ELMAGIC: Energy-Efficient Lean Model for Reliable Medical Image Generation and Classification Using Forward Forward Algorithm

This work presents a novel framework for energy-efficient medical data analysis and image synthesis by leveraging the Forward Forward algorithm with knowledge distillation to achieve significant reductions in memory and computational demands for medical data analysis. In the domain of image synthesis, autoencoders and diffusion models are utilized. The effectiveness of this approach is substantiated by the achieved F1 score of 87% in the classification of labels on the ODIRSK dataset. Moreover, the Frechet Inception Distance (FID) scores for the generated images are 0.06 for positive examples and 0.182 for negative ones, indicating a high degree of congruence between the generated and real images. The study further explores the concept of multi-teacher knowledge distillation. Here, the student model exhibits superior accuracy compared to Teacher2, albeit it does not surpass Teacher1. The research also applies iterative magnitude pruning in tandem with the Forward-Forward algorithm to optimize the state-of-the-art model, thereby augmenting its energy efficiency. Additionally, an attempt has been made to enhance the diffusion model by incorporating the Forward-Forward algorithm, aiming to reduce the denoising steps. The results are promising, with a slight trade-off in accuracy. In essence, this approach underscores the potential for energy-efficient image generation, augmentation, and classification while preserving high levels of accuracy and reliability. This study, therefore, serves as a testament to the advancements in the field of energy-efficient medical data analysis and image synthesis.