PATL: Pool-Based Active Twin Learner from Oracle with Imitation Learning for Early Epidemic Detection

Early identification of epidemic (such as COVID19) from medical imaging (e.g., Chest X-rays (CXRs) and CT scans) is crucial for timely intervention and controlling the spread of infections. However, during the early stages of an epidemic, labeled data is scarce, and the manual labeling of medical images by experts is labor-intensive. To address these challenges, we propose a novel framework, Pool-based Active Twin Learner (PATL), which leverages Active Learning (AL) and Imitation Learning (IL) in a combined architecture. Inspired by the concept of a digital twin, we use a trainee twin model that learns from an oraclean expert or a higher-performing model-through imitation learning, while actively querying for additional labeled data when necessary. The framework uses K-means clustering for unsupervised learning, maximizing limited labeled data by identifying patterns. We demonstrate the effectiveness of this approach in detecting epidemic diseases from CXRs and CT scans using the VGG16 deep learning model. Experimental results demonstrate that our approach achieves $0.999,0.951$ accuracy on prediction of COVID19 using the chest CT scan dataset and CXR dataset respectively, and 0.980 accuracy on prediction of Pneumonia using CT scan dataset offering comparable performance to traditional methods in both labeling efficiency and classification accuracy. This highlights its potential as an effective solution for epidemic detection with limited labeled data.