Medical Image Segmentation: A Review of Modern Architectures

Medical image segmentation involves identifying regions of interest in medical images. In modern times, there is a great need to develop robust computer vision algorithms to perform this task in order to reduce the time and cost of diagnosis and thus to aid quicker prevention and treatment of a variety of diseases. The approaches presented so far, mainly follow the U-type architecture proposed along with the UNet model, they implement encoder-decoder type architectures with fully convolutional networks, and also transformer architectures, exploiting both attention mechanisms and residual learning, and emphasizing information gathering at different resolution scales. Many of these architectural variants achieve significant improvements in quantitative and qualitative results in comparison to the pioneer UNet, while some fail to outperform it. In this work, 11 models designed for medical image segmentation, as well as other types of segmentation, are trained, tested and evaluated on specific evaluation metrics, on four publicly available datasets related to gastric polyps and cell nuclei, which are first augmented to increase their size in an attempt to address the problem of the lack of a large amount of medical data. In addition, their generalizability and the effect of data augmentation on the scores of the experiments are also examined. Finally, conclusions on the performance of the models are provided and future extensions that can improve their performance in the task of medical image segmentation are discussed.