Placenta preposed postpartum hemorrhage risk prediction system

Abstract

The invention relates to the field of medical image and machine learning, and discloses a placenta preposed postpartum hemorrhage risk prediction system. Comprising a radiomics feature acquisition unit used for acquiring MRI radiomics features of a preplacenta postpartum hemorrhage patient; the clinical feature acquisition unit is used for acquiring clinical features of the preplacenta postpartum hemorrhage patient; the subset construction unit is used for constructing a feature subset based on the clinical features and the MRI radiomics features; and the prediction unit is used for inputting the feature subset into a TabPFN model and carrying out postpartum hemorrhage probability prediction and hemorrhage quantity quantitative prediction. According to the method, automatic segmentation of placenta MRI is realized by adopting fine-tuned nnU-Net, and the average Dice coefficient reaches 0.83. The problems of non-uniform intensity distribution, placenta thickness variation and adjacent anatomical structure interference in the placenta MRI image are effectively solved. According to the method, accurate risk stratification and quantitative evaluation of preplacental postpartum hemorrhage are realized by integrating MRI image omics characteristics and clinical data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical imaging and machine learning, and in particular to a placenta previa postpartum hemorrhage risk prediction system. BACKGROUND

[0002] As a major risk factor for postpartum hemorrhage (PPH) in cesarean section, the clinical management of placenta previa highly depends on preoperative risk stratification.

[0003] Currently, clinical practice mainly relies on ultrasound assessment and traditional scoring systems (such as placenta accreta spectrum score), but there are significant limitations: ultrasound assessment is highly influenced by operator experience and is highly subjective, often leading to risk misjudgment; existing prediction models are prone to overfitting and have poor generalization ability due to small sample size and unbalanced data; MRI can provide more accurate anatomical information, but relies on manual segmentation, which is time-consuming and has inter-observer differences, making it difficult to scale up.

[0004] In addition, existing methods can only perform qualitative risk classification and cannot quantify the amount of bleeding, making it difficult to guide specific resource allocation. These defects result in insufficient accuracy in clinical assessment of the risk of hemorrhage in patients with placenta previa, which can lead to excessive intervention or inadequate preparation, increasing the risk of maternal and infant complications. SUMMARY

[0005] The main purpose of the present application is to solve the technical problem of insufficient accuracy in hemorrhage risk assessment in the prior art. A placenta previa postpartum hemorrhage risk prediction system comprises: an imageomics feature acquisition unit for acquiring MRI imageomics features of a placenta previa postpartum hemorrhage patient; a clinical feature acquisition unit for acquiring clinical features of a placenta previa postpartum hemorrhage patient; a subset construction unit for constructing a feature subset based on clinical features and MRI imageomics features; a prediction unit for inputting the feature subset into a TabPFN model to predict the probability of postpartum hemorrhage and the amount of bleeding.

[0006] The feature subset is obtained through multi-modal feature screening based on clinical features and MRI imageomics features; The MRI imageomics features are obtained by extracting features using PyRadiomics v3.0.1 based on an automatic segmentation placenta mask; The clinical features are extracted from electronic medical records, including demographic, obstetric history, current pregnancy parameters, and laboratory indicators.

[0007] The TabPFN model does not require additional parameter tuning, uses pre-trained weights for fine-tuning, and uses stratified sampling to ensure class balance in the training / test set.

[0008] A method for constructing a preplacental postpartum hemorrhage risk prediction system, comprising: Obtaining MRI imageomics features of preplacental postpartum hemorrhage patients; Obtaining clinical features of preplacental postpartum hemorrhage patients; Constructing a feature subset based on clinical features and MRI imageomics features; Inputting the feature subset into the TabPFN model to predict the probability of postpartum hemorrhage and the amount of hemorrhage.

[0009] The present application has the following beneficial effects: The present application constructs a multi-modal fusion model based on deep learning automatic segmentation, which realizes accurate risk stratification and quantitative evaluation of preplacental postpartum hemorrhage by integrating MRI imageomics features and clinical data.

[0010] The present application uses fine-tuned nnU-Net to realize automatic segmentation of placental MRI, with an average Dice coefficient of 0.83. It effectively overcomes the problems of uneven intensity distribution, placental thickness variation and adjacent anatomical structure interference in placental MRI images.

[0011] The present application extracts imageomics features (such as texture, morphological parameters) from the automatic segmentation results, and combines clinical data (such as gestational age, cesarean section history, preoperative hemoglobin, etc.), and selects 28 clinical features and 59 imageomics features through multi-method feature selection (including correlation analysis, LASSO, random forest, recursive feature elimination), reducing redundancy and retaining key information.

[0012] The present application uses a pre-trained TabPFN model to process the screened multi-modal features. This model is optimized for small sample data and can achieve robust risk classification (determine whether ≥1000 mL of bleeding) and bleeding amount quantitative regression in limited samples (n=308).

[0013] The present application analyzes the model output through SHAP value, identifies key influencing factors, and enhances clinical interpretability. This model has both risk stratification and quantitative prediction functions, and can provide direct basis for preoperative resource allocation. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The entire training evaluation flowchart; Figure 2 The feature screening result; Figure 3 The classification prediction hemorrhage risk evaluation result ROC chart; Figure 4For regression prediction of bleeding volume assessment result evaluation chart; Figure 5 For prospective six case classification and regression chart; Figure 6 For prospective six case specific prediction table. DETAILED DESCRIPTION

[0015] The terms "first", "second", "third", "fourth" and the like in the description and claims of the present application and above figures (if any) are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so construed herein can interchange, so that, for example, without limitation, the embodiments described herein can operate in other sequences than illustrated or described herein. Further, the terms "comprise" or "comprising", and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, product or apparatus that comprises a list of steps or units is not necessarily limited to those steps or units that are expressly listed, but can include additional steps or units that are not expressly listed or inherent to such process, method, product or apparatus.

[0016] For the convenience of understanding, the specific flow of the embodiments of the present application is described below, please refer to Figure 1 The first embodiment of the placenta previa postpartum hemorrhage risk prediction system in the embodiments of the present application comprises: An imageomics feature acquisition unit is configured to acquire MRI imageomics features of a patient with placenta previa postpartum hemorrhage. A clinical feature acquisition unit is configured to acquire clinical features of the patient with placenta previa postpartum hemorrhage. A subset construction unit is configured to construct a feature subset based on the clinical features and the MRI imageomics features. A prediction unit is configured to input the feature subset into a TabPFN model to perform postpartum hemorrhage probability prediction and bleeding volume quantitative prediction.

[0017] The present application also relates to a method for constructing a placenta previa postpartum hemorrhage risk prediction system, comprising: Acquiring MRI imageomics features of a patient with placenta previa postpartum hemorrhage. Acquiring clinical features of the patient with placenta previa postpartum hemorrhage. Constructing a feature subset based on the clinical features and the MRI imageomics features. Inputting the feature subset into a TabPFN model to perform postpartum hemorrhage probability prediction and bleeding volume quantitative prediction.

[0018] The feature subset is composed of 28 clinical features and 59 imageomics features, which are respectively: The feature subset is obtained by screening a multi-modal feature based on clinical features and MRI imageomics features; The MRI imageomics features are obtained by extracting features using PyRadiomics v3.0.1 based on an automatically segmented placenta mask; The clinical features are obtained by extracting from electronic medical records, including demographic, obstetric history, current pregnancy parameters, and laboratory indicators.

[0019] The automatic segmentation of the placenta mask is obtained in the following manner: To achieve accurate segmentation of the placenta, the present application fine-tunes the 3D U-Net architecture of the nnU-Net framework on the following parameters, and establishes a complete technical process covering network design, data preprocessing, model training and performance verification. Among them, the network uses an improved 3D U-Net backbone network with a seven-level hierarchical structure, and the number of channels at each level is [32, 64, 128, 256, 320, 320, 320] in turn. Through differentiated convolution kernel size (range from [1,3,3] to [3,3,3]) and step size settings (range from [1,1,1] to [1,2,2]) to realize multi-scale feature extraction, and each level contains two convolution blocks, supplemented by instance normalization with learnable affine parameters (ε=10 -5 ) and Leaky ReLU activation function (α=0.01) to improve network stability; in the data preprocessing stage, first use N4 algorithm to correct the intensity inhomogeneity of the MRI image, then perform Z-score normalization based on muscle signal, and then resample the image to the target interval of [6.0, 0.62, 0.62] mm, wherein the image uses third-order B-spline interpolation, and the mask uses nearest neighbor interpolation; during model training, use Nesterov stochastic gradient descent method with momentum (μ=0.99), set the initial learning rate to 0.01 and use a polynomial decay strategy, and through weight decay (λ=3×10 -5To mitigate overfitting, the loss function combines the Dice coefficient and cross-entropy with deep supervision. Simultaneously, data diversity is enhanced through rigid transformation and intensity shift data augmentation. Training is performed on an NVIDIA RTX A6000 GPU with a batch size of 2 and 1000 training epochs. After five-fold cross-validation, the model achieves an average Dice coefficient of 0.83 on a set of 40 independent test cases, effectively overcoming the problems of uneven intensity distribution, placental thickness variation (1.5-4.3cm), and interference from adjacent anatomical structures in placental MRI images. Feature selection strategy: Four methods are combined for selection (Figure 2): ① Correlation analysis (CORR): Calculating the Pearson correlation coefficient between features and hemorrhage volume; ② LASSO regression: Determining the regularization strength through five-fold cross-validation and selecting features with non-zero coefficients; ③ Random forest (RF): Random forest importance ranking based on 200 trees; ④ Recursive feature elimination (RFE): Using random forest as the base classifier, progressively removing the 10% of the least important features. Each method was repeated 10 times (Gaussian noise was added to enhance stability), and the top-m features that appeared frequently were retained. Finally, the union of the screening results of the four methods was taken to obtain 28 clinical features and 59 radiomics features.

[0020] The TabPFN model requires no additional parameter tuning, utilizes pre-trained weights for fine-tuning, and employs stratified sampling to ensure class balance in the training / test sets.

[0021] The specific prediction process of this invention includes: 1. Data collection and preprocessing (Figure 1) Data source: The study included 308 patients with placenta previa as a retrospective cohort (261 in the training set and 47 in the independent test set), and 6 prospective patients were included for clinical validation. All patients met the following criteria: placenta previa diagnosed by color Doppler ultrasound, gestational age ≥28 weeks, cesarean section with preoperative T2-weighted MRI, and complete clinical data (including demographic characteristics, obstetric history, laboratory tests, and actual intraoperative blood loss). Exclusion criteria were: incomplete clinical records, preoperative stillbirth, or poor MRI image quality that prevented segmentation.

[0022] 2. Automatic Placental Segmentation and Optimization We used partial data to manually label placental segmentation and then fine-tuned the training process. The following are the specific training and fine-tuning scheme and example parameter details: A 7-level encoder-decoder architecture was adopted, with channel widths of [32, 64, 128, 256, 320, 320, 320]. The convolutional kernels gradually transitioned from anisotropic [1,3,3] to isotropic [3,3,3] (adapting to the morphological characteristics of the placenta in the sagittal plane), and the downsampling stride was adjusted accordingly from [1,1,1] to [1,2,2]. After 5-fold cross-validation, the model achieved an average Dice similarity coefficient of 0.83 on the test set (Figure 2), indicating that the segmentation results were highly consistent with manual annotation (completed by radiologists using 3DSlicer), and the batch processing time was reduced to 4.2 ± 1.3 minutes per case, significantly better than manual segmentation (which required 30-60 minutes per case).

[0023] An open-source technology adaptation solution for placenta previa scenarios: When applying the nnU-Net automatic segmentation algorithm to placenta previa MRI, considering the special characteristics of placental tissue (such as the signal difference between the placenta and surrounding tissues in T2-weighted images), the preprocessing parameters (such as the intensity of N4 bias correction and Z-score normalization for muscle signal reference) and post-segmentation processing logic (such as placental boundary smoothing strategy) are optimized. This enables the open-source algorithm to achieve stable segmentation (Dice coefficient ≥ 0.83) in the specific pathological scenario of placenta previa, solving the problem of insufficient segmentation accuracy when the open-source model is directly applied to pathological placenta.

[0024] 3. Feature extraction and screening Radiomics Feature Extraction: Based on an automatically segmented placental mask, PyRadiomics v3.0.1 was used to extract features, including: ① morphological features (such as volume and surface area); ② first-order statistics (such as mean and skewness); ③ texture features (GLCM, GLRLM, GLSZM, etc.). Before extraction, image enhancement (Laplacian Gaussian filtering, wavelet decomposition) and outlier processing (truncated to the 99th percentile) were performed, initially generating 989 features.

[0025] Clinical feature collection: 39 preoperative variables were extracted from electronic medical records, covering demographics (age, BMI), obstetric history (number of cesarean sections, history of uterine surgeries, number of intrauterine procedures), current pregnancy parameters (IVF conception), and laboratory indicators (preoperative hemoglobin, coagulation function, D-dimer), etc.; Feature screening strategy: Four methods were used in combination for screening ( Figure 2The four methods were: ① Correlation Analysis (CORR): Calculating the Pearson correlation coefficient between features and bleeding volume, repeated 10 times (Gaussian noise was added to enhance stability), retaining the top-m features that appeared frequently; ② LASSO Regression: Determining the regularization strength through 5-fold cross-validation, and screening features with non-zero coefficients; ③ Random Forest (RF): Ranking the importance of 200 trees using Gini importance ranking; ④ Recursive Feature Elimination (RFE): Using Random Forest as the base classifier, progressively removing the 10% of the least important features. Finally, the union of the results from the four methods was used to obtain 28 clinical features (such as number of cesarean sections and preoperative hemoglobin) and 59 radiomics features (such as wavelet transform texture parameters), which both preserved key clinical bleeding-related signals and avoided the bias of a single method.

[0026] The multimodal feature fusion-supported quantitative regression model of this invention constructs a multimodal feature set based on clinical features (such as cesarean section history and preoperative hemoglobin) and MRI radiomics features (such as wavelet transform texture parameters). Through a combination strategy of "correlation analysis + LASSO + random forest + recursive feature elimination", a subset of features highly correlated with bleeding volume (28 clinical features + 59 radiomics features) is selected and input into the TabPFN model to achieve quantitative regression prediction of bleeding volume under small sample conditions. This feature fusion and regression modeling method is unique to the placenta previa scenario and fills the gap of existing technologies that can only qualitatively stratify and cannot quantitatively predict bleeding volume.

[0027] The feature subset consists of 28 clinical features and 59 radiomics features, as follows: 28 clinical features: G (Gravidity): Parity - Total number of pregnancies.

[0028] P (Parity): Parity - The number of times a live fetus has been delivered after 28 weeks of gestation.

[0029] CS (Caesarean Section): Number of previous cesarean sections.

[0030] Cavity: refers to the total number of previous induced abortions, diagnostic curettages, and hysteroscopic surgeries.

[0031] IVF (In Vitro Fertilization): Whether or not pregnancy is achieved through assisted reproductive technology.

[0032] GDM (Gestational Diabetes Mellitus): Gestational diabetes mellitus - diabetes that develops during pregnancy.

[0033] OC: Other Chronic Diseases (0 None / 1 Yes) PH (Past History / Pulmonary Hypertension): refers to bleeding events during the current pregnancy (from 28 weeks of gestation to delivery) such as bleeding related to placenta previa or placental abruption.

[0034] GW (Gestational Weeks): Gestational Weeks - Current number of days into pregnancy.

[0035] DM: Delivery Mode: Delivery Method (0 Elective Cesarean Section / 1 Emergency Cesarean Section) WBC (White Blood Cell count): White blood cell count - reflects the immune system and infection status.

[0036] HB_before (Hemoglobin before): Pretreatment hemoglobin - reflects anemia status.

[0037] PLT (Platelet count): Platelet count - related to coagulation function.

[0038] FIB (Fibrinogen): Fibrinogen - an important clotting factor.

[0039] APTT (Activated Partial Thromboplastin Time): Activated partial thromboplastin time - detects the intrinsic coagulation pathway.

[0040] TT (Thrombin Time): Thrombin time - a common pathway for detecting coagulation.

[0041] DD (D-Dimer): D-dimer is an important indicator reflecting thrombus formation and dissolution. An elevated level often suggests a hypercoagulable state or thrombosis.

[0042] TBIL (Total Bilirubin): Reflects liver function, biliary tract condition, and hemolytic diseases.

[0043] AST (Aspartate Aminotransferase): Aspartate aminotransferase - an enzymatic indicator of liver and myocardial function.

[0044] ALP (Alkaline Phosphatase): Alkaline phosphatase - associated with liver and biliary diseases and bone diseases.

[0045] TBA (Total Bile Acid): Elevated bile acid levels during pregnancy are a key diagnostic indicator of intrahepatic cholestasis of pregnancy (ICP), posing a serious risk to the fetus.

[0046] ALB (Albumin): Albumin - the main protein in the blood, reflecting nutritional and liver function status.

[0047] GLO (Globulin): Globulin - associated with immunity and inflammation.

[0048] PA (Prealbumin): Prealbumin - reflects short-term nutritional status and liver function.

[0049] LDH (Lactate Dehydrogenase): Elevated levels of lactate dehydrogenase are seen in tissue damage, such as hemolysis, liver disease, tumors, and preeclampsia.

[0050] Urea: a kidney function indicator.

[0051] Scr (Serum Creatinine): Serum creatinine - a core indicator of renal function.

[0052] UA (Uric Acid): Uric acid - a kidney function indicator, often elevated in gestational hypertension.

[0053] 59 radiomics features: Use the naming convention of [Image Filter]-[Feature Category]_[Feature Name]: log-sigma-1-0-mm-3D_firstorder_Energy Image filter: LoG filter with σ=1.0mm (emphasizing fine structure); Feature category: First-order statistics; Feature name: Energy; Meaning: Calculates the sum of squares of all pixel values ​​on an image emphasizing fine structure. A higher value indicates a more uniform or stable distribution of pixel intensity.

[0054] log-sigma-2-0-mm-3D_firstorder_Range Image filter: LoG filter with σ=2.0mm (emphasizing a medium to fine structure); Feature category: First-order statistics; Feature name: Range; Meaning: Calculates the difference between the maximum and minimum pixel values ​​on the processed image, reflecting the total span of intensity.

[0055] log-sigma-2-0-mm-3D_firstorder_Energy Image filter: LoG filter with σ=2.0mm (emphasizing medium-fine structures); Feature category: First-order statistics; Feature name: Energy; Meaning: Calculates the sum of squares of pixel values ​​on an image emphasizing medium-fine structures to measure the uniformity of intensity.

[0056] log-sigma-3-0-mm-3D_firstorder_Range Image filter: LoG filter with σ=3.0mm (emphasizing medium-coarse structures); Feature category: First-order statistics; Feature name: Range; Meaning: Calculates the difference between the maximum and minimum pixel intensities on an image emphasizing medium-coarse structures.

[0057] log-sigma-3-0-mm-3D_firstorder_Kurtosis Image filter: LoG filter with σ=3.0mm (emphasizing a medium-thickness structure); Feature category: First-order statistical feature; Feature name: Kurtosis; Meaning: Calculates the steepness of the distribution of all pixel values ​​on the processed image.

[0058] log-sigma-3-0-mm-3D_firstorder_Energy Image filter: LoG filter with σ=3.0mm (emphasizing medium-coarse structure); Feature category: First-order statistics; Feature name: Energy; Meaning: Calculates the sum of squares of pixel values ​​on an image emphasizing medium-coarse structure.

[0059] wavelet-LHL_firstorder_Energy Image Filter: Wavelet Transform (LHL: Low-pass filtering on the X-axis, High-pass filtering on the Y-axis, and Low-pass filtering on the Z-axis); Feature Category: First-order Statistics; Feature Name: Energy; Meaning: Calculate the sum of squares of pixel intensities on the image after wavelet decomposition in a specific direction.

[0060] wavelet-LHL_gldm_DependenceVariance Image filter: Wavelet transform (LHL); Feature category: Gray-level dependency matrix; Feature name: Dependency variance; Meaning: Calculates the variance of gray-level dependency on the filtered image to measure the heterogeneity of texture.

[0061] wavelet-LHH_glszm_SizeZoneNonUniformity Image Filter: Wavelet Transform (LHH: Low-pass filtering on the X-axis, high-pass filtering on the Y-axis, and high-pass filtering on the Z-axis); Feature Category: Gray-level Region Size Matrix; Feature Name: Region Size Non-uniformity; Meaning: Calculates the degree of non-uniformity in the distribution of continuous region sizes on the filtered image. A high value indicates large differences in region sizes and strong image heterogeneity.

[0062] wavelet-HLL_firstorder_Energy Image Filter: Wavelet Transform (HLL: High-pass filtering on the X-axis, Low-pass filtering on the Y-axis, and Low-pass filtering on the Z-axis); Feature Category: First-order Statistics; Feature Name: Energy; Meaning: Calculate the sum of squares of pixel intensities on the image after wavelet decomposition in a specific direction.

[0063] wavelet-HLL_glrlm_GrayLevelNonUniformity Image Filter: Wavelet Transform (HLL); Feature Category: Gray-Level Run-Length Matrix; Feature Name: Gray-Level Non-uniformity; Meaning: Calculates the uniformity of run-length distribution across different gray intensities in the filtered image. High values ​​indicate uneven intensity distribution and heterogeneous texture.

[0064] wavelet-HLL_glrlm_LongRunEmphasis Image filter: Wavelet Transform (HLL); Feature category: Gray-level run-length matrix; Feature name: Long run-length emphasis; Meaning: Emphasizes long consecutive pixel runs. A high value indicates the presence of large, uniformly textured areas in the image.

[0065] wavelet-HLL_glrlm_LongRunHighGrayLevelEmphasis Image filter: Wavelet Transform (HLL); Feature category: Gray-level run-length matrix; Feature name: Long run, high gray-level emphasis; Meaning: Simultaneously emphasizes both "long" and "bright" continuous areas. High values ​​indicate the presence of large, bright, and uniform areas in the image.

[0066] wavelet-HLL_glrlm_LongRunLowGrayLevelEmphasis Image filter: Wavelet Transform (HLL); Feature category: Gray-level run-length matrix; Feature name: Long run, low gray-level emphasis; Meaning: Simultaneously emphasizes both "long" and "dark" continuous regions. High values ​​indicate the presence of large, uniformly dark areas in the image.

[0067] wavelet-HLL_glrlm_RunLengthNonUniformity Image filter: Wavelet transform (HLL); Feature category: Gray-level run-length matrix; Feature name: Run-length nonuniformity; Meaning: Calculates the degree of nonuniformity in the run-length distribution. A high value indicates large variations in run-length and heterogeneous texture.

[0068] wavelet-HLL_glrlm_RunVariance Image filter: Wavelet transform (HLL); Feature category: Gray-scale run matrix; Feature name: Run variance; Meaning: Calculates the variance of the run length, measuring the degree of variation in the run length.

[0069] wavelet-HLL_glszm_LargeAreaEmphasis Image filter: Wavelet Transform (HLL); Feature category: Gray-level region size matrix; Feature name: Large region emphasis; Meaning: Emphasizes large, continuous regions in the image. A high value indicates that the image consists of large, continuous regions.

[0070] wavelet-HLL_glszm_LargeAreaLowGrayLevelEmphasis Image filter: Wavelet Transform (HLL); Feature category: Gray-level region size matrix; Feature name: Large area low gray-level emphasis; Meaning: Simultaneously emphasizes "large" and "dark" continuous regions. High values ​​indicate the presence of large, dark areas in the image.

[0071] wavelet-HLL_glszm_SizeZoneNonUniformity Image filter: Wavelet transform (HLL); Feature category: Gray-scale region size matrix; Feature name: Region size non-uniformity; Meaning: Calculates the degree of non-uniformity in the distribution of continuous region sizes, reflecting heterogeneity.

[0072] wavelet-HLL_gldm_DependenceVariance Image filter: Wavelet transform (HLL); Feature category: Gray-level dependency matrix; Feature name: Dependency variance; Meaning: Calculates the variance of gray-level dependency, measuring the heterogeneity of texture.

[0073] wavelet-HLH_firstorder_Kurtosis Image Filter: Wavelet Transform (HLH: High-pass filtering on the X-axis, low-pass filtering on the Y-axis, and high-pass filtering on the Z-axis); Feature Category: First-order Statistics; Feature Name: Kurtosis; Meaning: Calculates the steepness of the pixel intensity distribution on the filtered image.

[0074] wavelet-HLH_firstorder_Energy Image Filter: Wavelet Transform (HLH); Feature Category: First-order Statistics; Feature Name: Energy; Meaning: Calculates the sum of squares of pixel intensities on the filtered image.

[0075] wavelet-HLH_glrlm_GrayLevelNonUniformity Image filter: Wavelet transform (HLH); Feature category: Gray-level run matrix; Feature name: Gray-level non-uniformity; Meaning: Calculates the uniformity of run length distribution across different gray intensities, reflecting intensity heterogeneity.

[0076] wavelet-HLH_glrlm_LongRunEmphasis Image filter: Wavelet transform (HLH); Feature category: Gray-level run-length matrix; Feature name: Long run-length emphasis; Meaning: Emphasizes long consecutive pixel runs, high values ​​indicate coarse texture.

[0077] wavelet-HLH_glrlm_LongRunHighGrayLevelEmphasis Image filter: Wavelet transform (HLH); Feature category: Gray-level run-length matrix; Feature name: Long run-length high gray-level emphasis; Meaning: Simultaneously emphasizes "long" and "bright" continuous areas.

[0078] wavelet-HLH_glrlm_RunLengthNonUniformity Image filter: Wavelet transform (HLH); Feature category: Gray-level run-length matrix; Feature name: Run-length nonuniformity; Meaning: Calculates the degree of nonuniformity in the run-length distribution, reflecting texture heterogeneity.

[0079] wavelet-HHH_glrlm_RunVariance Image Filter: Wavelet Transform (HHH: High-pass filtering on the X, Y, and Z axes to emphasize high-frequency details); Feature Category: Gray-scale run-length matrix; Feature Name: Run-length variance; Meaning: Calculates the variance of the run length on an image that emphasizes details.

[0080] wavelet2-LHL_firstorder_Maximum Image Filter: Wavelet Transform (LHL, possibly referring to another generation of wavelet decomposition or a specific implementation); Feature Category: First-order Statistics; Feature Name: Maximum Value; Meaning: Find the brightest pixel intensity value in the filtered image.

[0081] wavelet2-LHL_firstorder_Skewness Image Filter: Wavelet Transform (LHL); Feature Category: First-order Statistics; Feature Name: Skewness; Meaning: Calculates the asymmetry of pixel intensity distribution on the filtered image.

[0082] wavelet2-LHL_firstorder_Energy Image filter: Wavelet transform (LHL); Feature category: First-order statistics; Feature name: Energy; Meaning: Calculates the sum of squares of pixel intensities on the filtered image.

[0083] wavelet2-LHL_glszm_SizeZoneNonUniformity Image filter: Wavelet transform (LHL); Feature category: Gray-scale region size matrix; Feature name: Region size non-uniformity; Meaning: Calculates the degree of non-uniformity in the distribution of continuous region sizes, reflecting heterogeneity.

[0084] wavelet2-LHH_glrlm_RunVariance Image filter: Wavelet transform (LHH); Feature category: Gray-scale run-length matrix; Feature name: Run-length variance; Meaning: Calculates the variance of the run length, measuring the degree of variation in the run length.

[0085] wavelet2-HLL_firstorder_Skewness Image Filter: Wavelet Transform (HLL); Feature Category: First-order Statistics; Feature Name: Skewness; Meaning: Calculates the asymmetry of pixel intensity distribution on the filtered image.

[0086] wavelet2-HLL_firstorder_Energy Image Filter: Wavelet Transform (HLL); Feature Category: First-order Statistics; Feature Name: Energy; Meaning: Calculates the sum of squares of pixel intensities on the filtered image.

[0087] wavelet2-HLH_firstorder_Skewness Image Filter: Wavelet Transform (HLH); Feature Category: First-order Statistics; Feature Name: Skewness; Meaning: Calculates the asymmetry of pixel intensity distribution on the filtered image.

[0088] wavelet2-HLH_firstorder_Kurtosis Image Filter: Wavelet Transform (HLH); Feature Category: First-order Statistics; Feature Name: Kurtosis; Meaning: Calculates the steepness of the pixel intensity distribution on the filtered image.

[0089] wavelet2-LLL_firstorder_Mean Image filter: Wavelet transform (LLL: low-pass filtering on the X, Y, and Z axes, approximating the original image); Feature category: First-order statistics; Feature name: Mean; Meaning: Calculates the average intensity of all pixels on a smoothed version that approximates the original image.

[0090] wavelet2-LLL_firstorder_Median Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Median; Meaning: Find the median of pixel intensity values ​​on a smooth image.

[0091] wavelet2-LLL_firstorder_Minimum Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Minimum; Meaning: Find the darkest pixel intensity value on a smooth image.

[0092] wavelet2-LLL_firstorder_Maximum Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Maximum value; Meaning: Find the brightest pixel intensity value on a smooth image.

[0093] wavelet2-LLL_firstorder_Variance Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Variance; Meaning: Calculates the degree to which pixel intensity values ​​deviate from the mean on a smooth image, measuring dispersion.

[0094] wavelet2-LLL_firstorder_Skewness Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Skewness; Meaning: Calculates the asymmetry of pixel intensity distribution on a smooth image.

[0095] wavelet2-LLL_firstorder_Energy Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Energy; Meaning: Calculates the sum of squares of pixel intensities on a smooth image.

[0096] wavelet2-LLL_firstorder_RootMeanSquared Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Root mean square; Meaning: Calculates the square root of the average of the squares of pixel intensity values ​​on a smooth image.

[0097] wavelet2-LLL_firstorder_RobustMeanAbsoluteDeviation Image filter: Wavelet Transform (LLL); Feature category: First-order statistics; Feature name: Robust Mean Absolute Deviation; Meaning: Calculates the average of the median absolute values ​​of the deviations of all pixel intensity values ​​from the median on a smoothed image. Insensitive to outliers.

[0098] wavelet2-LLL_firstorder_InterquartileRange Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Interquartile range; Meaning: Calculates the difference between the 75th percentile and the 25th percentile on a smooth image, reflecting the dispersion of the middle 50% of the data.

[0099] wavelet2-LLL_firstorder_MeanAbsoluteDeviation Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: Mean absolute deviation; Meaning: Calculates the average of the absolute values ​​of the deviations of all pixel intensity values ​​from the mean value on a smooth image.

[0100] wavelet2-LLL_firstorder_90Percentile Image filter: Wavelet transform (LLL); Feature category: First-order statistics; Feature name: 90th percentile; Meaning: In a smooth image, find a pixel value that is lower than 90% of the pixel values.

[0101] wavelet2-LLL_glcm_Autocorrelation Image filter: Wavelet transform (LLL); Feature category: Gray-level co-occurrence matrix; Feature name: Autocorrelation; Meaning: Measures the fineness and coarseness of image texture. High values ​​indicate coarse texture with periodic patterns.

[0102] wavelet2-LLL_glcm_ClusterProminence Image filter: Wavelet transform (LLL); Feature category: Gray-level co-occurrence matrix; Feature name: Cluster prominence; Meaning: Measures the asymmetry and skewness of the gray-level co-occurrence matrix. A high value indicates matrix asymmetry and asymmetrical gray-level distribution in the image.

[0103] wavelet2-LLL_glcm_JointEntropy Image filter: Wavelet transform (LLL); Feature category: Gray-level co-occurrence matrix; Feature name: Joint entropy; Meaning: Measures the randomness and complexity of image texture. A high value indicates that the intensity combinations between pixel pairs are very random, and the texture is complex.

[0104] wavelet2-LLL_glrlm_GrayLevelNonUniformity Image filter: Wavelet transform (LLL); Feature category: Gray-level run length matrix; Feature name: Gray-level non-uniformity; Meaning: Calculates the uniformity of run length distribution across different gray intensities in a smooth image, reflecting intensity heterogeneity.

[0105] wavelet2-LLL_glrlm_HighGrayLevelRunEmphasis Image filter: Wavelet Transform (LLL); Feature category: Gray-level run-length matrix; Feature name: High gray-level run-length emphasis; Meaning: Emphasizes bright consecutive pixel runs. High values ​​indicate the presence of bright areas in the image.

[0106] wavelet2-LLL_glrlm_RunLengthNonUniformity Image filter: Wavelet transform (LLL); Feature category: Gray-level run-length matrix; Feature name: Run-length nonuniformity; Meaning: Calculates the degree of nonuniformity in the run-length distribution on a smooth image.

[0107] gradient_firstorder_Maximum Image Filter: Gradient Filter (Edge Emphasis); Feature Category: First-Order Statistics; Feature Name: Maximum Value; Meaning: Find the brightest pixel intensity value (i.e., the most significant edge intensity) on an image with prominent edges.

[0108] gradient_firstorder_Range Image filter: Gradient filter; Feature category: First-order statistics; Feature name: Range; Meaning: Calculates the intensity difference between the strongest and weakest edges on the gradient image.

[0109] gradient_firstorder_Energy Image filter: Gradient filter; Feature category: First-order statistics; Feature name: Energy; Meaning: Calculates the sum of squares of all pixel values ​​on the gradient image. A high value indicates a higher overall edge intensity or a concentrated distribution.

[0110] gradient_glszm_LargeAreaEmphasis Image Filter: Gradient Filter; Feature Category: Gray-level Region Size Matrix; Feature Name: Large Region Emphasis; Meaning: Emphasizes large, continuous regions in a gradient image. High values ​​indicate the presence of large areas with similar edge characteristics in the image.

[0111] gradient_glszm_LargeAreaLowGrayLevelEmphasis Image Filter: Gradient Filter; Feature Category: Gray-Level Region Size Matrix; Feature Name: Large Region Low Gray-Level Emphasis; Meaning: In a gradient image, simultaneously emphasizes continuous regions with both "large" and "low edge intensity." A high value indicates the presence of large, uniform regions with indistinct edges in the image.

[0112] 4. Model Building and Performance Validation Classification Model (Risk Stratification): Six models were compared, with TabPFN (pre-trained tabular base model) as the core. Input: 28 clinical features and 59 radiomics features were fused. Training: No additional parameter tuning was required; fine-tuning was performed directly using pre-trained weights. Stratified sampling was used to ensure class balance between the training and test sets. Performance: An AUC of 0.89 (sensitivity 79.2%, specificity 78.3%) was achieved on the test set. The specificity decreased by only 8.1% compared to the training set (far lower than XGBoost's 30.4%). The calibration curve showed that the predicted probability was highly consistent with the actual bleeding risk (Figure 3).

[0113] The regression model of this invention (quantitative prediction of bleeding volume): Constructing a regression task based on TabPFN: Input: Multimodal features of the same classification model; Performance: Test set R 2 The residuals were 0.51, RMSE 0.642, and uniformly distributed (Figure 5). Since all data underwent natural logarithmic processing, this figure only shows a qualitative comparison of different models. However, based on actual prospective clinical prediction results, the prediction performance is excellent, and the error is within a clinically acceptable range. Multimodal prediction workflow in small sample scenarios: The TabPFN model was applied to the multimodal feature (clinical + radiomics) fusion task of placenta previa. By optimizing the feature input format (such as standardized alignment of clinical variables and radiomics features) and model inference parameters (such as batch processing window size), it simultaneously achieves bleeding risk stratification (≥1000 mL) and quantitative prediction of bleeding volume in small samples (n≤300). This solves the problem of overfitting in traditional models with small samples and high-dimensional multimodal data. Furthermore, this workflow is specifically designed for placenta previa diagnosis and treatment scenarios and is irreplaceable. The "stratified + quantitative" dual-output clinical decision-making system integrates the above-mentioned segmentation, feature fusion and prediction models to form a complete preoperative assessment process. It can simultaneously output the risk stratification result of "whether it is severe postpartum hemorrhage (≥1000 mL)" and the specific blood loss estimate, which directly supports the allocation of clinical resources (such as blood product reserves and priority of interventional equipment preparation). This dual-output function and its connection logic with clinical decision-making are innovative points not covered by existing technologies.

[0114] It provides a convenient end-to-end training process for small sample scenarios (less than 500 cases) and offers a simple fine-tuning training process for single-center scenarios lacking experienced doctors.

[0115] Based on the previous validation, doctors can automatically segment the placenta's weights using the finely tuned nnUnet we provide, skipping the feature selection step, and directly put the collected or extracted clinical or radiomics features into TabPFN for fine-tuning to adapt to specific single-center application scenarios.

[0116] 5. Prospective clinical validation The model's applicability was validated in 6 prospective patients (Figure 5): Predictive bleeding risk (classification) effectiveness: Low-risk cases (2 cases): ultrasound PAS score 2-3, model predicted bleeding probability 0.31-0.36, routine cesarean section was performed clinically, actual blood loss was 400-600 mL, no blood transfusion required; High-risk cases (2 cases): One case had a history of two cesarean sections and MRI showed placenta accreta, with a model prediction probability of 0.82. Preoperative preparation included an internal iliac artery balloon and autologous blood. Actual blood loss was 2330 mL, which was successfully managed according to the contingency plan. The other case was conceived through IVF and had preoperative anemia, with a model prediction probability of 0.71, indicating high risk. Actual blood loss was 1500 mL, and 4 units of blood were transfused in time. Borderline risk cases (2 cases): predicted probability 0.55-0.61, imaging suggested suspicious implantation, clinical placement of Bakri balloon for prevention, actual bleeding 600-800 mL, avoid excessive intervention.

[0117] Quantitative prediction of blood loss (regression) effect (Figure 6): The error range between the regression prediction results and the actual values ​​in the six cases was (43.2-372.4 mL), which is within the clinically acceptable range. Using 1000 mL as the clinically recognized definition of PPH as the intervention threshold, perfect intervention can be achieved, and corresponding medical intervention can be carried out according to the predicted bleeding volume to avoid waste of resources.

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for predicting the risk of postpartum hemorrhage in placenta previa, characterized in that, include: The radiomics feature acquisition unit is used to acquire MRI radiomics features of patients with postpartum hemorrhage due to placenta previa. Clinical feature acquisition unit, used to acquire clinical features of patients with postpartum hemorrhage due to placenta previa; Subset construction unit, used to construct feature subsets based on clinical features and MRI radiomics features; The prediction unit is used to input the feature subset into the TabPFN model to predict the probability of postpartum hemorrhage and the amount of bleeding.

2. The postpartum hemorrhage risk prediction system for placenta previa according to claim 1, characterized in that, The feature subset was obtained through multimodal feature screening based on clinical features and MRI radiomics features; The MRI radiomics features were obtained by extracting features using PyRadiomics v3.0.1 based on an automatically segmented placental mask.

3. The postpartum hemorrhage risk prediction system for placenta previa according to claim 2, characterized in that, The clinical features were extracted from electronic medical records and included demographic data, obstetric history, current pregnancy parameters, and laboratory indicators.

4. The postpartum hemorrhage risk prediction system for placenta previa according to claim 3, characterized in that, The method for obtaining the automatically segmented placental mask is as follows: A 7-level encoder-decoder architecture was used to automatically segment MRI images to obtain a placental mask; The channel width of the 7-level encoder-decoder is [32, 64, 128, 256, 320, 320, 320], the convolution kernel gradually transitions from anisotropic [1,3,3] to isotropic [3,3,3], and the downsampling step size is adjusted accordingly from [1,1,1] to [1,2,2].

5. A postpartum hemorrhage risk prediction system for placenta previa according to claim 4, characterized in that, The feature subset was obtained by combining and filtering four methods: ① Correlation analysis (CORR): Calculate the Pearson correlation coefficient between the characteristic and the amount of bleeding; ② LASSO Regression: The regularization strength is determined by 5-fold cross-validation, and features with non-zero coefficients are selected; ③ Random Forest (RF): Based on Gini importance ranking of 200 trees; ④ Recursive Feature Elimination (RFE): Using a random forest as the base classifier, the 10% least important features are gradually removed; Each method was repeated 10 times, and Gaussian noise was added to enhance stability. The top-m features that appeared frequently were retained. Finally, the union of the screening results of the four methods was taken to obtain a feature subset including 28 clinical features and 59 radiomics features.

6. The postpartum hemorrhage risk prediction system for placenta previa according to claim 1, characterized in that, The TabPFN model requires no additional parameter tuning, utilizes pre-trained weights for fine-tuning, and employs stratified sampling to ensure class balance in the training / test sets.

7. The postpartum hemorrhage risk prediction system for placenta previa according to claim 1, characterized in that, The placenta previa postpartum hemorrhage risk prediction system is applied to scenarios with fewer than 500 samples.

8. The postpartum hemorrhage risk prediction system for placenta previa according to claim 1, characterized in that, The placenta previa postpartum hemorrhage risk prediction system is applied to scenarios with fewer than 300 samples.

9. A method for constructing a postpartum hemorrhage risk prediction system for placenta previa according to claim 1, comprising: To obtain MRI radiomics characteristics of patients with postpartum hemorrhage due to placenta previa; To obtain the clinical characteristics of patients with postpartum hemorrhage due to placenta previa; Construct a feature subset based on clinical features and MRI radiomics features; The feature subset is input into the TabPFN model to predict the probability of postpartum hemorrhage and the amount of bleeding.

Images