Adaptive hyperparameter and regularization strategies for improved elastic weight consolidation in lung cancer classification

Accurate and reliable categorization of lung cancer histopathology is considered challenging due to incremental data arrival, distributional shifts, and the tendency of conventional convolutional neural networks to suffer from overfitting and catastrophic forgetting when new subclasses are introduced. To address these issues, a continual learning framework was developed to preserve previously acquired knowledge while maintaining high diagnostic performance. A ResNet50 backbone regularized with Elastic Weight Consolidation (EWC) was employed, in which the Fisher-weighted penalty coefficient (λ) was updated dynamically through a curvature-aware schedule. By doing so, stability and plasticity were balanced adaptively without the need for manual grid search. The framework was trained and evaluated on 15,000 pathologically confirmed lung tissue images comprising adenocarcinoma (aca), squamous cell carcinoma (scc), and benign (n) classes. The experimental setup was organized as a continual learning sequence over three phases of training, where new classes were introduced incrementally while earlier ones were revisited. The proposed approach achieved external test accuracy of 100%, 97.2%, and 98.96% across the three phases, demonstrating its capability to sustain high accuracy while effectively preventing catastrophic forgetting. These results indicate that the integration of adaptive hyperparameter tuning with a curvature-driven EWC schedule can enhance continual learning performance in lung cancer classification. It is concluded that the ResNet50-EWC framework provides a scalable and clinically relevant solution that eliminates the need for full retraining when data expands, and future validation on additional histological subtypes and external whole-slide cohorts is planned.