Conventional centralised machine learning frameworks used for early prediction of diseases frequently prove ineffective to address the privacy concerns of patients’ medical data and health records, as they necessitate the aggregation of raw data at a single site, hence increasing the danger of data breaches and regulatory non-compliance. This study presents a unified framework that advances algorithmic innovation and domain-specific applications through the consolidation and critical analysis of recent research on Federated Learning (FL) as a secure, decentralised healthcare intelligence solution.
What is Federated Learning in Healthcare?
Federated Learning (FL) trains machine learning models collaboratively across several healthcare facilities or edge devices while preserving sensitive information by aggregating model parameters or gradients on a global scale without direct data exchange. However, traditional FL techniques like “FedAvg, FedProx, FedAdam, FedYogi, and FedAdagrad” has limitations to perform under non uniform distribution of data across the sites, have high communication costs, and lack in mechanisms for equitable client selection when dealing with heterogeneous non-IID data distributions.
Limitations of Traditional Centralised Machine Learning
Such constraints limit convergence rates, scaling of models and applicability of models in the real healthcare environment. In order to eliminate these bottlenecks, this research combines three new contributions of the reviewed journal articles:
1) Performance analysis of decentralised aggregation technique for breast cancer diagnostics
2) CFWL (Custom Federated Weighted Learning) model of melanoma detection.
3) Priority-based Client Selection Federated Averaging (PC_FedAvg) model of brain tumor classification along with Selection of transfer learning models.
All of them presuppose deep transfer learning models including CNN, VGG16, VGG19, DenseNet201, and Xception within a FL setting to enhance feature presentation, speed up the convergence process, and decrease the communication cost. Some of the top computer engineering colleges offer specialisation programs in machine learning to train future engineers to take up the challenge of decentralising machine learning through classification algorithms.
In breast cancer diagnosis, it is assured that adoption to a decentralised framework is an appropriate choice and it validates the pros and cons of FedAvg and FedProx aggregation techniques. The CFWL model employs adaptive client weighting (by combined FedAvg and FedProx strategy) to prioritise high-performing nodes, significantly improving classification accuracy and resilience in melanoma prediction under non-IID conditions. To prioritise high-performing nodes, significantly improving classification accuracy and resilience in melanoma prediction under non-IID conditions. Experimental evaluations demonstrate that integrating VGG16 into the federated setup yields 90% accuracy over CNN under non-IID data settings, superior precision, recall, and F1-scores compared to conventional FL baselines with respect to aggregation algorithms FedAvg, FedProx, FedAdagrad, while simultaneously minimising privacy leakage risks.
For brain tumor classification, the PC_FedAvg algorithm incorporates a priority-driven client selection mechanism that restricts participation to top-performing clients, thereby accelerating convergence and reducing communication rounds without compromising accuracy. When benchmarked on the Figshare, SARTAJ, and Br35H datasets, PC_FedAvg combined with the Xception model achieved 99.4% accuracy. When combined with PC_FedAvg strategy along with federated set up it achieves 91.8 % accuracy and substantial reductions in communication cost compared to FedAvg, FedAdam, and FedProx.
Methodology: Transfer Learning for image classificationDenseNet201, VGG16, CNN,Xception model.
Datasets: Bain MRI and ISIC2018 image dataset
Size of datasets:
BrainMRI: 2GB (6695 images)
ISIC:10.4GB (10,015 images)
Aggregation Techniques: FedAvg, FedProx and proposed Hybrid
Platform/Framework: Flower and Tender Flow Federated (TFF) with Colab
Scope and Future Applications of Federated Learning in Healthcare
This research focuses primarily on developing and evaluating high-accuracy, communication-efficient, and privacy-preserving FL frameworks to be used in different health applications. The project will solve the underlying challenges of data heterogeneity, regulatory requirements, communication overhead, and fairness problems that currently do not allow FL to find widespread application in the healthcare sector.The end goal is to create federated models that will deliver reliable diagnostic and prognostic capabilities in numerous medical fields as well as ensure data privacy and deliver clinical-quality performance. This project aims to provide the foundations of the safe, scalable, and ethical use of federated learning in healthcare decision-support systems.
Conclusion
A significant contribution of the work was the design and validation of customised federated learning (Custom Federated Weighted Learning (CFWL) to detect melanoma, PC_Federated averaging (PC_FedAvg) to classify brain tumors and decentralised aggregation strategies to identify breast cancer. All these frameworks were compared to already state-of-the-art algorithms, including FedAvg, FedProx, and FedAdam. According to the findings, the proposed methods were not only competitive and more accurate but also used fewer communication rounds, thereby overcoming the problem of computational and communication overheads, which is especially important in large-scale healthcare systems.
Pursuing a degree program like B.Tech in Artificial Intelligence and Machine Learning can help you better understand this topic. The experimental results also verified that the more sophisticated deep learning models such as Xception, VGG19, and DenseNet201 when deployed in the federated settings could attain a high level of generalisation performance on several datasets. Notably, the thesis emphasised that the diagnostic reliability was not to be judged by the accuracy only. Measures like precision, recall, F1-score, and communication effectiveness were thoroughly taken into account to give a comprehensive analysis.