CNN-based automatic segmentations and radiomics feature reliability on contrast-enhanced ultrasound images for renal tumors

• DOI: 10.3389/fonc.2023.1166988
• Publication Date: 2023-06-02T14:22:09Z
• Authors: Yin Yang; Fei Chen; Hongmei Liang; Yun Bai; Zhen Wang; Lei Zhao; Sai Ma; Qinghua Niu; Fan Li; Tianwu Xie; Yingyu Cai
• Keywords: Ground truth; Gold standard (test); Contrast-enhanced ultrasound

To investigate the feasibility and efficiency of automatic segmentation contrast-enhanced ultrasound (CEUS) images in renal tumors by convolutional neural network (CNN) based models their further application radiomic analysis.From 94 pathologically confirmed tumor cases, 3355 CEUS were extracted randomly divided into training set (3020 images) test (335 images). According to histological subtypes cell carcinoma, was split clear carcinoma (ccRCC) (225 images), angiomyolipoma (AML) (77 other (33 Manual gold standard serves as ground truth. Seven CNN-based including DeepLabV3+, UNet, UNet++, UNet3+, SegNet, MultilResUNet Attention UNet used for segmentation. Python 3.7.0 Pyradiomics package 3.0.1 feature extraction. Performance all approaches evaluated metrics mean intersection over union (mIOU), dice similarity coefficient (DSC), precision, recall. Reliability reproducibility radiomics features Pearson intraclass correlation (ICC).All seven achieved good performance with mIOU, DSC, precision recall ranging between 81.97%-93.04%, 78.67%-92.70%, 93.92%-97.56%, 85.29%-95.17%, respectively. The average coefficients ranged from 0.81 0.95, ICCs 0.77 0.92. UNet++ model showed best 93.04%, 92.70%, 97.43% 95.17%, For ccRCC, AML subtypes, reliability analysis derived automatically segmented excellent, 0.96 0.96, different 0.91, 0.93 0.94, respectively.This retrospective single-center study that had on tumors, especially model. feasible reliable, validation multi-center research is necessary.