion types or skin tones may result in missed or incorrect diagnoses for vulnerable populations.

Research objectives

The major goal of this research was to develop and evaluate methods for addressing class imbalance in datasets to improve the performance of machine learning models on imbalanced datasets:

• To develop and implement a fairness-aware federated learning algorithm that dynamically adjusts client contributions.

• To compare the performance of the proposed fairness-aware federated learning algorithm with other algorithms using a real-world skin image dataset such as Fitzpatrick17k.

• To assess the proposed fairness-aware federated learning model using a comprehensive set of performance metrics—including accuracy, area under the receiver operating characteristic curve (AUC-ROC), precision, recall, and F1-score across multiple client distributions and skin disease categories.

• To test the significance of the proposed fairness-aware federated learning algorithm using a statistical significance test by considering all performance metrics.

Federated learning in medical imaging

The research work (Xu et al., 2022) addresses global dermatological challenges, advocating for early diagnosis through smartphone apps using federated learning for privacy-preserving model aggregation. Their fairness-aware framework improves diagnostic accuracy across diverse skin tones, validated on the Fitzpatrick 17k dataset. The article introduces FL as a solution for privacy-preserving machine learning, emphasizing its potential for collective intelligence while protecting data privacy. It highlights challenges, especially involving business competitors in FL federations, which can lead to delays and unequal benefit distribution. To address these issues, the authors propose the FLI payoff-sharing scheme, designed to maximize collective benefit and mitigate inequalities among data owners, supported by experimental comparisons with existing methods (Truex et al., 2019). The authors present FedCM, a federated learning framework for early detection and classification of Alzheimer’s disease (AD) using non-invasive data. FedCM employs model distillation to maintain model heterogeneity and prevent privacy breaches, focusing on sharing predictions rather than raw data or model weights. Tested on sMRI data from multiple datasets, FedCM outperforms previous FL and centralized learning systems in accuracy metrics and attention visualization on relevant brain regions, showcasing its viability in AD classification while addressing data privacy and distribution variability challenges (Huang, Yang & Lee, 2021). This study combines an upgraded version of the faster R-CNN with FL to propose a method for identifying multiple pests in orchards. FL allows data integration from various parties without centralizing data, reducing communication costs. Using ResNet-101 in the faster R-CNN improves detection accuracy for small targets, while multi-size fusion of feature maps enhances detection accuracy for pests and diseases of various sizes. The authors introduce the Soft-NMS algorithm to address shading issues, achieving 90.27% average accuracy and reducing detection time to 0.05 seconds per image. FL further boosts mean average precision (mAP) to 89.34% while decreasing model training time by 59%. Future work includes refining the FL algorithm, adding adaptive strategies to improve accuracy and convergence rates, and enhancing dense object detection (Deng et al., 2022). The research work (Wu et al., 2021) presents a federated contrastive learning (FCL) framework for dermatological disease diagnosis on mobile devices, integrating FL with contrastive learning (CL). FCL enables pre-training on distributed, unlabeled data followed by fine-tuning on limited labeled data without compromising privacy. By sharing characteristics during pre-training across devices, the approach enhances recall and precision for diagnosing dermatological diseases across varied skin tones. Experimental results show significant improvements over existing techniques, underscoring the effectiveness of FCL in enhancing diagnostic accuracy and privacy preservation. FedHealth (Chen et al., 2020) is a federated transfer learning system designed for wearable medical devices, combining data from different organizations while preserving privacy. It uses federated learning to create personalized models, achieving a 5.3% accuracy increase in human activity recognition over baseline methods. This system enhances wearable healthcare, especially for diagnosing conditions like Parkinson’s disease, by integrating large volumes of health data with advanced machine learning, addressing data fragmentation and model personalization issues. Federated transfer learning (FTL) (Saha & Ahmad, 2021) tackles data isolation and privacy in AI by allowing knowledge transfer across different user bases while maintaining privacy. FTL’s applications and future directions emphasize the need for advanced techniques and real-world datasets to improve effectiveness. Secure FTL advances efficiency and security in data federation by using secret sharing (SS) and secure multiparty computing (MPC) with the SPDZ protocol. This model enhances collaborative training efficiency and protects against malicious actors, reducing runtime and communication costs significantly (Sharma et al., 2019).

Federated learning for real-world problem-solving

The FOLB algorithm (Nguyen et al., 2020), designed for distributed mobile devices, achieves rapid convergence and improved accuracy by addressing the statistical heterogeneity of the system. Its adaptive aggregation effectively minimizes loss, reducing the number of communication rounds needed. Future research is suggested to refine device selection strategies to enhance performance. The research work by Liu et al. (2020) introduced a framework for privacy-preserving machine learning in decentralized data environments. It enables knowledge transfer among enterprises while protecting data privacy using techniques like homomorphic encryption and secret sharing. The framework facilitates accurate model generation without data exposure, outperforming traditional federated learning and achieving accuracy comparable to non-privacy-preserving methods. Future research aims to improve performance through distributed computing, enhancing the scalability of large data federations. FL is a promising solution for Internet of Things (IoT) networks, addressing distributed and privacy-sensitive data challenges (Khan et al., 2021). This approach, essential for applications in smart industries, intelligent transportation, and healthcare, leverages advances in FL, focusing on robustness, privacy, and communication efficiency. Integrating FL with emerging 6G networks could enhance IoT applications while ensuring privacy and security. Current FL systems are analyzed, detailing their functional design, distributed training methods, and data processing techniques. A four-layer architecture covers presentation, user services, FL training, and infrastructure. The study discusses central, hierarchical, and decentralized aggregation methods and data manipulation strategies like as compression and RPCs. FL systems such as TensorFlow Federated and FATE are reviewed, with research areas, including interpretability and handling unbalanced data highlighted for future exploration (Liu et al., 2022). Federated learning faces challenges with non-IID datasets, particularly in domains like medical imaging and object detection. The article introduces model-contrastive learning (MOON) to enhance FL performance on non-IID datasets through contrastive learning at the model level. MOON improves collaboration in multiparty training without raw data exchange, showing superiority over current methods in diverse image categorization tasks. Its potential extends beyond vision-related problems, addressing the heterogeneity of local data distribution effectively (Li, He & Song, 2021).

Fairness in federated learning

In the context of FL, the author explores fairness challenges beyond privacy and communication costs. They introduce AgnosticFair, a fairness-aware FL architecture that addresses unknown testing data distributions using kernel reweighing functions. This approach ensures fairness and high accuracy across diverse local datasets, demonstrating effectiveness on real-world data scenarios. Future research aims to expand this framework with additional fairness considerations and optimized kernel functions (Du et al., 2021). Auto-FedAvg addresses the dynamically adjusting weights based on data distribution and training progress. It demonstrates improved performance over existing FL methods in both general and medical image analysis tasks (Xia et al., 2021). A novel algorithm is introduced for fair resource allocation in federated learning, evaluating participants’ contributions to model performance without sharing data. It uses weighted accuracy improvements to allocate resources proportionally, surpassing traditional methods in fairness and scalability for large-scale applications (Li et al., 2019). Federated learning coordinates model training across decentralized clients to preserve data privacy, tackling challenges like statistical heterogeneity and limited communication bandwidth. FedFa, a proposed algorithm combining double momentum gradient and weighting techniques, enhances convergence and fairness in heterogeneous networks (Huang et al., 2020).

Machine learning deep learning models for medical image classification

Bello, Ng & Leung (2024) introduced a skin cancer diagnostic fine-tuned deep learning model in their research work. Their presented model was helpful to achieve better performance as compared to baseline models. Yaman et al. (2022) introduced a hybrid deep feature extraction model based on five pre-trained deep learning models and an ImRMR-based feature selection model for skin cancer classification. Baygin, Tuncer & Dogan (2022) presented a discrete wavelet transform, local phase quantization (LPQ), local binary pattern (LBP), and pre-trained DarkNet models for feature extraction learning-based skin cancer classification model. The study (Wang et al., 2021) tackles long-tailed image classification, where data imbalance makes learning challenging. The authors propose a hybrid network combining cross-entropy loss for classifiers and supervised contrastive loss for feature learning. The method transitions from feature to classifier learning over time. They explore supervised contrastive loss (SC loss) and prototype supervised contrastive loss (PSC loss), with PSC being more memory efficient. Experiments on three datasets show the hybrid network’s superior performance, making it the first study to apply supervised contrastive learning to long-tailed classification, with future work focusing on PSC loss optimization. Tuncer et al. (2024) introduced a fewer trainable parameters based lightweight CNN model in his research work. The authors (Tao et al., 2019) propose a novel self-adaptive support vector machine (SVM) cost-sensitive ensemble for imbalanced data classification, addressing challenges where the majority class overwhelms the minority samples. This approach enhances prediction accuracy by adjusting SVM decision boundaries to favor the minority classes, crucial for tasks like text categorization and intrusion detection. Their method, validated on real-world datasets, demonstrates robust performance and improves generalization by focusing on borderline the minority cases. Future directions include refining cost-sensitive strategies and optimizing boosting techniques to further enhance classifier performance in imbalanced datasets. The research work (Tao et al., 2020) proposes ACFSVM, an affinity and class probability-based fuzzy support vector machine, to handle imbalanced datasets. Traditional SVMs tend to favor majority classes, especially in noisy or outlier-rich datasets. ACFSVM uses an affinity measure computed in kernel space with an SVDD model on the majority class samples, coupled with class probabilities from kernel k-nearest neighbors. This approach identifies potential outliers and border samples in the majority class, emphasizing the significance of the minority class samples. Experimental results on UCI datasets show that ACFSVM outperforms existing methods in G-Mean, F-Measure and AUC, showcasing its effectiveness in imbalanced dataset classification. The study (Lee, Jun & Lee, 2017) addresses degradation issues in imbalanced data with their proposed method, which leverages SVM’s ability to construct nonlinear boundaries and detect minor disjuncts and data shifts. They enhance SVM’s performance in handling class imbalance by incorporating a weight adjustment factor into a weighted SVM used in AdaBoost. This factor focuses on borderline instances and positive noise, adjusting weights based on SVM margin categorization. Evaluation on 10 real-world datasets demonstrates the method’s superior performance over standard SVM and various sampling and boosting techniques in terms of F-Measure and AUC. Ensemble learning techniques for classifying imbalanced data using metrics like accuracy, F1-score, g-mean, minutiae cylinder-code (MCC), Cohen’s Kappa, and AUC, particularly in the context of cabbage image classification (Wardhani et al., 2019). Imbalanced data poses challenges in machine learning, especially in domains like medical and plant disease classification, where minority classes are underrepresented. Conventional metrics can be misleading, necessitating the use of metrics like AUC, MCC, and Kappa to accurately assess classifier performance. Ensemble learning proves effective in mitigating imbalanced data issues, ensuring robust classification performance across various metrics.

Materials

The experimentation and analysis, we have used an Intel Core i5 processor, 8 GB of RAM, and 500 GB computational resources for model training and evaluation. We have used the Windows-10 operating system, and the Google-Colab cloud service was used for the experimentation. All experiments, including model training and evaluation, were conducted on Google Colab using its cloud-based GPU environment for consistency and performance. The HP laptop was used solely for code development, preliminary testing, and result visualization. No model training or benchmarking was performed locally. For seamless execution, the Python environment in Colab was pre-configured with the necessary libraries such as TensorFlow, Keras, NumPy, Pandas, and Matplotlib. The data set used in this research work, is available at the link https://github.com/mattgroh/fitzpatrick17k (Groh, 2021). The Fitzpatrick17k dataset is used to evaluate the suggested approach. The Fitzpatrick17k dataset contains 16,577 images with six skin types, numbered 1 through 6, are labeled on the dataset, with smaller labels designating lighter skin and larger labels designating darker skin. Three separate categories make up the Fitzpatrick17k dataset: “Benign,” “Malignant,” and “Non-Neoplastic.” Fitzpatrick17k sample images that illustrate benign, malignant, and non-neoplastic skin conditions are shown in Fig. 1.

Methods

In federated learning, normally, data is divided into rounds. We have equally divided the training and testing datasets into 20 rounds. During the pre-processing, the dataset was divided into a training set (70%), a validation set (10%), and a test set (20%). The data used in the training, validation, and testing was also normalized. During the pre-processing, the dataset was loaded and then split the data based on Fitzpatrick scale values. To ensure consistency across images, the following pre-processing pipeline was applied. All images were resized to 224 $\times$ 224 pixels for compatibility with most CNN architectures. Pixel intensity values were normalized to the range [0, 1]. For the federated learning setup, a non-IID client partitioning strategy was applied to simulate realistic data heterogeneity across six virtual clients. The images available for the Fitzpatrick17k dataset were resized images and then we split these images to training, validation, and test sets as a part of pre-processing.

Automatic weight adjusting FL

The proposed algorithm is a federated learning approach designed to ensure fairness among participating clients while maintaining model performance. The list of parameters used in the algorithm is given in Table 1. The pseudocode of Algorithm 1 begins by initializing necessary parameters, including the client set C, the total number of communication rounds T, the number of local epochs E, the batch size B, the global model $\theta$, a hyperparameter Q that influences fairness weighting, and an upper bound M for the fairness adjustment counter m, which is initialized to 1. For each communication round, the global model $\theta$ is sent to all clients. Each client initializes its local model ${\theta }_c$ with the global model and performs local training using the LocalUpdate function. During this process, the client splits its dataset into mini-batches of size B, iterates through E local epochs, and updates the model parameters ${\theta }_k$ using gradient descent. The function also computes a fairness metric ${\lambda }_c$ based on the standard deviation of the class distribution within the client’s data. A smaller standard deviation, indicating more balanced data, results in a higher fairness contribution. After local updates, the server collects the updated models and fairness metrics from all clients. If $m$, the fairness adjustment counter, is below the upper bound M, it is incremented by 1. The server then calculates fairness-based weights $w_c$ for each client using an exponential weighting function, where the contribution of each client depends on $m-{\lambda }_c$. Clients with higher ${\lambda }_c$, signifying less fairness, contribute less to the global model update. Finally, the server aggregates the client models into the global model $\theta$ using the calculated weights. This algorithm is particularly suited for Federated Learning scenarios with heterogeneous and imbalanced data distributions. By incorporating a fairness metric into the aggregation process, it ensures equitable contributions from all clients, promoting balanced learning outcomes while preserving data privacy, as only models and fairness metrics are shared with the server. Class weighting is a key technique to address class imbalance in the model training, where minority class instances receive higher weights and majority class instances lower weights, balancing their influence. We refine this with a class imbalance ratio factor and an exponential parameter to fine-tune the weighting’s impact, adjusted using the standard deviation of class frequencies. This ensures precise calibration of weights according to dataset characteristics, enhancing the model’s ability to generalize and predict accurately across all classes by maintaining fairness and robustness in diverse applications.

Experimental setup

In this research work, we have used the Fitzpatrick17k dataset for the skin cancer image classification problem using a federated learning technique. The goal is to cooperatively train a DenseNet201 architecture across several clients while maintaining data privacy by utilizing the federated average technique. Google Colab is used for the implementation, utilizing its powerful computational power to run the code quickly. An HP laptop is used to facilitate the experimentation process, providing a handy platform for overseeing and administering the federated learning process. The federated learning process consists of multiple essential steps in each iteration. To simulate a distributed learning environment, the dataset is first divided among several clients. To avoid exchanging sensitive data with third parties, the clients train their models independently on local data for every training cycle. This decentralized method fosters a cooperative model improvement while allaying privacy concerns. The models are trained using client-specific data for several epochs throughout the training phase of each round. By applying methods like image augmentation and transfer learning from DenseNet201 models that have already been trained, the models are taught to identify images into groups like benign, malignant, and non-neoplastic. Metrics like accuracy and loss are tracked throughout training to evaluate the model’s convergence and performance. The federated averaging phase is essential for combining the knowledge from individual client models after local training. A consensus model that represents the combined intelligence of the participating clients is created by averaging the weights of the last layer across all clients. With the protection of privacy and secrecy, this federated averaging approach makes sure that the global model gains from the various data distributions found across various clients. The modified consensus model is then used as the basis for training on client data in the next rounds. The procedure is repeated several times, improving the model’s performance through cooperative learning at each iteration. Extensive experimentation and analysis yield insights into the federated learning approach’s scalability, performance improvements, and convergence behavior. All things considered, this study highlights the potential of federated learning in practical applications and its consequences for machine learning techniques that protect privacy. Client-wise partition of the specified dataset used in the experimentation is given in the Table 2. This non-IID partition simulates real-world data heterogeneity in federated learning environments for six federated clients.

Transfer learning models, architectures and hyperparameters

Transfer learning is a machine learning technique where a pre-trained model, developed on one task, is adapted for use on another related task. This method is particularly beneficial when there is limited data available for the target task, as it leverages the knowledge learned from a larger dataset. In transfer learning, the model’s architecture is typically retained while weights from the source domain are fine-tuned on a smaller target-domain dataset. This approach can significantly reduce training time and improve the model’s performance, especially in tasks such as image classification, object detection, and natural language processing.

Transfer learning is commonly used with deep learning models, where the base model (e.g., VGG16, ResNet50, DenseNet201) has already been trained on large datasets like ImageNet. The pre-trained model can be used either as a feature extractor or by fine-tuning some of its layers. This approach is now standard in machine learning, helping models generalize and handle limited labeled data.

Fairness in federated learning

Fairness in FL is a critical consideration as it involves training a model collaboratively across multiple decentralized devices or organizations. FL systems must ensure that the contributions from each participating entity are fairly accounted for, and no single party’s data disproportionately influences the model’s performance. This is particularly challenging when dealing with data heterogeneity, imbalanced class distributions, or unequal participation rates. Fairness metrics, such as equal opportunity and demographic parity, can be used to evaluate FL systems. Moreover, techniques like re-weighting loss functions, personalized FL models, and secure aggregation protocols help address fairness concerns, fostering equitable and unbiased outcomes across all participants. DenseNet201 architecture is used as a backbone architecture in the proposed algorithm because it is helpful to reduce overfitting and suitable to handle imbalanced client data. DenseNet201 pretrained model efficiently passes features and gradient flow between connected layers of the model. DenseNet201 is considered a robust model for the solution of medical imaging problems and is suitable for federated learning for heterogeneous clients with limited data.

The proposed algorithm incurs minimal communication overhead, as only a limited number of scalar parameters are exchanged per communication round. Moreover, the integration of fairness optimization results in a linear computational cost, which is negligible in comparison to the overall training workload and does not adversely impact runtime efficiency.

Precison

The precision comparison graph reported in Fig. 3 shows that the proposed model consistently achieves higher accuracy across benign, malignant, and non-neoplastic categories compared to AgnosticFL, DenseNet201, ResNet50, Vgg16, FedAvg and InceptionV3. This highlights the effectiveness of the proposed model in minimizing false positives, with notable precision in the non-neoplastic category, underscoring its robust performance in accurately identifying this group.

Recall

The recall comparison graph reported in Fig. 4 reveals that the proposed model consistently outperforms other deep learning models in the benign, malignant, and nonneoplastic categories, achieving the highest recall values across all categories. This indicates that the proposed model is more sensitive in recognizing true positives and effectively detects a greater percentage of cases in each category, especially excelling in the non-neoplastic category, where it significantly outperforms the other models.

F1-score

The F1-score comparison graph given in Fig. 5 shows the harmonic mean of precision and recall for various deep learning models across benign, malignant, and non-neoplastic categories, highlighting that the proposed model achieves the highest F1-scores in each category. Compared to AgnosticFL, DenseNet201, ResNet50, Vgg16, FedAvg and InceptionV3, the proposed model demonstrates superior overall performance by balancing precision and recall effectively. This is particularly evident in the nonneoplastic category, where achieving a high F1-score is crucial for a fair evaluation of the model’s capability in accurately identifying cases while minimizing both false positives and false negatives.

Accuracy

The accuracy comparison shown in Fig. 6 illustrates that across benign, malignant, and non-neoplastic categories, the proposed model consistently outperforms AgnosticFL DenseNet201, ResNet50, Vgg16, FedAvg, and InceptionV3. This underscores the modelś capability to accurately classify examples from all three classes, making it a strong candidate for the current classification task. The suggested approach demonstrates effectiveness in delivering reliable and precise classification results, evident from its significant accuracy improvement, particularly in distinguishing between different classes.

AUC-ROC

The AUC-ROC comparison reported in Fig. 7 illustrates that across benign, malignant, and non neoplastic categories, the proposed model consistently outperforms AgnosticFL DenseNet201, ResNet50, Vgg16, FedAvg and InceptionV3.

Confusion matrix of fairness-based federated learning algorithm

The confusion matrix given in Fig. 8 illustrates the class-wise performance of the proposed fairness-based federated learning algorithm for the skin lesion classification problem. It can be observed from Fig. 8 that fairness-based federated learning algorithms significantly improve the performance for the majority as well as minority classes. These results reflect enhanced generalization and reliability over existing algorithms.

Statistical analysis

In this research work, the following null and alternate hypothesis are defined to test the significance of the FbFedFAuto algorithm. The null hypothesis states that there is no significant difference between the performance of the proposed algorithm and other state-of-the-art models. The alternate hypothesis states that there is a significant difference in the performance. It is recommended to use a one-tailed t-test and small value of the level of significance when dealing with a small sample size of data in order to avoid Type-II error during the significance test. We have used both 5% and 10% levels of significance to perform statistical analysis for fair comparison. The following Table 5 shows a comparison of FbFedFAuto using p-values of one-tailed t-test along with test-statistic with other baseline algorithms.

It is evident from Table 5 that the performance of the proposed model is significantly different as compared to DenseNet201, Resnet50, VGG16, InceptionV3, AgnosticFL, and FedAvg across all performance metrics for 10% level of significance and significant performance in most of the cases for 5% level of significance. It can be observed that the performance of the proposed model is consistent across accuracy, precision, recall, F1-measure, and AUC-ROC curve.

Ablation study

This section presents the ablation study to assess and quantify the contribution of the proposed fairness-based federated learning algorithm for two widely used Fitzpatrick17k dermatology image datasets. The simple federated averaging learning algorithm, FedAvg algorithm and the proposed algorithm, FbFedFAuto are compared using the same parameter settings. In all three models, we have used DenseNet201, which is known as a strong feature propagation and strong gradient flow as a backbone architecture. The proposed algorithm is compared with the DenseNet201, ResNet50, VGG16, InceptionV3, FedAvg, and AgnosticFL algorithms for accuracy, precision, recall, F1-Measure, and AUC-ROC performance metrics. In this section, the average performance of all performance metrics is compared. It can be observed that, compared to DenseNet201 the proposed algorithm improves the average accuracy by about 4.72%, 4.72%, 4.51%, 3.66% for accuracy, precision, recall, and F1-measure, respectively. Similarly as compared to Resnet50 the metrics improves 4.72%, 4.93%, 5.54%, 5.30%; for VGG16 the metrics improves 8.13%, 7.91%, 7.23%, 5.42%; for Inceptionv3, the metrics improves 5.77%, 5.77%, 5.44%, 4.47% ; for FedAvg, the metrics improves 37.43%, 47.38%, 34.33%, 39.44% and for AgnosticFL, the metrics improves 3.95%, 3.98%, 3.66% and 4.27% respectively for average performance of three classes of benign, malignant and non-neoplastic. It can be summarized that the FbFedFAuto model consistently outperformed the baseline algorithms by improving the classification performance.

Research findings and advantages

This research work introduced a fairness-based federated learning algorithm for a skin cancer dataset. It is evident from the experimental results that incorporating fairness in the federated learning algorithm is helpful to achieve the highest accuracy and other metrics and outperform the DenseNet201, ResNet50, VGG16, InceptionV3, FedAvg, and AgnosticFl algorithms. Statistical tests showed consistent performance of the fairness-based federated learning algorithm compared to other algorithms for all performance metrics.

Potential applications and limitation

The ability to handle the diverse data of the proposed model makes it suitable for mobile health applications, telemedicine, and global diagnostics in remote areas to provide initial diagnostics so that people may consult for dermatology services in a timely manner. Although the proposed framework demonstrates strong performance, its generalization may be constrained by the limited diversity of evaluation datasets.

Future work

Future work will focus on validating the proposed model across additional dermatology datasets such as ISIC and DermNet to enhance generalizability.