2.1 Animal Model
The study was approved by institutional board of Animal Care and Use Committee, the Second People's Hospital of Wuhu City, Anhui Province, China, in compliance with China national guidelines for the care and use of animals.Ninety male SD rats (initial weight: 140-170g) were selected. Choosing male SD rats was to avoid hormonal fluctuations, as female rats undergo hormonal fluctuations during their estrous cycles, which could potentially affect experimental outcomes.After a three-day adaptation period and numbering, the rats were randomly divided into two groups: a normal control group consisting of 30 rats fed a standard diet and a high-fat group consisting of 60 rats fed a high-fat diet.The high-fat diet used was the internationally certified D12492 high-fat diet, with the following nutritional composition: 20% protein, 20% carbohydrates, and 60% fat.The high-fat group was administered a mixture of CCl4 in oil (1:4) from the 8th week until the 14th week (0.2 mL/kg; twice a week; intraperitoneally).In the normal control group, 10 rats were randomly selected at the 4th, 10th, and 14th weeks for imaging examination. In the high-fat group, 10 rats were randomly selected at the 4th, 6th, 8th, 10th, 12th, and 14th weeks after anesthesia for imaging examination. Subsequently, blood samples were collected from the abdominal aorta, and finally, under anesthesia, the rats were euthanized by cervical dislocation, and liver tissue was obtained.
2.2Ultrasound Data Acquisition
Examination is performed using the ACUSON Sequoia real-time shear wave elastography ultrasound diagnostic system (Brand: Siemens, Origin: USA), equipped with a standard linear array 10L4 transducer (4-10MHz), and an animal experiment-specific V6 ultrasound diagnostic system (Brand: FiNo, Origin: China), equipped with an X4-12L transducer (4-12MHz).All rats were fasted for one day before the examinations and then anesthetized by intraperitoneal injection of pentobarbital sodium (3% saline solution, 40mg/kg),After successful anesthesia, the rats' abdominal fur was removed, and they were securely positioned on the operating table for the ultrasound scan.Grayscale ultrasound was used to observe liver size, shape, and echogenicity. Two-dimensional (2D) images of the central lobe of the liver were uniformly stored. Subsequently, the system was switched to 2D-SWE and P-SWE modes. A region of interest (ROI) was uniformly selected in the central lobe of the liver, avoiding large blood vessels and bile ducts, and located 1 cm below the body surface. Emean of the ROI was measured five times, and the mean value was calculated.Additionally, in the 2D-SWE mode, adjacent sections of the liver and kidney were selected, and the elasticity values of the liver and kidney were measured to calculate their liver-to-kidney elasticity ratio (ratio).
2.3Blood and Tissue Sampling
Following the completion of the ultrasound scans, the rats' body weights were measured. Subsequently, blood samples were collected through abdominal aorta puncture, and the rats were euthanized.Blood was collected into anticoagulant tubes, centrifuged at 4000rpm for 15 minutes, and the serum was stored at -80°C for later use. Simultaneously, the liver was quickly separated, rinsed with physiological saline, and the central lobe of the liver was fixed in 10% neutral formalin for further processing.
2.4 Serum and Histopathological Analysis
The collected serum samples were labeled and sent to the laboratory at our Hospital for analysis.The serum samples were tested for the following parameters: alanine transaminase (ALT), aspartate transaminase (AST), AST/ALT ratio, gamma-glutamyl transferase (GGT), total cholesterol (CHOL), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C).
The liver tissues were fixed in 10% neutral formalin for 24 hours, followed by routine tissue processing and staining with hematoxylin-eosin (HE). Pathological diagnosis was performed using the SAF scoring system for NAFLD diagnosis.①Degree of Fat Deposition (0-3 points): Based on the percentage of hepatocytes containing large or medium-sized lipid droplets, it was divided into S0-3, corresponding to 0-3 points. S0 (<5%, none); S1 (5-33%, mild); S2 (>33-66%, moderate); S3 (>66%, severe).②Degree of Fibrosis (0-4 points): 0: None; F1: Peri-sinusoidal or peri-portal; F2: Both peri-sinusoidal and peri-portal; F3: Bridging fibrosis; F4: Cirrhosis.③Degree of Inflammation (0-4 points, sum of hepatocellular ballooning score and lobular inflammation score):Hepatocellular Ballooning: 0-2 points (0: Hepatocytes are cuboidal with normal size and pink, acidophilic cytoplasm; 1: Hepatocytes are round with pale, reticulated cytoplasm, clustered, and similar in size to normal hepatocytes; 2: Hepatocytes are at least twice the size compared to grade 1).Lobular Inflammation: 0-2 points: Presence of two or more foci of inflammatory cell aggregates within lobules. Count the number of aggregates observed at 20x magnification (0: None; 1:≤2 foci of aggregates; 2:﹥2 foci of aggregates).
According to the SAF scoring system, NAFLD is categorized into Non-Alcoholic Fatty Liver (NAFL) and Non-Alcoholic Steatohepatitis (NASH).NASH is diagnosed when there is concurrent hepatocellular fat deposition, ballooning, and lobular inflammation, which means a score of 1 for each criterion, with a total score of 3 points or higher.
2.5 Radiomic Feature Extraction and Selection
The ultrasound images were exported from our imaging system .Then, ITK-SNAP 3.8.0 (www.itksnap.org/) was used to delineate the contours of the Region of Interest (ROI) for radiomic feature extraction, model building, and evaluation.Radiomic features were automatically extracted from each image using the "pyradiomics" toolkit.To select radiomic features with good reproducibility and low redundancy, the following steps were taken:①Independent sample t-tests were performed on all extracted features, and features with p > 0.05 were removed.②For highly repetitive features, the Pearson correlation coefficient was calculated to express the relationship between features, and only one of any pair of features with a correlation coefficient >0.9 was retained.③The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was applied, and ten-fold cross-validation was used to determine the optimal λ value. Based on the model corresponding to the best λ value, non-zero coefficient radiomic features were selected.④All selected features were standardized using the Z-score method.⑤Finally, radiomic features and their corresponding coefficients were selected based on the LASSO algorithm.
2.6 Establishment of Clinical Models, Radiomic Models, and Clinical Radiomic Models.
After feature selection, various machine learning (ML) classification algorithms were used to establish clinical models, radiomic models, and clinical radiomic models. These algorithms included Logistic Regression (LR), Support Vector Machine (SVM), K Nearest Neighbor (KNN), Random Forest (RF), Extremely Randomized Trees (ExtraTree), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Multi-Layer Perceptron (MLP).Figure 1 in the analysis pipeline illustrates the workflow for establishing the clinical radiomic model.
2.7 Model Evaluation
Receiver Operating Characteristic (ROC) curve area under the curve (AUC) and Decision Curve Analysis (DCA) were used to evaluate the performance and clinical utility of the radiomic model, clinical model, and clinical radiomic model in both the training and validation sets. Additionally, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were also assessed.
2.8 Statistical Analysis
Statistical analysis was conducted using R software (version 4.3.2) and Python (version 3.7.2).Continuous data were evaluated for consistency between the training and validation sets using independent sample t-tests.The performance of the models was evaluated using ROC curves, and the AUC of each prediction model was compared using the DeLong test. The clinical value of various prediction models was compared using Decision Curve Analysis (DCA). Model fitting was assessed using calibration curves. The calculations for DCA mainly utilized the "rms" and "rmda" packages in R. Statistical significance for all two-tailed tests was set at P < 0.05.