Lightweight Continual Learning for Cervical Cancer Diagnosis with Synthetic Augmentation and Explainable AI

Sequential task learning in medical diagnosis often leads to catastrophic forgetting, where a model’s accuracy on earlier disease patterns sharply declines after learning new data (Rios & Itti, 2018). Recent advances use generative models to replay past data and mitigate forgetting (Rios & Itti, 2018). We propose the Lightweight Continual Cervical Diagnosis GAN (LCC-DGAN), a novel framework that combines closed-loop generative replay with risk factor interpretability for cervical cancer screening. Our approach trains a compact generative adversarial network (GAN) to synthesize diverse patient risk profiles from earlier training stages, stored in a small memory buffer. This synthetic augmentation is interwoven with incremental learning of a lightweight classifier, preventing forgetting of previously learned risk patterns while adapting to new data. First, we show that preserving a diverse subset of prior patient samples in a memory unit allows the model to nearly match the performance of training on all data at once (Rios & Itti, 2018). Next, we demonstrate that continual refresh of the GAN-generated samples during training yields further performance gains, sustaining accuracy on earlier cohorts without large memory overhead (Rios & Itti, 2018). We compare LCC-DGAN against fine-tuning and regularization-based approaches (e.g. Elastic Weight Consolidation), observing superior retention of diagnostic accuracy on prior cases. Finally, to ensure clinical trust, we integrate explainable AI modules: Grad-CAM visualizations highlight salient regions in any imaging data (Selvaraju et al., 2017), and SHAP values quantify each risk factor’s influence on model predictions (Shakil, Islam & Akter, 2024). This closed-loop system provides not only robust continual learning for cervical cancer diagnosis, but also interpretable insights into how factors like age, sexual history, and HPV status contribute to each decision (Shakil, Islam & Akter, 2024). The result is a computationally efficient, deployable diagnostic tool that continually improves with new data while retaining knowledge and providing clinicians with understandable explanations of its predictions.