Experiments with Semi-Supervised Learning: from Cityscapes to Medical Images

Deep neural networks have proved exceptional performance in supervised learning tasks, particularly in image segmentation and classification, when trained on large sets of labeled data, like ImageNet. However, creating such extensive datasets requires significant effort, which may not be feasible in many real-world settings. This limitation slows down the process of adopting a solution based on deep learning in various contexts. To address this challenge and develop more data-efficient deep learning methods, there is an increasing interest in semi-supervised learning, which aims to reduce the amount of labeled data required. In this paper, we discuss different methods of data augmentation to be used in semi-supervised learning and report on experiments with Generative Adversarial Networks (GANs) and Domain Adaptation. We apply a supervised method on two datasets from two different domains, to obtain a better perspective on the performance of the approach, and then test how we can improve it using fewer real data through procedures such as synthetic data generation, augmentation, and domain adaptation.

I. Introduction

Semi-supervised learning combines labeled and unlabeled data during training to improve model performance. It can be important for applications where the labeled data is inexistent or expensive to obtain. When doing semantic segmentation on medical images, semi-supervised learning can help with the fact that is hard to obtain so much sensitive data from real patients. Using this method, we can reduce human effort and time required for annotation, allowing for faster and more efficient workflow.