Abstract:

The use of machine learning to medical imaging is being held down by the significant shortage of data. In this study, we analyze the impact of variety viable approaches to the problem of scarce data have on the pace at which training may be completed and the quality of the models produced. The first approach utilizes the more conventional kind of transfer learning. We show that transfer learning may have a considerable influence on the performance of a model when applied to the issue of cell instance segmentation. In particular, pre-training on generic cell datasets improves the performance of neuronal cell segmentation. The second technique (GAN-Based) is preliminary training that is based on the GAN. As part of this approach, a generative adversarial network, also known as a GAN, is constructed, and the discriminator is use as a pre-trained model for the first task. We show that GAN pre-training has a large beneficial influence on the classification of brain tumors using MRI images, even when no extra data is provided in the GAN pre-training stage.