Colonoscopy Image Synthesis: Transforming Medical Image Classification With Enhanced Quality and Computational Efficiency

This work introduces a novel approach to generate synthetic medical images from textual prompts using fine-tuned DreamBooth and LoRa. The study focuses on image synthesis and image classification, evaluating the quality and diversity of generated images using Inception Scores and Fréchet Inception Distance. The results show consistent average Inception Scores of 2.36 across datasets, with FID scores (0.11 for single-center, 0.073 for multicenter, and 0.076 for combined) indicating that the generated images closely resemble real ones. In the combined data set, EfficientNet shows superior performance, achieving 89% training accuracy and 91 % validation accuracy, significantly outperforming other models. In particular, all models exhibit slightly higher validation accuracies compared to their training accuracies, Indicating a good generalization. The substantial performance gap between EfficientNet and the other models, with a difference of at least 19 percentage points, underscores its exceptional effectiveness on this dataset. This research highlights the promise of AI-generated medical images, while emphasizing the need for ethical considerations and strategies to use them in healthcare applications.