An evolvable adversarial network with gradient penalty for COVID-19 infection segmentation

Abstract

COVID-19 infection segmentation has essential applications in determining the severity of a COVID-19 patient and can provide a necessary basis for doctors to adopt a treatment scheme. However, in clinical applications, infection segmentation is performed by human beings, which is time-consuming and generally introduces bias. In this paper, we developed a novel evolvable adversarial framework for COVID-19 infection segmentation. Three generator networks compose an evolutionary population to accommodate the current discriminator, i.e., generator networks evolved with different mutations instead of the single adversarial objective to provide sufficient gradient feedback Compared with the existing work that enforces a Lipschitz constraint by weight clipping, which may lead to gradient exploding or vanishing, the proposed model also incorporates the gradient penalty into the network, penalizing the discriminator’s gradient norm input. Experiments on several COVID-19 CT scan datasets verified that the proposed method achieved superior effectiveness and stability for COVID-19 infection segmentation.

CHS-Net: A Deep Learning Approach for Hierarchical Segmentation of COVID-19 via CT Images

“…The authors achieved promising results with a dice value of 0.78 while outperforming other state-of-the-art models. In similar work, an evolvable adversarial learning strategy is proposed by He et al [ 41 ], where three different mutation operators are utilized to train the generator with added gradient penalty for producing stable COVID-19 infection segmentation. However, for real word implications of such approaches, it is critical to quantify the prediction uncertainty of the model [ 42 ].…”

COV-DLS: Prediction of COVID-19 from X-Rays Using Enhanced Deep Transfer Learning Techniques

“…The performance of this method can be further enhanced through the concept of noise annotations. He et al [ 24 ] presented an adversarial framework for discriminating COVID-19-infected patients using chest CT images. Three mutation operators were used to modify the generator for segmentation, and a gradient penalty was used to eliminate gradient vanishing.…”

TSRNet: Diagnosis of COVID-19 based on self-supervised learning and hybrid ensemble model

“…To solve the problem of relying more on real image label information in previous deep learning segmentation methods, Xiaoming Liu et al [ 21 ] proposed a weakly supervised segmentation method for COVID-19 patient chest CT images. Juanjuan He et al [ 22 ] developed an evolvable adversarial framework for COVID-19 patients. The method used three different mutation-evolving generator networks and incorporated gradient penalties into the model to achieve excellent performance in segmenting chest CT images of patients with COVID-19.…”