Using The VAE GAN Structure as Pre-Training Technique to Improve Deep Learning Performance for Medical Image Classification

Background and aims: Artificial intelligence has made great progress in the field of medicalimage processing in recent years. One of the problems of the current deep learning models is theneed for a lot of labeled data for proper performance. On the other hand, there is little suitable andlabeled data in the medical field, and creating labels is time-consuming and expensive. We suggesta self-supervised pre-training method for Covid-۱۹ identification to overcome this challenge.Method: In the first stage of this study, a Variational Auto-Encoder Generative Adversarial Network(VAE-GAN) architecture is used. The images are first fed into an encoder network, and thenthe image data is fed into the latent space. The vector is then sent to the generator, which attemptsto reconstruct the same image using transposed convolution. The reconstructed image is then sentto the discriminator network, which attempts to determine whether the created image is real orfake. Following training, the initial layers of the encoder and discriminator, as well as the weights,are separated from the network and their output is fed into an MLP network, which is then trainedagain using the labels.Results: The results of the study show that our proposed model can perform better in detectingCovid-۱۹ with X-ray images than common models with limited labeled data.Conclusion: To address the problem of a lack of labeled datasets for pneumonia detection, weshowed that an unsupervised pretraining on unlabeled data can learn useful representations fromChest X-ray images and that only a few labeled data samples are needed to achieve the higher accuracyof a supervised model learned on a smaller annotated dataset. The proposed model can beapplied to other fields where there are insufficient labeled images to ensure proper performance.

کلیدواژه ها:

artificial intelligence ، medicine ، VAE-GAN ، self-supervised learning ، covid-۱۹