Federated learning for privacy-preserving skin cancer classification using deep neural networks

Skin cancer is a significant health issue in the entire world and there is a need to have diagnostic systems that are precise, enlargeable and privacy safeguarding. The heterogeneity of the institution and the issue of patient confidentiality frequently restrict the use of centralized deep learning in the medical imaging field because of data-sharing limitations and heterogeneity across different institutions. The solution is federation learning (FL), which allows joint training of models without transfer of unprocessed clinical data.

In this work, it is suggested to implement a privacy-sensitive FL architecture in skin cancer classification on the basis of decentralized clinical environments. The framework has compared two aggregation topologies, centralized Federated Averaging (FedAvg) and a decentralized ring-based peer-to-peer strategy on five simulated non-IID clinical client data distributions. In order to capture the institutional heterogeneity, we constructed a heterogeneity sensitive client partitioning algorithm that is measured by KullbackLeibler divergence. The suggested framework was also compared to standard FL algorithms, such as FedProx and SCAFFOLD, and was evaluated at various non-IID levels of severity. The backbone architectures implemented were MobileNetV2 and VGG16 and tested on two benchmark data sets, which were ISIC and Skin Cancer: Malignant vs. Benign. It was reported that the performance is measured in several random seeds and has 95% confidence intervals, and the balance of privacy and utility is measured using the settings of differential privacy.

MobileNetV2 with the ring-based FL topology was the best performing of all the considered configurations. This setting achieved an accuracy of 98.88% (95% CI: 97.92-99.84%) and F1-score of 98.80% which is significantly better than the centralized baseline (p < 0.01). In both datasets, MobileNetV2 has been able to achieve higher convergence rates, reduced communication overhead, and greater performance in terms of resistance to heterogeneous client distributions, compared to VGG16 in federation. The degradation of performance with increasing severity of non-IID was gradual, which validates that the planned framework is scalable and robust.

The results show that federated learning is able to be as effective as or even more effective than centralized learning in skin cancer classification without sacrificing patient privacy. The MobileNetV2 performance is higher, implying that lightweight architectures are specifically suited in FL due to their effective gradient propagation and low communication cost. Altogether, the suggested framework will be a feasible and regulation-conformed solution of AI-aided dermatological diagnosis in distributed healthcare settings.