Animals and drug treatmentSprague Dawley female rats (8 weeks old, weighing approximately 200–240 g) were purchased from the Vital River Laboratory Animal Technology. All rats were housed in a relatively consistent environment at room temperature (18–20 °C) and normal humidity (50–60%), with food and water available ad libitum, and were reared and bred in individually ventilated cages. After one week of acclimatization, the female rats were mated with sexually mature males (2:1), and the first day of gestation (gestation day, GD) was defined as the day when the vaginal plug was found. All pregnant rats were anesthetized by intraperitoneal injection of sodium pentobarbital (30 mg/kg). Rats were euthanized intravenously at the 19 days of gestation (GD19), fetal rats were subsequently evaluated by ultrasound; then, relevant specimens were collected for preservation. The experimental protocol was approved by the Institutional Animal Care and Use Committee (Ethical approval was obtained from the hospital institution Ethical Review Board, No. 2021–20; Date of approval, January 27, 2021). The animal experiments were performed were in accordance with the National Institutes of Health Guidelines for Use and Care of Animals (National Institutes of Health Publication No. 85–23). The authors complied with the ARRIVE guidelines.The pregnant rats were randomly divided into two groups: control group (sham operation, n = 6), which received a 0.9% intraperitoneal injection of normal saline daily from GD10 to GD18; and the L-NAME group (n = 6), which received an intraperitoneal injection of 125 mg/kg/day L-NAME (NG-Nitro-L-arginine Methyl Ester, L-NAME, Sigma-Aldrich, Merck KGaA, Darmstadt, Germany) daily from GD10 to GD18 intraperitoneally. The blood pressure of the rats was measured using a BP-2000 Animal Noninvasive Blood Pressure Analysis System27. The systolic BP (SBP) reported by the system represented the average of at least six valid BP measurements, featuring a normal wave pattern. The SBP of the rats in the two groups was detected at progestation (GD0), GD1 and GD18. Each pregnant rat was placed in a metabolic cage at GD0, GD1, or GD18, for the collection of 24h urine samples (see the supplemental details).Histopathological examinationThe specimen of the placenta was collected and then observed in its entirety, paying particular attention to its general appearance and basic morphology. After adequate fixation (4% paraformaldehyde), the tissue was dehydrated in gradient ethanol solution and immersed in wax according to a standard protocol. Subsequently, the wax blocks were carefully sliced (thickness 4 μm) and placed on slides. After dewaxing, hematoxylin–eosin staining was performed. The overall morphology and microstructure of the placenta were evaluated under a light microscope.Western blottingWestern blotting was conducted according to the specifications of the antibody manufacturer. Briefly, 20 μg aliquots of lysates were subjected to SDS-PAGE. The resulting protein bands were then transferred onto a PVDF membrane. Antibodies against sFlt-1 and PIGF, along with a Phototope-HRP western blot kit, were used for the western blotting assay.Evaluation of placental functionThe placentas were categorized into two distinct subgroups: the normal function group and the dysfunction group. A smooth pregnancy course, devoid of L-NAME drug intervention, served as a prerequisite for inclusion in the normal function category. Conversely, successful induction of preeclampsia was necessary for classification into the dysfunction group. Any of the following items indicated inclusion in dysfunction group: abnormal placental pathological structure, aberrant placental tissue expression, or deviant fetal rat development. Notably, animals in the control group that exhibited any of the three aforementioned anomalies were systematically excluded from analysis.The procedures of ultrasound-based radiomics
(i)
Placental ultrasound image acquisition
On the 19th day of gestation in pregnant rats, the pregnant rats were fixed in a supine after anesthesia, and their abdomen was disinfected after hair removal. Then ultrasonic examination was performed using an instrument (EPIQ Elite W, Intera, Philips, The Netherlands) with the linear array probe eL18-4 (the center frequency was 9.0 MHz). A sonographer with ≥ 5 years of experience in obstetrics and gynecology ultrasound imaging performed all of the examinations. Four complete images were randomly acquired for each fetal rat placenta. The placenta area was not blocked by fetal rat limbs or other body parts that produce an acoustic shadow, and tried to place the placenta area in the near field area to obtain a complete and accurate image of the placenta. The data were reviewed by a senior physician of 10 years and those with substandard or incomplete images were excluded. The operators and auditors involved in data collection were blinded to the groups. There were no annotations of arrows or any measured values on the images, and the images were collected and stored in DICOM format (.DCM).
(ii)
Placenta delineation
The regions of interest (ROIs) on the placental images were delineated using ITK-SNAP software (3.2.0-RC1) by automatically drawing boundaries between the ROI and background. The masks of the segmented placenta were reviewed and manually adjusted by a radiologist experienced in obstetric imaging. A detailed illustration of the steps of semiautomatic placenta segmentation is provided in Fig. 1.Fig. 1The procedures of utrasound-based radiomics.
(iii)
Feature extraction
The feature catalog was divided into six main categories, containing a total of 93 features. Then, texture extraction of the 93 features was carried out on 6 components respectively, and 558 features were obtained (please refer to the supplemental materials for details). All features are numbered sequentially (1–558) so that each feature is associated with a feature serial number (FSN).
(iv)
Features analysis
Multifactor analysis was carried out to remove redundant features, preventing overfitting of the model. The Boruta method was used for feature selection, and a random forest algorithm was used to measure the feature importance, ultimately reducing the dimensionality of the input model.
(v)
Model construction
After feature selection, the model was constructed with the training set as the inputs. Through experimental comparison, the extreme random tree algorithm28 was selected as the classifier. Using Bayesian search theory and tenfold cross-validation as the standard, iterative optimization was carried out, and the optimal phenotypic model was ultimately established and trained.Statistical analysesIBM SPSS Statistics 20.0 (SPSS, Inc., New York) was used for statistical analysis in this study. Continuous variables were statistically analyzed with independent t tests or nonparametric rank sum tests. Categorical variables were analyzed with the chi-square test. Redundant features were removed based on the Pearson correlation coefficient, and the Boruta feature selection method was utilized to reduce the feature dimension of the input model. This was done to select useful features for the model. For bilateral tests, P < 0.05 was considered to indicate statistical significance.The microscopic observation and image analysis of placental pathology were conducted under the guidance of two pathologists.