Liver fibrosis is the progressive process of repeated injury of hepatocytes by different causes, and excessive deposition and abnormal distribution of extracellular matrix in the liver [1]. Liver fibrosis is a key step in the progression of chronic liver disease to cirrhosis [1,2]. It had been demonstrated that fibrosis can be reversed after removal of injury factors in the past few decades [1]. Five stages of liver fibrosis are estimated by METAVIR scoring system [3]. METAVIR F0 is no liver fibrosis (NLF), F1 corresponds to fibrosis with expansion of portal zones, F2 represents fibrosis with expansion of most portal zones and occasional bridging, F3 represents fibrosis with expansion of most portal zones and marked bridging and occasional nodules. The last stage METAVIR F4 is cirrhosis. METAVIR F1-F2 are regarded as early liver fibrosis (ELF) and METAVIR F3-F4 are considered to be progressive liver fibrosis (PLF). Biopsy is the “gold standard” for assessing liver fibrosis [4], while it is traumatic and risky.

Deep learning (DL), a sub-branch of machine learning (ML), can be used for automatically learning data to get valid feature representations for organ segmentation，tumor detection and disease staging [5]. DL has been gradually used in improving the accuracy of detecting disease [6,7]. DL can predict liver fibrosis stages by analyzing systematically liver morphology and texture changes [8]. The models adopted by deep learning include Convolutional Neural Network (CNN), Fully Convolutional Network, U-Net model [[9], [10], [11]] and others. Choi et al. validated a deep learning system based on CNN for staging liver fibrosis using portal venous phase CT images. The result showed a staging accuracy of 79.4% in the test data set [7]. Xue et al. proposed a deep CNN by transfer learning to analyze liver fibrosis images of multimodal gray scale modality and the AUCs of elastogram modality were 0.950, 0.932 and 0.930 for classifying S4, ≥S3, and ≥ S2, respectively [12]. Yu et al. used a CNN and transfer learning-based algorithm to stage liver fibrosis by Second Harmonic Generation microscopy images [13]. They showed the balanced area under receiver operating characteristic values of CNN was up to 0.85–0.95. Hectors et al. compared the VGG16 model from ImageNet with MR elastography (MRE), and they found that there was no significant difference between DL and MRE for staging liver fibrosis [14]. Few studies evaluated liver fibrosis by using deep learning in rats. Cheng et al. staged liver fibrosis by deep learning models based on ultrasound radio frequency signals in rats [15], and the accuracy of the deep learning networks using the training and validation data was above 0.83 and 0.80 and the AUCs were higher than 0.95 and 0.93. Obviously, deep learning is playing a vital role in the automatic segmentation of organs and the classification of diseases [16].

For the task of organic segmentation, nnU-Net is a newly-developed deep learning neural network for medical image segmentation in recent years built on the fully convolutional network [[17], [18], [19]]. For feature extraction and segmentation map generation, the nnU-Net architecture consists of a down-sampling path and an up-sampling path [18]. In the process of down-sampling, feature channels are gradually added in order to obtain the feature maps; the segmented images with high resolution can be obtained by the up-sampling architecture. In the recent years, nnU-Net shows excellent performance in lung lesions segmentation [19], liver segmentation [20], atherosclerotic plaque segmentation [21], whole breast and fibroglandular tissue segmentation [22] and myocardial scar tissue segmentation [23]. The nnU-Net neural network model could automatically adapt architecture to different image [18] and is the improved version based on standard U-Net model. The nnU-Net algorithm, as an emerging convolutional neural network method for medical image segmentation, had recently been validated as an accurate tool for segmentation from 2D or 3D image modalities [17,20].

This study validates the feasibility of using a segmentation-classification cascaded deep learning model for fully automated liver fibrosis grading on T1-weighted MRI images of rats. Further, this study is expected to show DL model helping clinicians improving clinical practice with high accuracy.