An Efficient Method for Lung Cancer Classification Using EfficientNet and Learner Cooking Optimization

Abstract

Computational approaches are increasingly important for accurate and privacy-preserving analysis of medical imaging data. This study proposes a federated deep learning framework that integrates optimized deep feature extraction, feature selection, and privacy-preserving collaborative learning for multiclass lung cancer classification from CT images. The proposed method uses convolutional neural network with compound scaling by performing hierarchical feature extraction and then feature selection strategy to minimize redundancy and improve the class separability is based on optimization. The training of the model occurs in a federated (between several clinical nodes) distributed environment in which a global classifier is jointly trained (learned) without any actual exchange of patient images, a factor that guarantees the data privacy. Experiments with the publicly available LIDC-IDRI prove that the proposed framework has a total accuracy of 99.73, precision of 99.62, sensitivity of 99.82, specificity of 99.43, and F1-score of 99.38 with four lung cancer types. The architecture of the proposed federated feature aggregation framework is not specific to lung CT imaging, the pattern of combining a pretrained deep feature extractor, a bio-inspired feature optimizer, and secure additive masking is structurally applicable to other privacy-sensitive medical imaging tasks, though validation in those domains would require independent experimental evidence and is identified as future work.