Exploring the Intersection of Computer Vision and Generative Adversarial Networks in Medical Image Synthesis

Abstract

Pixel-wise similarity metrics, including Euclidean norms and mean squared errors, measure sample distortion and constitute a popular option in computer vision applications, such as super-resolution, style transformation, and image-to-image translation. However, this approach is not suitable for several domains, including psychology research. The sufficient discrimination between the types of image content that fulfill complex constraints, such as the segmentation of individual organs and vessels within the human body, can only be indicated by a learned distance metric. Consequently, recent state-of-the-art computer vision tasks in medical imaging are supported by learned data representations of deep neural networks. This chapter investigates the intersection between computer vision models and generative adversarial networks in the field of medical images. Specifically, we examine the performance and properties of conditional generative adversarial networks and their Wasserstein extension for real-time multi-organ segmentation of X-ray computed tomography.

The supervised training of machine learning models for computer vision tasks has become increasingly dependent on the availability of high-quality labeled data. However, annotations are often labor-intensive to obtain as they are typically requested from domain experts (e.g., physicians in the case of medical images) and thus generally limited in size. To overcome the restrictions imposed by the necessary presence of labeled data, generative adversarial networks were introduced for unsupervised learning and widely adopted by the research community. Generative image models aim to maximize the similarity between real and generated samples, enabling the production of synthetic data that emulates training samples. Such synthesized image datasets can then be employed to train a more generalized computer vision model.