CNN-EWC: A continuous deep learning approach for lung cancer classification

Abstract Problem In continuous deep learning scenarios, where the model is required to learn new tasks without losing knowledge from prior tasks, catastrophic forgetting is a significant limitation of continuous deep learning models. This problem has the highest significance in the field of medical image analysis, as it is imperative to ensure the reliability of diagnostics and prediction through consistent and accurate classification. This study solves the challenge of catastrophic amnesia in deep learning by utilizing the conjunction of Convolutional Neural Networks (CNNs) with the Elastic Weight Consolidation (EWC) called CNN-EWC model. Aims This research seeks to develop a continuous learning pipeline for classifying histopathology images of lung cancer into lung benign tissue ( n ), lung adenocarcinoma (ACA), and lung squamous cell carcinoma (SCC). A dataset of 15,000 images equally across these three classes is used to comprehensively evaluate the proposed method in a realistic medical diagnostic scenario. Methods The suggested CNN-EWC method mitigates catastrophic forgetting in continuous deep learning by safeguarding critical parameters across tasks. The model is trained successively on three tasks, utilizing EWC to inhibit substantial alterations to essential parameters. This ensures a preservation of knowledge from previous tasks while accommodating new tasks. The methodology preserves elevated classification precision in medical diagnostics by balancing stability and adaptability. Results The CNN-EWC model experimental results achieved testing accuracy from 98 to 99.6%. Notwithstanding a diversity in dataset size (3,000, 4,500, and 6,000 images), the model performed excellently for lung n, lung ACA, and SCC, its good statistics hold true for all three diseases tested on the testing dataset. External tests were conducted on (3–1,500) images to verify the accuracy of the model’s classification and prediction, and the accuracy reached 100%, which indicates that the problem of catastrophic forgetting that occurs in continuing education has been addressed. Conclusions The CNN-EWC model has proven its effectiveness in continuous training in terms of accuracy, speed, saving time, and correct classification of images as it solves catastrophic forgetting but also maintains very high classification performance. In medical imaging, the simple switching logic effectively solves several simulations of catastrophic forgetting by avoiding back propagation for large numbers of images or by direct calculation.