Lung cancer is the leading cause of death by cancer in the world, thus demanding an urgent requirement for reliable and reproducible diagnostic tools for its early detection and differentiation into subtypes. Most classical deep learning methods suffer from catastrophic forgetting and have trouble with continuous learning, especially with an increasing amount of clinical data. This work investigates these issues by introducing a best multi-class lung cancer classification framework consisting of Residual Network architecture (Resnet-50) and Continuous Deep Learning (CDL), and Elastic Weight Consolidation (EWC). The proposed model was trained sequentially using two tasks: Task1 consisted of 6,800 histopathological images containing different lung cancer subtypes, and Task2 involved an additional set of 3,800 images with new subtype information. EWC was used to protect the important model properties acquired in Task1 and consequently mitigate catastrophic forgetting when learning Task2. Experimental results showed the model achieved a training accuracy of 99.7 % on Task1 and 96.4% on Task2. Importantly, it attained perfect test accuracy in testing for each task (test sets range from 1 to 1200 images), confirming the strength of its generalization, retention capability, and stability throughout sequential learning. CDL and EWC incorporated in the ResNet-50 architecture establish an effective pipeline for continued multicategory lung cancer classification. This research highlights a practical and gradual development path of AI-based diagnostic systems that interact with the knowledge embedded in the presented clinical data on the form of real healthcare situation, for more precise and adjustable decision-making logic.
Muhammed et al. (Tue,) studied this question.