Performance of a generative adversarial network using ultrasound images to stage liver fibrosis and predict cirrhosis based on a deep-learning radiomics nomogram

AIM: To investigate the performance of a generative adversarial network (GAN) model for staging liver fibrosis and its radiomics-based nomogram for predicting cirrhosis.MATERIALS AND METHODS: This two-centre retrospective study included 434 patients for whom input data of ultrasound images and histopathological data (obtained within 1 month of ultrasound examinations) were assigned to the training cohort (249 patients), the internal cohort (92 patients), and the external (93 patients) cohort. A data augmentation method based on a GAN model was used. The discriminative performance was evaluated for classifying fibrosis of S4 and >= S3. Deep-learning radiomics features were extracted for the prediction of cirrhosis (S4). To perform feature reduction and selection, the least absolute shrinkage and selection operator (LASSO) algorithm was applied. Radiomics scores, along with clinical factors, were incorporated into a nomogram using multivariable logistic regression analysis. The performance of the models was estimated with respect to discrimination power, calibration, and clinical benefits.RESULTS: The areas under the receiver operating characteristic curve (AUCs) values of the GAN were 0.832/0.762 (>= S3), and 0.867/0.835 (S4) for internal/external test sets, respectively. The radiomics nomogram that intergrated radiomics scores and clinical factors showed good calibration and discrimination ability of 0.922 (AUC) in the training dataset, 0.896 in the in-ternal dataset, and 0.861 in the external dataset. Decision curve analysis (DCA) demonstrated that the nomogram outperformed radiologist and haematological indices in terms of the most clinical benefits.CONCLUSIONS: The GAN model could be applied to discriminate fibrosis stages, and a favourable predictive accuracy for diagnosing cirrhosis was achieved using a deep-learning radiomics nomogram. (c) 2022 The Royal College of Radiologists. Published by Elsevier Ltd. All rights reserved.