Bayesian-Optimised Latent Encoding and Agent-Based Simulation for Enhanced Medical Image Character Recognition

This paper presents a Bayesian-optimised Conditional Variational Autoencoder (CVAE) for synthetic data augmentation, embedded within an agent-based simulation framework. The CVAE systematically refines latent-space representations, generating high-quality synthetic character images that enhance dataset diversity and reduce the risk of overfitting. Bayesian optimisation ensures optimal latent variable selection, improving reconstruction accuracy while enabling scalable Medical Image Character Recognition (MICR) training. The proposed agent-based system introduces autonomous agents: patient agents, doctor agents, imaging device agents, and recognition agents that collaborate to simulate real-world MICR workflows. This structured pipeline enables dynamic dataset augmentation while supporting medical diagnostics and automated text extraction. Experimental evaluations demonstrate significant performance improvements, with CNN models achieving accuracy gains of +3.2%, +3.5%, and +1.79% on the public dataset and +2.41%, +6.85%, and +1.60% on the private dataset when augmented with 50, 100, and 150 synthetic images per class, respectively. This research validates the effectiveness of Bayesian-tuned latent-space encoding and a supporting agent-based data augmentation, offering a scalable, computationally efficient solution for MICR enhancement.