CycleGAN-Based Transfer Learning for Efficient Pathogenic Bacteria Classification

The rapid and accurate identification of bacterial species is crucial in fields such as medicine, food safety, and agriculture. Traditional methods of bacterial strain identification are time-consuming, labor-intensive, and require expert knowledge. With the advancement of deep learning technologies, automated bacterial classification has emerged as a promising solution. This study proposes a transfer learning method based on Cycle Generative Adversarial Networks (CycleGAN) to improve the classification and recognition of common pathogenic bacteria. By leveraging macroscopic images to guide microscopic bacterial identification, the model facilitates efficient knowledge transfer between the macroscopic and microscopic domains. The proposed model includes two transformation functions, feature extractors, discriminators, and several loss functions to optimize the image translation process. We evaluate the model using a dataset from DIBaS, which includes images of Escherichia coli, Staphylococcus aureus, and Candida albicans under oil immersion conditions. The results show that the model achieves an average accuracy of 96.4%, outperforming several classic supervised models, including VGGNet, ResNet, and GoogLeNet, as well as some newer models. This result demonstrates the model's efficiency and accuracy, particularly in the recognition of microscopic images. Ablation studies confirm the importance of each component in the model, highlighting the role of recognition loss, macro auxiliary loss, and the cyclic transfer learning process. Despite the promising results, the method still requires a large amount of labeled data, which could be mitigated by introducing unsupervised learning techniques to reduce the dependence on labeled data, greatly improving efficiency. Additionally, the model's performance depends on the subjective selection of macroscopic objects, suggesting that future work should adopt more objective and accurate evaluation methods, such as using automated algorithms to assess similarity, or exploring strategies that combine unsupervised and self-supervised learning.