FedMobileNet: A Resource-Conscious Federated Learning Framework for Imbalanced Lung Cancer CT Image Classification

Abstract

Federated learning enables distributed training without exchanging original images, but non-IID data and class imbalance can degrade model performance. In this study, FedMobileNet is presented as a decentralized machine learning system based on MobileNetV2 for multiclass classification. The training instances were spread among four simulated clients using a Dirichlet-based non-IID partitioning method (with α = 1.0) so that each client had at least five samples per class. The model used a multiclass focal loss with class weights set to α = [2.0, 1.0, 1.0, 1.0] and γ = 2.0; minority oversampling was not applied. The model that achieved the highest validation AUC at the checkpoint was selected and then evaluated once on the separate test dataset. In the test evaluation described here, FedMobileNet reached an accuracy of 89.4% and a macro-AUC of 0.97; the class-wise F1-scores were 0.81 for adenocarcinoma, 0.77 for large cell carcinoma, 0.97 for normal tissue, and 0.83 for squamous cell carcinoma. These results correspond directly to findings from a specific simulated federated configuration. They should not be interpreted as proof of clinical effectiveness or as evidence that the method is ready for deployment. The study uses a single public CT dataset, a single fixed random seed, and a specific client partition. It did not include analysis with repeated seeds and different partitions, external clinical validation, or direct measurements of memory, energy, communication cost, and inference latency. For this reason, the framework is described as resource-conscious rather than having been experimentally validated as resource-efficient for clinical or edge deployment.