Cerebral hemorrhage hematoma enlargement prediction system and method based on deep learning

Abstract

The invention provides a cerebral hemorrhage hematoma enlargement prediction system and method based on deep learning, and the system comprises a data obtaining module which is used for obtaining a baseline CT image and clinical data of a patient; the image preprocessing module is used for carrying out standardization processing on the CT image; the feature extraction module is used for extracting image features from the CT image by adopting a deep convolutional neural network; the feature fusion module is used for fusing the image features, the clinical features and the radiology features; and the prediction module is used for predicting the hematoma expansion risk based on the fused features. The system captures 3D space information to the maximum extent by extracting the maximum lesion level, shows high reliability, high interpretability and clinical availability in the aspect of predicting hematoma expansion, is remarkably superior to prediction of clinicians, achieves reasonable consistency of prediction probability and actual probability on multiple data sets, and has good application prospects. The system can be used as a clinical decision support system and has potential to improve patient prognosis.

Description

Technical Field

[0001] The present invention relates to the technical field of the integration of artificial intelligence and biomedicine, and particularly to a system and method for predicting hematoma expansion in cerebral hemorrhage based on deep learning. Background Art

[0002] Intracerebral hemorrhage (ICH) is a severe type of stroke characterized by bleeding within the brain. Hematoma expansion (HE) is an important treatment goal and surrogate endpoint for acute ICH. Currently, clinical methods for identifying HE rely primarily on physician judgment and assessment based on clinical scoring systems, which are subject to high subjectivity and limited accuracy. Although some studies have attempted to use 2D deep learning techniques to predict the risk of HE, these models have several limitations: they do not consider the revised definition of hematoma expansion (RHE), lack multicenter validation, and have limited model interpretability. Identifying patients most at risk for HE is crucial for improving the assessment of treatment effects in future trials. Consequently, researchers worldwide have devoted significant research efforts to identifying predictors of HE. In addition to age, Glasgow Coma Scale (GCS), time from onset to CT scan, subarachnoid hemorrhage, hematoma volume, and location, signs of ICH on unenhanced CT have also been identified as predictors of HE. Various prediction scores have been developed based on these factors, but there is currently no clear consensus on which tool has the best predictive power. Furthermore, whether these scores can accurately predict RHE remains unknown. The lack of reliable predictors of hematoma expansion hinders physicians' early detection and subsequent aggressive intervention. Therefore, prospectively identifying patients at high risk for hematoma expansion remains a priority.

[0003] In the field of medical image analysis, several related technologies have been proposed. For example, CN101916443A discloses a CT image processing system for identifying abnormal signals in the subarachnoid space. However, its technical solution is primarily based on traditional image processing algorithms and does not involve deep learning techniques. CN112700451A discloses a method for automatically segmenting 3D medical images, using a two-stage segmentation strategy, but its application scenario differs from the present invention.

[0004] The above-mentioned record of background technology knowledge is intended to help ordinary technicians in this field understand the existing technology that is relatively close to the present invention, and at the same time facilitate the understanding of the inventive concept and technical solution of the present invention. It should be clear that in the absence of clear evidence that the above-mentioned content has been disclosed before the filing date of this patent application, the above-mentioned background technology should not be used to evaluate the novelty of the technical solution of this application. Summary of the Invention

[0005] Technical issues In order to solve the above problems, the purpose of the present invention is to provide a deep learning-based system and method for predicting hematoma expansion in cerebral hemorrhage. The system integrates the capabilities of 2D transfer learning and maximizes the capture of 3D spatial information by extracting the largest level of the lesion. It shows strong reliability, high interpretability and clinical usability in predicting hematoma expansion, which is significantly better than the predictions of clinicians. The predicted probability and actual probability have achieved reasonable consistency in multiple data sets, providing a visual explanation for the decision-making process. It can be used as a clinical decision support system, timely intervention and transfer to intensive care units for close monitoring and active management, and has the potential to improve patient prognosis.

[0006] Technical Solution That is, the present invention includes the following technical solutions.

[0007] (1) A deep learning-based system for predicting hematoma expansion in cerebral hemorrhage, comprising: A data acquisition module, used to obtain baseline CT images and clinical data of patients; An image preprocessing module, used for performing standardization processing on the CT image; A feature extraction module, which uses a 2.5D deep convolutional neural network to extract image features from the CT image; a feature fusion module for fusing the image features with clinical features and radiological features extracted from clinical data; A prediction module predicts the risk of hematoma expansion based on the fused features, wherein the hematoma expansion includes ventricular hemorrhage.

[0008] Compared with pure 3D models, this application uses a 2.5D architecture to reduce computational complexity by approximately 40-60% while maintaining spatial information coherence, which enables the system to achieve real-time reasoning (3-5 seconds per case) on conventional clinical workstations. This balance is particularly critical in emergency scenarios because the prediction of hematoma expansion needs to be completed in a very short time. Anisotropic processing (initial 2D convolution to process high-resolution cross-sections, and later convolution to fuse multi-section information) enables the model to adapt to the differences in layer thickness (1-5mm) of CT scanners in different hospitals, thereby improving clinical applicability. The traditional definition of hematoma expansion (CHE) does not take into account the important factor of ventricular hemorrhage expansion, while the revised definition of hematoma expansion (RHE) takes ventricular hemorrhage expansion into consideration, showing superior predictive ability.

[0009] In some specific implementation schemes, the processing performed by the image preprocessing module includes: Limit CT values to the range of 45-100HU; A fixed-resolution resampling method was used to unify the voxel spacing to 1 mm × 1 mm × 1 mm; Extract the largest cross-section of the lesion area and its adjacent slices to form the 2.5D input; Use min-max transformation to adjust the value range to [-1,1]; Resize the image to 224×224.

[0010] In some specific implementations, the feature extraction module utilizes a 2.5D deep convolutional neural network with a ResNet101 backbone network. Input is five consecutive CT slices, including the slice along the maximum lesion area and two adjacent slices above and below it. A parallel backbone network extracts a 2048-dimensional feature vector, which is then subjected to cross-slice max pooling to yield a single feature vector. This cross-slice max pooling operation essentially captures the growth pattern characteristics of the hematoma in three-dimensional space and enables the network to learn the dynamic temporal characteristics of hematoma evolution. This pattern is more sensitive to small hematomas (<6 ml), resulting in a system prediction AUC of 0.882 for RHE1 on the YY dataset, with a peak sensitivity of 0.956.

[0011] In some specific implementation plans, the clinical characteristics include age, gender, time from onset to CT scan, and GCS score at admission; the radiological characteristics include intracerebral hemorrhage and ventricular hemorrhage volume, cerebral hemorrhage location, cerebral hemorrhage signs, and subarachnoid hemorrhage.

[0012] In some specific embodiments, the prediction results output by the prediction module include the following three defined risks of hematoma expansion: RHE1: intracerebral hemorrhage >6 ml or increased by 33% or any expansion of ventricular hemorrhage; RHE2: intracerebral hemorrhage > 6 ml or increased by 33% or intraventricular hemorrhage > 1 ml; CHE: Parenchymal hemorrhage >6 mL increases the risk of hematoma expansion by 33%. The aforementioned CHE considers traditional hematoma expansion, while RHE1 and RHE2 both account for ventricular hemorrhage. These three are not three categories of the model, but rather three outcomes. Parenchymal hemorrhage >6 mL meets the three aforementioned hematoma expansion risks. It should be noted that RHE1 / RHE2 have superior predictive contribution in cases where only ventricular hemorrhage expands.

[0013] In some specific implementation schemes, the 2.5D deep convolutional neural network of the feature extraction module adopts the following strategy during training: (1) Randomly select the central slice of the 2.5D input from the first three slices with the largest lesion area; (2) Apply a random translation within 10% and a random scaling within the range of [0.9, 1.1] to the ROI; (3) The cropped ROI is mirror-flipped horizontally and vertically with a probability of 50%.

[0014] In some specific embodiments, the learning rate of the 2.5D deep convolutional neural network backbone is based on the minimum learning rate and the maximum learning rate Sure, Set to 0, Set to 0.01; The learning rate for the task-specific part is calculated as follows:

[0015] The learning rate of the backbone is determined as follows:

[0016] in, Indicates the number of iteration cycles; Indicates the current iteration number.

[0017] In some specific implementation schemes, the system also includes a visual explanation module for generating visual explanations of model decisions. In some specific implementation schemes, the visual explanation module uses gradient-weighted class activation mapping technology to generate an attention heat map that shows the image areas that the model focuses on when making predictions.

[0018] (2) A method for predicting hematoma expansion of cerebral hemorrhage based on deep learning for diagnosis and treatment of non-diseases, the method being based on the aforementioned system and comprising the following steps: Obtain the patient's baseline CT images and clinical data; performing standardized preprocessing on the CT image; Use 2.5D deep convolutional neural network to extract image features from preprocessed CT images; fusing the imaging features with clinical features; Predict the risk of hematoma expansion based on post-fusion features; Generate visual explanations of model decisions.

[0019] In some specific implementations, the following data augmentation strategy is implemented during the training of a 2.5D deep convolutional neural network: the central slice of the 2.5D input is randomly selected from the first three slices with the largest lesion area; a random translation within 10% and a random scaling within the range of [0.9, 1.1] are applied to the ROI; and the cropped ROI is mirrored horizontally and vertically with a probability of 50%. This random selection of slices, random processing of the ROI, and flipping with a certain probability can significantly avoid overfitting of the deep convolutional neural network model during training for all tasks, thereby improving the model's generalization ability, its prediction accuracy, and enhancing the model's interpretability.

[0020] (3) A computer device comprising a memory, a processor, a communication interface, and a communication bus; wherein the memory, the processor, and the communication interface communicate with each other via the communication bus; the memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory, and when the processor executes the computer program, it implements at least one step of the aforementioned method for predicting hematoma expansion of cerebral hemorrhage based on deep learning for diagnosis and treatment of non-diseases.

[0021] (4) A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, at least one step of the above-mentioned method for predicting hematoma expansion of cerebral hemorrhage based on deep learning for diagnosis and treatment of non-diseases is implemented.

[0022] On the basis of conforming to the common sense in this field, the above-mentioned preferred conditions can be combined with each other to obtain a specific implementation method.

[0023] Beneficial effects This study introduces a novel 2.5D deep learning model that combines strong performance, high interpretability, and clinical usability. These advances have played a key role in improving HE risk prediction and helping to understand the mechanisms by which deep learning models predict HE. Given the results of this study, we believe that the HENet model will have a positive impact on clinical practice: deep learning technology can be integrated into electronic medical records and picture archiving and communication systems (PACS) for use as a clinical decision support system. HENet accurately identifies patients at risk for hematoma expansion, enabling timely intervention and transfer to intensive care units for close monitoring and active management. Deep learning-enhanced clinician decision-making has the potential to improve outcomes for patients with ICH.

[0024] Compared to pure 3D models, the 2.5D architecture reduces computational complexity by approximately 40-60% while maintaining spatial information coherence. This enables the model to achieve real-time inference (3-5 seconds per case) on conventional clinical workstations (such as workstations equipped with NVIDIA RTX 50 series). This balance is particularly critical in emergency scenarios, as the prediction of hematoma expansion needs to be completed in a very short time. Anisotropic processing (initial 2D convolution processes high-resolution cross-sections, and later convolution fuses multi-section information) enables the model to adapt to the differences in layer thickness (1-5mm) of CT scanners in different hospitals, improving clinical applicability.

[0025] The system implicitly achieves multi-scale feature fusion, from local signs (such as island signs) to global features (such as overall hematoma volume changes). This fusion enhances the model's sensitivity to morphological features, resulting in prediction accuracy superior to physician interpretation. The combination of clinical features (such as GCS score) and deep learning features produces a super-additive effect: on the YY dataset, the combined features achieved an IDI value of 0.36, significantly higher than that of single-modality features (p < 0.001), indicating that clinical data fills the gaps in imaging features in certain pathophysiological dimensions. The system significantly outperformed physicians' predictions in terms of the net reclassification index (NRI) and integrated discrimination index (IDI) for RHE1 and RHE2 outcomes.

[0026] The feature representations learned by the model demonstrated cross-center knowledge transfer capabilities. In the migration from the HS dataset (training set) to the YY and LA datasets (external validation sets), the model performance decreased by less than 8%, significantly lower than the 15-20% achieved by traditional machine learning methods. This transferability demonstrates that the model is able to capture universal pathological features that are device-independent, laying the foundation for the development of a universal hematoma expansion prediction model. The model's strong focus on the "low-density band" feature (37% contribution) suggests that it may be a novel imaging biomarker for hematoma expansion. This finding is highly consistent with the "avalanche" pathology model and provides a new perspective for studying the pathological mechanisms of hematoma expansion.

[0027] The present invention adopts the above technical solution to achieve the above purpose, which makes up for the shortcomings of the existing technology and has reasonable design and convenient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to make the above-mentioned and / or other purposes, features, advantages and examples of the present invention more obvious and easy to understand, the following is a brief introduction to the drawings required for use in the specific embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 Schematic diagram of the HENet model architecture of the present invention.

[0030] Figure 2 Receiver operating characteristic curves of HENet model, 2D model and doctors for predicting hematoma expansion; (Figures (A), (B) and (C) correspond to the receiver operating characteristic (ROC) curves of HENet model, 2D model and doctors for predicting RHE1, RHE2 and CHE on the YY dataset, respectively. Figures (D), (E) and (F) correspond to the ROC curves of HENet model, 2D model and doctors for predicting RHE1, RHE2 and CHE on the LA dataset, respectively).

[0031] Figure 3 Figures (A), (C) and (E) represent the calibration curves of the HENet model, 2D model and doctors for predicting hematoma expansion on the YY dataset, respectively. Figures (B), (D) and (F) represent the calibration curves of the HENet model, 2D model and doctors for predicting RHE1, RHE2 and CHE on the LA dataset, respectively. HL represents the Hosmer-Lemeshow Goodness-of-Fit Test).

[0032] Figure 4 Decision curve analysis diagram for predicting hematoma expansion; (Figures (A), (B) and (C) represent the decision curves of the HENet model, 2D model and doctors for predicting RHE1, RHE2 and CHE on the YY dataset, respectively. Figures (D), (E) and (F) represent the decision curves of the HENet model, 2D model and doctors for predicting RHE1, RHE2 and CHE on the LA dataset, respectively).

[0033] Figure 5 Figure 2. Model interpretability of patients with left basal ganglia intracerebral hemorrhage compared with Grad-CAM.

[0034] Figure 6 Figure 3. Interpretability of the model with Grad-CAM in patients with left basal ganglia hemorrhage and extended intraventricular hemorrhage. DETAILED DESCRIPTION

[0035] Those skilled in the art may refer to the contents herein and appropriately substitute and / or modify the process parameters to achieve the desired effect. However, it should be noted that all such substitutions and / or modifications are obvious to those skilled in the art and are considered to be included in the present invention. The products and preparation methods described herein have been described through preferred embodiments. It is obvious that those skilled in the art can modify or appropriately change and combine the products and preparation methods described herein to implement and apply the technology of the present invention without departing from the content, spirit, and scope of the present invention.

[0036] Unless otherwise defined, technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The present invention utilizes the methods and materials described herein; however, other suitable methods and materials known in the art may also be used. The materials, methods, and examples described herein are illustrative only and are not intended to be limiting. All publications, patent applications, patents, provisional applications, database entries, and other references mentioned herein are incorporated herein by reference in their entirety. In the event of a conflict, the present specification, including definitions, will control.

[0037] Unless otherwise specified, the materials, methods, and examples described herein are illustrative only and not limiting. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, suitable methods and materials are described herein.

[0038] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. At the same time, the embodiments in this application and the features in the embodiments can be combined with each other unless there is a conflict.

[0039] It should be understood that any technical solution claimed for protection in the present invention does not involve the diagnosis and treatment of diseases.

[0040] To facilitate understanding of the embodiments of the present invention, abbreviations and key terms that may be involved in the embodiments of the present invention are first explained or defined. Undefined abbreviations or key terms are generally understood by those skilled in the art.

[0041] NIfTI file: Neuroimaging Informatics Technology Initiative file, a file format used to store neuroimaging data, the full name of which is Neuroimaging Informatics Technology Initiative file.

[0042] NCCT sign: Non-Contrast Computed Tomography, refers to the non-enhanced CT sign.

[0043] Grad-CAM: Gradient-weighted Class Activation Mapping, a visualization technique for explaining the decision-making process of convolutional neural networks (CNNs).

[0044] Conv: Convolution, convolution.

[0045] GCS: Glasgow Coma Scale.

[0046] ICH: Intracerebral Hemorrhage.

[0047] IVH: Intraventricular Hemorrhage, Intraventricular Extension of IntracerebralHemorrhage.

[0048] SAH: Subarachnoid Hemorrhage.

[0049] DICOM: Digital Imaging and Communications in Medicine In addition, the experimental methods used in the examples are conventional methods unless otherwise specified. The materials and reagents used are commercially available unless otherwise specified. Reagents or instruments used without manufacturer indication are commercially available conventional products. All publications and other references cited herein are incorporated by reference in their entirety.

[0050] The present invention is described in detail below Example 1: A deep learning model, system and method for predicting hematoma expansion in cerebral hemorrhage is provided, which is identified as a HENet model. The HENet model architecture diagram is as follows Figure 1 As shown, the model combines the capabilities of 2D transfer learning and maximizes the capture of 3D spatial information by extracting the largest aspect of the lesion. The model integrates clinical features in an end-to-end manner, and its performance has been verified in two external validation cohorts, including discrimination ability, calibration curves, and clinical utility, achieving satisfactory results in predicting revised hematoma expansion (RHE1 and RHE2) and traditional hematoma expansion (CHE). In terms of model interpretability, the HENet model has superior interpretability, using Grad-CAM technology to visualize the contribution of different regions in the input image to HE prediction. This provides valuable insights into the model's decision-making process and enhances our understanding of its functionality without the need for complex modifications or retraining of the model. The details are as follows.

[0051] Patient data Data from patients with intracerebral hemorrhage were retrospectively collected and used to build and test deep learning models. The datasets included a training set (n=522) from three different hospitals: HS Hospital (February 2014 to December 2019; HS dataset), and two independent external testing datasets collected from YY Hospital (June 2017 to June 2021; YY dataset; n=100) and LA Hospital (August 2022 to August 2023; LA dataset; n=96), respectively. Due to the retrospective nature of the study, informed consent was waived. Eligible patients were those who underwent baseline (<6 hours after clinical onset) and 24-hour follow-up CT scans. Patient exclusion criteria were as follows: (a) surgery before follow-up CT scan and (b) incomplete examination or severe artifacts.

[0052] CT scans were obtained using six different CT scanners from four different vendors. Detailed information regarding CT scan acquisition can be found in Table 1. Manual segmentation was used to measure intraparenchymal hemorrhage volume and intraventricular hemorrhage volume. NCCT landmarks were also manually annotated. Further details regarding lesion segmentation and image interpretation are provided in the Supplementary Methods. Clinical data including age, sex, time interval from symptom onset to baseline CT scan, and Glasgow Coma Scale score on admission were collected.

[0053] Table 1. CT scan parameters of HS, YY and LA datasets

[0054] Doctor's interpretation Two physicians independently estimated the risk of three different definitions of hematoma expansion (RHE1, RHE2, and CHE) based on baseline clinical and imaging data presented in this study. RHE1: intraparenchymal hemorrhage >6 ml or a 33% increase or any expansion of intraventricular hemorrhage; RHE2: intraparenchymal hemorrhage >6 ml or a 33% increase or intraventricular hemorrhage >1 ml; and CHE: intraparenchymal hemorrhage >6 ml or a 33% increase.

[0055] Image interpretation Two radiologists evaluated nine non-contrast computed tomography (NCCT) signs of intracerebral hemorrhage based on baseline CT images. These signs included density (inhomogeneous density, swirl sign, hypodensity, black hole sign, mixed sign, fluid level) and shape (irregular shape, island sign, satellite sign). Two radiologists (XZ and HX) independently interpreted the NCCT signs without access to the patients' clinical data or follow-up CT information. During the analysis, the diagnostic criteria and illustrative cases of the NCCT signs were used as a reference template. Consensus between the two investigators determined the presence or absence of each ICH sign. In case of disagreement, a third radiologist (XF) independently reviewed the images, and the majority opinion prevailed.

[0056] Lesion segmentation Baseline and follow-up CT images were randomized and converted to NIfTI files before segmentation. Two experienced radiologists (XZ and HX) with 8 and 20 years of experience in neuroradiology performed lesion segmentation for parenchymal hemorrhage and hemoventricular effusion using ITK-SNAP software (version 3.8.0). Parenchymal hemorrhage and hemoventricular effusion lesions were manually delineated slice by slice along the lesion margins on the CT images. After segmentation, individual slices were reconstructed into a 3D representation of each region to calculate the volume of parenchymal hemorrhage and hemoventricular effusion. A senior radiologist (LS) confirmed all segmentations, and any disagreements were resolved by consensus.

[0057] Doctor's prediction Two physicians (JS and ZZ) with 5 and 20 years of neuroradiology experience, respectively, independently estimated the risk of three different definitions of hematoma expansion (RHE1, RHE2, CHE) for patients based on the clinical and imaging data presented in this study. They were instructed to rate the probability of hematoma expansion on a scale of 0 to 100 (where 0 indicates the absolute absence of hematoma expansion and 100 indicates definite progression of hematoma expansion). The reliability of the two physicians' continuous quantitative assessment of the risk of hematoma expansion (i.e., risk from 0 to 100) was examined using the intraclass correlation coefficient (ICC). To assess the discriminatory ability of the physicians' predictions, the mean of their independent predictions was compared with the predictive ability of the HENet model. A predicted risk equal to or greater than 50% was interpreted as predictive of hematoma expansion.

[0058] Data preprocessing This paper employs several key techniques to address challenges in medical image analysis. Pixel value truncation is performed by limiting pixel intensities to a range of 45 to 100, reducing the influence of extreme values and outliers. This normalization process standardizes the dataset and improves the reliability of subsequent analysis. To address inconsistencies in voxel spacing between volumes of interest, spatial normalization is implemented. A fixed-resolution resampling method is used for spatial normalization, ensuring a uniform voxel spacing of 1 mm × 1 mm × 1 mm for accurate comparison and evaluation.

[0059] In terms of data preparation, based on the annotated region of interest information, the maximum cross-section of the lesion area and its two slices (±2) above and below it were extracted to form the minimum bounding rectangle of all sub-regions. During this cropping process, some samples may lack slices with contextual information, such as when the maximum cross-section appears at the bottom of the ROI. In these cases, the missing slices are filled with the maximum cross-section, creating a 2.5D grayscale image dataset containing three images. After normalization by minimum-maximum transformation so that the value range reaches [-1,1], each cropped sub-region image is resized to 224×224 using the nearest neighbor interpolation method to ensure compatibility with the input requirements of the model.

[0060] Model building HENet is a robust 2.5D DCNN model built on the foundation of 2D models and employs residual blocks in its design. The model uses a pre-trained ResNet101 as a feature extractor and is applied to three binary classification tasks, namely RHE1, RHE2, and CHE. The HENet model takes a 2.5D input and focuses on the slice along the largest lesion area and its adjacent slices (two above and two below (±2)). Each CT sequence is treated as an independent image, and the ROI is defined as the smallest rectangle that contains the lesion area in all slices. After cropping and resizing these ROIs, they are input into a parallel backbone network to extract a 2048-dimensional feature vector. Max pooling across slices produces a single 2048-dimensional feature vector, which is then passed through a fully connected layer for binary classification.

[0061] In order to fully utilize the advantages of deep learning models, clinical (age, gender, CT time, GCS score) and radiological features (cerebral parenchymal hemorrhage and intraventricular hemorrhage volume, cerebral hemorrhage location, cerebral hemorrhage signs, subarachnoid hemorrhage) were combined with deep learning features to form a combined feature set. Figure 1 Instructions.

[0062] Model training process To address the overfitting problem of deep convolutional neural network models during training for all tasks, data augmentation techniques were implemented, including the following strategies: (1) randomly selecting the center slice of the 2.5D input from the first three slices with the largest lesion area; (2) applying a random translation within 10% and a random scaling in the range of [0.9, 1.1] to the ROI; and (3) mirror-flipping the cropped ROI horizontally and vertically with a probability of 50%.

[0063] The learning rate of the model backbone is based on the minimum learning rate and the maximum learning rate Sure. Set to 0, Set to 0.01. Parameters Represents the number of iteration cycles. Since the main part of the model uses pre-trained parameters, the current number of iterations Fine-tune it while learning to ensure effective knowledge transfer. Therefore, the learning rate of the task-specific part is calculated as follows:

[0064] The learning rate of the backbone is determined as follows:

[0065] The learning rate varied with the number of iterations. The model underwent approximately 2,967 iterations, equivalent to 90 epochs of training, with 32 rounds of early stopping. The batch size was configured to 16, and the SGD optimizer was used. All training processes were performed on an NVIDIA 4090 GPU using Onekey version 2.5.3 and PyTorch version 1.8.1. The learning rate calculation method for the task-specific part and the learning rate calculation method for the backbone part can force the model to escape the flat area and improve generalization ability. The performance of the model has been verified in two external validation cohorts, including discrimination ability, calibration curve, and clinical utility, and satisfactory results were achieved in predicting both revised hematoma expansion and traditional hematoma expansion.

[0066] Visual Explanation Generation After training and testing the HENet model, visual explanations were performed to understand the specific image components the model focused on when making its decisions. Grad-CAM was used as a means to generate these visual explanations for each image patch. The class localization map was generated by visualizing the gradients flowing into the network's final convolutional layer, which occurs before the fully connected layers. This specific layer was chosen for its ability to preserve spatial information associated with the input image's class, a property that can be compromised in fully connected layers.

[0067] Notably, the Grad-CAM technique is able to generate these graphs without requiring any modifications to the existing model architecture or additional model training.

[0068] Statistical analysis Categorical data are presented as absolute and relative frequencies. For continuous variables, normality was assessed using the Shapiro-Wilk test. If normality was violated, continuous variables were reported as medians and interquartile ranges. Differences in patient and hematoma characteristics among the three datasets were examined using the chi-square test or the Kruskal-Wallis test.

[0069] Regarding model performance, the predictive performance of the HENet model, 2D model, and physician predictions was evaluated using receiver operating characteristic (ROC) curves in the YY and LA datasets. Calibration and decision curve analyses were also performed to assess calibration performance and clinical utility. The Hosmer-Lemeshow goodness-of-fit test was used to assess the calibration capabilities of the HENet model, 2D model, and physician predictions.

[0070] For model comparison, the performance of the HENet model was compared with that of the 2D model and physician predictions using multiple measures of discriminability improvement: specifically, the increase in area under the curve (AUC) was assessed by the DeLong test, as well as the integrated discrimination index (IDI) and continuous net reclassification index (NRI).

[0071] Table 2 summarizes the patient and hematoma characteristics in the HS, YY, and LA datasets. The HS dataset included 522 subjects with a mean age of 62 years, while the YY and LA datasets included 100 and 96 subjects with a mean age of 57 and 68 years, respectively. The majority of subjects in all datasets were male. Age, time from onset to CT scan, and Glasgow Coma Scale (GCS) score showed significant differences between the different datasets (p < .001). With regard to the location of ICH, deep and lobar locations were the most prevalent. Significant variations were observed in the location of ICH (p = .002), intraparenchymal hemorrhage volume (p = .038), intraventricular hemorrhage volume (p = 0.033), and the presence of RHE1 (p = 0.006). No significant differences were observed with respect to NCCT markers or the presence of SAH, RHE2, and CHE.

[0072] Table 2. Baseline characteristics of the HS, YY, and LA datasets feature HS dataset (n=522) YY dataset (n=100) LA dataset (n=96) value age* 62(49-71) 57(47-66) 68(60-76) <0.001‡ gender 0.079 male 395(76) 67(67) 65(68) female 127(24) 33(33) 31(32) CT start time* 4(3-5) 2(1-4) 3(2-5) <0.001‡ GCS score* 15(11-15) 15(15-15) 14(12-15) <0.001‡ ICH Location 0.002‡ depth 317(61) 74(74) 56(58) lungs 141(27) 20(20) 27(28) brainstem 37(7) 1(1) 1(1) cerebellum 27(5) 5(5) 12(12) ICH volume* 20(9-45) 22(11-43) 14(7-34) 0.038‡ IVH volume*† 8(3-17) 8(2-22) 3(2-6) 0.033‡ Heterogeneous density 145(28) 28(28) 28(29) 0.962 Whirlpool sign 334(64) 64(64) 71(74) 0.161 low density 298(57) 62(62) 63(66) 0.235 Black hole sign 211(40) 42(42) 46(48) 0.391 Mixed logo 36(7) 9(9) 6(6) 0.71 Liquid level 43(8) 2(2) 5(5) 0.062 irregular shape 242(46) 42(42) 34(35) 0.124 Island Flag 125(24) 22(22) 17(18) 0.399 For administration 248(48) 43(43) 41(43) 0.537 SAH 56(11) 4(4) 12(12) 0.084 CHE 125(24) 31(31) 27(28) 0.272 RHE1 414(79) 68(68) 83(86) 0.006‡ RHE2 177(34) 42(42) 31(32) 0.255 Note: Unless otherwise indicated, data are numbers of patients, with percentages in brackets.

[0073] *Data are medians, with interquartile ranges in brackets.

[0074] †IVH was present in 241 patients in the HS dataset, in 43 patients in the YY dataset, and in 28 patients in the LA dataset.

[0075] ‡ indicates statistical significance; P < 0.05.

[0076] Abbreviations: CHE, conventional hematoma extension; GCS, Glasgow Coma Scale; ICH, intracerebral hemorrhage; IVH, intraventricular extension of intracerebral hemorrhage; RHE1, revised definition of hematoma extension one; RHE2, revised definition of hematoma extension two; SAH, subarachnoid hemorrhage.

[0077] Model performance The HENet model demonstrated significant predictive power in predicting hematoma expansion, as evidenced by its performance across different metrics. In predicting the revised definition of hematoma expansion one (RHE1), the HENet model achieved AUC values of 0.882 and 0.854 on the YY and LA datasets, respectively, with accuracy scores of 0.870 and 0.844. The model's sensitivity was particularly prominent, peaking at 0.956 on the YY dataset. In predicting the revised definition of hematoma expansion two (RHE2), the HENet model maintained consistent positive predictive values despite variations in sensitivity and specificity across datasets, indicating its reliable prediction of true positive cases. For conventional hematoma expansion (CHE), despite variations in values across datasets, the model's negative predictive value remained high, highlighting its strong ability to correctly identify non-expanding hematomas. These findings are summarized in Tables 3 and Figure 2 This is shown in the , which further confirms the effectiveness and reliability of the HENet model in predicting hematoma expansion.

[0078] Table 3. Physician performance in predicting hematoma expansion

[0079] Abbreviations: AUC: area under the curve; CHE: conventional hematoma expansion; NPV: negative predictive value; PPV: positive predictive value; RHE1: revised hematoma expansion definition one; RHE2: revised hematoma expansion definition two; PPV: positive predictive value; NPV: negative predictive value.

[0080] The calibration curve analysis of the HENet model showed that the predicted and actual probabilities achieved reasonable consistency in both datasets for all definitions of hematoma expansion, such as Figure 3As shown in the figure. In the YY dataset, the decision curve shows that when the threshold probability is 30%-95% in RHE1, 30%-70% in RHE2, and 20%-65% in CHE, the HENet model outperforms the "treat all patients" and "do not treat any patients" schemes in terms of net benefit. Similarly, the LA dataset shows that when the threshold probability is 65%-95% in RHE1, 0%-100% in RHE2, and 18%-65% in CHE, the HENet model can outperform the two schemes in terms of net benefit, as shown in the figure. Figure 4 shown.

[0081] Model comparison Table 4 shows the intraclass correlation coefficient of hematoma expansion (HE) predicted by two doctors. The performance of doctors’ prediction and the performance of 2D model are described in detail in Table 5 and Table 6, respectively. The ROC curves of doctors’ prediction and 2D model as well as the calibration curve and decision curve can be found in Figure 2 、 Figure 3 and Figure 4 Found in.

[0082] Table 4. Intraclass Correlation Coefficient (ICC) of Two Doctors in Predicting Hematoma Expansion

[0083] Table 5. Physician performance in predicting hematoma expansion

[0084] Table 6. Performance of 2D model in predicting hematoma expansion

[0085] Table 7 shows a comparative analysis of the HENet model and physician predictions for hematoma outcomes. The HENet model significantly outperformed physician predictions in terms of the net reclassification index (NRI) and integrated discrimination index (IDI) for RHE1 and RHE2 outcomes on both datasets. Although HENet showed higher values for CHE outcomes, these differences were not statistically significant, indicating comparable performance to physician predictions for this specific outcome.

[0086] Table 7. Comparative analysis of hematoma prediction results between HENet model and doctors

[0087] Note: Delong Test: Delong test; Continuous NRI: Continuous Net Reclassification Index; IDI Analysis: Integrated Discriminant Improvement Index Analysis.

[0088] Table 8 shows a comparative analysis of the HENet model and the 2D model in predicting hematoma expansion. For the RHE1 and RHE2 results, significant improvements were detected in both the NRI and IDI values across datasets, validated by p-values less than 0.001. Although the AUC values increased, these increases did not reach statistical significance, with p-values exceeding 0.05. In the YY dataset, all indicators showed significant increases (p<0.001 and p=0.006). Third, regarding the LA dataset, although the AUC value of the HENet model in predicting CHE decreased compared to the 2D model, this decrease was not statistically significant.

[0089] Table 8. Comparative analysis of the performance of HENet model and 2D model in predicting hematoma results

[0090] Note: Delong Test: Delong test; Continuous NRI: Continuous Net Reclassification Index; IDI Analysis: Integrated Discriminant Improvement Index Analysis.

[0091] Model Interpretation In this study, the Grad-CAM technique was applied to visualize the activations of the last convolutional layer for hematoma expansion (HE) prediction. By making the last convolutional layer transparent, the regions in the input image that contribute most to HE prediction are displayed. The colors on the heat map represent the degree of contribution of each region to the prediction result. Red areas indicate positive influence, meaning that the features in these regions are more likely to guide the model to predict a specific category. Blue areas indicate negative influence, suggesting that the features in these regions are more likely to prevent the model from predicting a specific category. This visualization provides valuable insights into the model's decision-making process without the need for complex architectural changes or model retraining, such as Figure 5 、 Figure 6 shown.

[0092] Figure 5 Figure 3. Model interpretability and Grad-CAM for a patient with a left basal ganglia intracerebral hemorrhage. The first row shows the patient's baseline CT image. In the second row, individual-specific explanations based on attention-based interpretability are displayed as high-resolution attention heat maps. The regions of high attention in the heat map (highlighted in red) correspond to morphological features that help the model predict hematoma expansion. The white arrow indicates the island sign. The last row shows the follow-up CT image, which reveals the expansion of the patient's intracerebral hemorrhage volume.

[0093] Figure 6The model interpretability and Grad-CAM of a patient with left basal ganglia cerebral hemorrhage with intraventricular hemorrhage expansion are shown, where the first row shows the patient's baseline CT image. In the second row, local explanations are provided for each patient, which are presented in the form of high-resolution attention heat maps, utilizing attention-based interpretability. The areas of increased attention shown in the heat map (indicated in red) correspond to morphological features that help the model predict hematoma expansion. The last row shows follow-up CT images, indicating the growth of the patient's cerebral hemorrhage and intraventricular hemorrhage expansion. It can be seen that the present application scheme helps to predict cases of ventricular hemorrhage expansion.

[0094] In this study, we introduced a deep learning model for hematoma expansion prediction, identified as the HENet model. The HENet model incorporates the power of 2D transfer learning and maximizes the capture of 3D spatial information by extracting the largest slice of the lesion. In addition, the model incorporates clinical features in an end-to-end manner. The performance of the model has been validated in two external validation cohorts in terms of discrimination ability, calibration curves, and clinical utility, achieving promising results in predicting revised hematoma expansion (RHE) and conventional hematoma expansion (CHE). In terms of model interpretability, the HENet model has superior interpretability, using the Grad-CAM technique to visualize the contribution of different regions in the input image to HE prediction. This provides valuable insights into the model's decision-making process and enhances our understanding of its functionality without the need for complex modifications or retraining of the model.

[0095] Few studies of deep learning models have focused on predicting RHE. This application fills this gap by focusing on RHE prediction. The HENet model demonstrated robust performance in predicting RHE1, with AUCs of 0.882 and 0.854 for the YY and LA datasets, respectively. It also performed reasonably well in predicting RHE2, with AUCs of 0.728 and 0.811 for the two datasets, respectively. Notably, the HENet model significantly outperformed clinicians' predictions of RHE outcomes in both cohorts. Compared with the 2D model, HENet achieved significant improvements in the net reclassification index (NRI) and integrated discrimination index (IDI) values for RHE outcomes, with significant p-values obtained. In addition, this application also performed external validation of the model on these two datasets. The HENet model demonstrated reasonable performance in predicting CHE outcomes, with AUCs of 0.712 for the YY dataset and 0.716 for the LA dataset.

[0096] Currently, the pathophysiology of hematoma expansion is unclear. Miller Fisher proposed an alternative "avalanche" model, suggesting that the initial hematoma displaces and ruptures surrounding vessels, leading to additional bleeding. Figure 5As shown, our model's interpretable results using Grad-CAM indicate that it focuses on the periphery of the hematoma, supporting the "avalanche" model. This location focus, determined by our model, echoes the findings of Trans et al., who used a deep learning model to predict hematoma growth but did not consider the expansion of intraventricular hemorrhage. In addition, Figure 5 In the N+2 layer of the Grad-CAM, the model shows concern for multiple small hemorrhages (island sign) surrounding the hematoma. One potential reason for the island sign is that it represents multiple active bleeding from several ruptured arterioles, which is also consistent with the "avalanche" model. In addition, for patients with intracerebral hemorrhage with enlarged ventricular hemorrhage, Figure 6 Grad-CAM analysis revealed that the HENet model particularly emphasizes the ventricular region. This observation indirectly reinforces the view that the HENet model considers IVH growth as a significant complication of ICH, significantly impacting patient outcomes. Analysis using Grad-CAM revealed significant instability in the ICH and IVH regions, indicating that these areas are prone to active bleeding and further expansion. This insight provides a key insight into the mechanisms of hematoma expansion.

[0097] We introduced a novel 2.5D deep learning model that combines strong performance, high interpretability, and clinical usability. These advances played a key role in improving HE risk prediction and helping to understand the mechanisms by which deep learning models predict HE. Given the results of this study, we believe that the HENet model will have a positive impact on clinical practice: deep learning technology can be integrated into electronic medical records and picture archiving and communication systems (PACS) for use as a clinical decision support system. HENet accurately identifies patients at risk for hematoma expansion, enabling timely intervention and transfer to intensive care units for close monitoring and active management. Deep learning-enhanced clinician decision-making has the potential to improve outcomes for patients with ICH.

[0098] Example 2: Based on the above embodiments, to address the issue of multi-center data discrepancies, especially device discrepancies, a dynamic DICOM adaptation and federated learning module (HE-FedDICOM) is added, specifically including: DICOM metadata perception unit: automatically reads CT slice thickness, tube voltage and other parameters; Dynamic normalization: adjust the grayscale normalization range according to the KV value; Adaptive resampling: intelligent interpolation of data with different layer thicknesses (1mm-5mm); Privacy protection architecture: Front-end: Desensitization processing based on the DICOM standard automatically removes sensitive fields such as the patient name; Back-end: A lightweight segmentation model (Pruned U-Net) runs locally in the medical institution; Federated learning mechanism: The central server aggregates the model parameters of each hospital, adopts differential privacy-protected gradient transmission, designs a hematoma-specific aggregation algorithm, and pays weighted attention to key layer parameters.

[0099] Verification results show that the system's robustness to differences in slice thickness has been improved. On mixed data of 1mm / 3mm / 5mm, the AUC fluctuation is ≤5% (traditional methods fluctuate up to 15%). Robustness to tube voltage differences has also been improved. As shown in Table 9, the performance difference has been reduced from 14.3% to 7.7% for data collected by different GE / Philips / Siemens devices. In addition, only model parameters need to be uploaded instead of the original DICOM data, so the data transmission volume is reduced by 83%. The hospital's local inference speed is <4 seconds per case, meeting the real-time needs of clinical emergency. The optimization proposes the first federated learning system for the prediction of cerebral hemorrhage, which achieves cross-device and cross-center model generalization through DICOM metadata awareness. Privacy protection is implemented throughout the entire process from data desensitization to model training to solve the problem of sensitive data sharing in the implementation of medical AI.

[0100] Table 9. Robustness of HE-FedDICOM on multi-center data

[0101] Example 3: Based on the above embodiment, a computer-readable storage medium is further provided, in which a computer program that can be executed by a processor is stored. When the computer program is executed by the processor, at least one step of the above method is executed and the same technical effect can be achieved. To avoid repetition, this embodiment will not be described in detail.

[0102] Example 4: Based on the aforementioned embodiment, a computer device is provided, comprising a memory, a processor, a communication interface, and a communication bus; wherein the memory, the processor, and the communication interface communicate with each other via the communication bus; the memory is used to store computer programs; the processor is used to execute the computer programs stored in the memory, and when the processor executes the computer program, it implements at least one step of the aforementioned method and can achieve the same technical effect. To avoid repetition, this embodiment will not be described in detail.

[0103] Computer-readable media include permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0104] The conventional techniques in the above embodiments are prior arts known to those skilled in the art, and thus will not be described in detail here.

[0105] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope defined by the appended claims.

[0106] Although the present invention has been described in detail and certain specific embodiments have been cited, it will be apparent to those skilled in the art that various changes and modifications can be made without departing from the spirit and scope of the invention.

[0107] Although the above-mentioned specific embodiments have shown, described and pointed out the novel features applied to various embodiments, it should be understood that various omissions, replacements and changes can be made to the form and details of the described devices or methods without departing from the spirit of the present disclosure. In addition, the various features and methods described above can be used independently of each other, or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of the present disclosure. Many of the above-mentioned embodiments include similar components, and therefore, these similar components are interchangeable in different embodiments. Although the present invention has been disclosed in the context of certain embodiments and examples, it should be understood by those skilled in the art that the present invention can extend beyond the specifically disclosed embodiments to other alternative embodiments and / or applications and their obvious modifications and equivalents. Therefore, the present invention is not intended to be limited by the specific disclosure of the preferred embodiments herein.

[0108] Matters not covered in the present invention are all known technologies.

Claims

1. A deep learning-based system for predicting hematoma expansion in cerebral hemorrhage, characterized in that: include: A data acquisition module, used to obtain baseline CT images and clinical data of patients; An image preprocessing module, used for performing standardization processing on the CT image; A feature extraction module, which uses a 2.5D deep convolutional neural network to extract image features from the CT image; a feature fusion module for fusing the image features with clinical features and radiological features extracted from clinical data; A prediction module predicts the risk of hematoma expansion based on the fused features, wherein the hematoma expansion includes ventricular hemorrhage.

2. The system according to claim 1, wherein: The processing performed by the image preprocessing module includes: Limit CT values to the range of 45-100HU; A fixed-resolution resampling method was used to unify the voxel spacing to 1 mm × 1 mm × 1 mm; Extract the largest cross-section of the lesion area and its adjacent slices to form the 2.5D input; Use min-max transformation to adjust the value range to [-1,1]; Resize the image to 224×224.

3. The system according to claim 1, wherein: The feature extraction module uses a 2.5D deep convolutional neural network with ResNet101 as the backbone network. The input is 5 consecutive CT slices, which include the slice along the maximum lesion area and two adjacent slices above and below it. A 2048-dimensional feature vector is extracted through a parallel backbone network, and a single feature vector is obtained after cross-slice maximum pooling.

4. The system according to claim 1, wherein: The clinical characteristics included age, sex, time from onset to CT scan, and GCS score on admission; The radiological features include intracerebral hemorrhage and intraventricular hemorrhage volume, cerebral hemorrhage location, cerebral hemorrhage signs, and subarachnoid hemorrhage.

5. The system according to claim 1, wherein: The prediction results output by the prediction module include the following three risks of hematoma expansion: RHE1: intracerebral hemorrhage >6 ml or increased by 33% or any expansion of ventricular hemorrhage; RHE2: intracerebral hemorrhage > 6 ml or increased by 33% or intraventricular hemorrhage > 1 ml; CHE: intracerebral hemorrhage > 6 ml or increased by 33%.

6. The system according to claim 1, wherein: The 2.5D deep convolutional neural network of the feature extraction module adopts the following strategy during training: (1) Randomly select the central slice of the 2.5D input from the first three slices with the largest lesion area; (2) Apply a random translation within 10% and a random scaling within the range of [0.9, 1.1] to the ROI; (3) The cropped ROI is mirror-flipped horizontally and vertically with a probability of 50%.

7. The system according to claim 1, wherein: The learning rate of the backbone of the 2.5D deep convolutional neural network is based on the minimum learning rate and the maximum learning rate Sure, Set to 0, Set to 0.01; The learning rate for the task-specific part is calculated as follows: The learning rate of the backbone is determined as follows: in, Indicates the number of iteration cycles; Indicates the current iteration number.

8. The system according to claim 1, wherein: It also includes a visual explanation module that uses gradient-weighted class activation mapping technology to generate attention heat maps that show the image areas that the model pays attention to when making predictions.

9. The system according to any one of claims 1 to 8, characterized in that: It also includes dynamic DICOM adaptation and federated learning modules, including: DICOM metadata awareness unit: automatically reads CT device parameters; Dynamic normalization: adjust the grayscale normalization range according to the KV value; Adaptive resampling: intelligent interpolation of data of different layer thicknesses; Privacy protection architecture: Front-end: Desensitization processing based on the DICOM standard, automatically deleting sensitive fields; Back-end: Lightweight segmentation model runs locally in medical institutions; Federated learning mechanism: The central server aggregates the model parameters of each hospital, adopts differential privacy-protected gradient transmission, designs a hematoma-specific aggregation algorithm, and pays weighted attention to key layer parameters.

10. A deep learning-based method for predicting hematoma expansion in cerebral hemorrhage for non-disease diagnosis and treatment, characterized by: The execution of the method is based on the system according to any one of claims 1 to 9; The method comprises the following steps: Obtain the patient's baseline CT images and clinical data; performing standardized preprocessing on the CT image; Use 2.5D deep convolutional neural network to extract image features from preprocessed CT images; fusing the imaging features with clinical features; Predict the risk of hematoma expansion based on post-fusion features; Generate visual explanations of model decisions.

Images