Fatty liver auxiliary evaluation method and system based on whole liver multi-dimensional characteristics and machine learning model

Abstract

The application belongs to the field of fatty liver evaluation, and discloses a fatty liver auxiliary evaluation method and system based on whole-liver multi-dimensional features and a machine learning model. After obtaining three-dimensional masks of a liver and a spleen, multi-dimensional features related to fatty liver grading and clinical and pathological features are extracted from CT image data of a target region, so that fatty liver evaluation is no longer limited to a single liver-to-spleen ratio, but comprehensively depicts overall features of the liver from multiple angles. By using multi-dimensional information in the whole liver range, comprehensive quantification of the degree of fatty infiltration is realized, thereby overcoming the problems of insufficient information utilization and difficulty in reflecting the overall picture of the disease in the prior art. The extracted multi-dimensional features are integrated and standardized, so that features of different dimensions and different scales are in a unified numerical space, and the consistency and comparability of model input data are ensured.

Description

Technical Field

[0001] This invention belongs to the field of fatty liver assessment, specifically involving a method and system for auxiliary assessment of fatty liver based on multidimensional features of the whole liver and machine learning models. Background Technology

[0002] Fatty liver disease, also known as fatty liver disease, is a pathological condition caused by abnormal fat deposition in liver cells and has become one of the most common chronic liver diseases worldwide. Its severity is closely related to the risk of cirrhosis, liver cancer, and cardiovascular events. Therefore, accurate and objective severity grading of fatty liver is of great significance for developing individualized intervention strategies, evaluating treatment effects, and predicting patient prognosis.

[0003] Currently, abdominal computed tomography (CT) is the most commonly used and important non-invasive imaging method for clinical assessment of fatty liver. In current clinical practice, the liver fat content is usually indirectly reflected by measuring the CT values ​​of the liver and spleen and calculating the liver-spleen CT value ratio. According to the generally accepted grading standards, a liver-spleen ratio greater than 0.7 indicates mild fatty liver, between 0.5 and 0.7 indicates moderate fatty liver, and less than or equal to 0.5 indicates severe fatty liver. This method is relatively simple to operate and has been widely used in clinical practice.

[0004] However, existing technologies heavily rely on manual operation in their implementation. Doctors typically need to select a representative image from the same CT slice, manually or semi-automatically delineate the regions of interest (ROIs) of the liver and spleen on that slice, calculate their average CT values, and then determine the liver-spleen ratio. This method has the following drawbacks: First, the measurement process is highly subjective and lacks consistency. Different doctors may differ in selecting scanning planes, determining the location, size, and shape of the region of interest (ROI), and even measurements taken by the same doctor at different times may produce deviations, leading to insufficient repeatability of results. In long-term follow-up or efficacy evaluation, this human error may be amplified, affecting the comparability of results before and after.

[0005] Secondly, existing methods are usually based on single-level, localized measurements, which have limited information utilization and are difficult to reflect the overall three-dimensional structure of the liver and the characteristics of fat distribution. Fat is often unevenly distributed in the liver, and single-point measurements cannot fully reflect the extent, proportion, and density distribution of fat infiltration, nor can they sensitively capture subtle changes that occur during treatment.

[0006] Furthermore, the existing grading model uses a three-level classification method with fixed thresholds, resulting in discrete grading results that are difficult to accurately reflect the continuous changes in fatty liver disease. When the condition is within the threshold range, even slight fluctuations in the numerical values ​​can cause a jump in the grading, which is not conducive to the refined monitoring of the dynamic evolution of the disease.

[0007] In addition, manual case-by-case drawing and calculation is inefficient and can easily increase the workload of doctors in large-scale physical examinations or long-term follow-up management scenarios, making it difficult to meet the needs of large-scale and standardized applications.

[0008] In summary, existing CT grading techniques for fatty liver based on manual single-layer measurement suffer from problems such as strong subjectivity, insufficient information utilization, crude grading methods, and low efficiency, making it difficult to achieve objective, accurate, and continuous dynamic assessment of the severity of fatty liver. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems of existing methods, such as strong subjectivity, insufficient information utilization, coarse grading and low efficiency, which make it difficult to achieve objective, accurate and continuous dynamic assessment of the severity of fatty liver. The invention provides a method and system for auxiliary assessment of fatty liver based on multidimensional features of the whole liver and machine learning models.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an auxiliary assessment method for fatty liver based on multidimensional features of the whole liver and a machine learning model, comprising the following steps: Acquire CT image data of the target area, preprocess the CT image data, and segment out the three-dimensional mask of the liver and spleen; Based on the three-dimensional mask of the liver and spleen, multidimensional features related to the clinical and pathological classification of fatty liver are extracted from the CT image data of the target area. All extracted clinically and pathologically relevant multidimensional features related to fatty liver grading were integrated and standardized. A grading model was then constructed based on the standardized multidimensional features using a random forest classifier. A grading model was used to evaluate CT image data during the follow-up period to obtain auxiliary evaluation results.

[0011] A further improvement of this invention lies in the following method for acquiring CT image data of the target area, preprocessing the CT image data, and segmenting out the three-dimensional mask of the liver and spleen: Acquire CT image data of the target area; The CT image data is linearly mapped using preset abdominal window width and window level to complete the window width and window level adjustment; The CT image data after window width and window level adjustment are normalized and resampled to a uniform spatial resolution. A pre-trained 3D U-Net was used as a segmentation model to process the resampled CT image data and segment out the 3D masks of the liver and spleen.

[0012] A further improvement of this invention is that the pre-trained 3D U-Net as a segmentation model includes: The encoder consists of 4–5 downsampling stages. Each stage contains two 3×3×3 convolutional layers, a ReLU activation function, a batch normalization, followed by 2×2×2 max pooling for downsampling. The decoder uses transposed convolution or interpolation to upsample each upsampling stage, and concatenates the feature maps of the corresponding stages of the encoder through skip connections, and then goes through two 3×3×3 convolutional layers. The output layer uses 1×1×1 convolution and the Softmax activation function to output a binary probability map of the liver and spleen.

[0013] A further improvement of this invention lies in the following method for training the pre-trained 3D U-Net as a segmentation model: Several manually annotated 3D masks of the liver and spleen were obtained as standard data; The default loss function is a combination of Dice Loss and cross-entropy loss; The standard data U-Net model is trained by randomly applying rotation, scaling, translation, and elastic deformation during training. The Adam optimizer is used, and the initial learning rate of the U-Net model is set to 1e-4. The learning rate is reduced to half every 50 rounds until the training result meets the requirements. Training is then stopped to obtain the pre-trained 3D U-Net.

[0014] A further improvement of this invention lies in that, based on a three-dimensional mask of the liver and spleen, multidimensional features related to the clinical and pathological classification of fatty liver are extracted from the CT image data of the target area. These multidimensional features include: The ratio characteristics are the liver-spleen ratio and the liver-kidney ratio; Morphological characteristics include volume, sphericity and ellipsoidity, and discrete packing density; Global density features include mean density, median density, and mode density; Key interval characteristics include the proportion of low-density intervals, the proportion of high-density intervals, and the average density of specific intervals; The distribution shape characteristics are skewness and kurtosis.

[0015] A further improvement of this invention lies in the feature integration and standardization of all extracted clinically and pathologically relevant multidimensional features related to fatty liver grading, as follows: We acquire clinically and pathologically relevant multidimensional features related to fatty liver grading, integrate all clinically and pathologically relevant multidimensional features related to fatty liver grading into a high-dimensional feature vector, and use Z-score standardization to normalize the features and eliminate the influence of dimensions.

[0016] A further improvement of this invention lies in the following specific method for constructing a gradation model based on a random forest classifier using standardized multidimensional features: Pre-set several pre-labeled tag data; Divide the label data into training and validation sets according to the required proportions; A random forest classification model is constructed, which consists of several decision trees. Subsamples are randomly drawn from the training set to construct each decision tree through bootstrapping. At the same time, some features are randomly selected for division when splitting nodes. During the training phase, the random forest classification model learns the ability of different features to distinguish fatty liver severity, and outputs the final classification result through a multi-decision-tree voting mechanism, while generating probability estimates for each category. A grid search combined with cross-validation is used to perform combined traversal of parameters including the number of decision trees, maximum tree depth, minimum number of split samples, minimum number of leaf node samples, and maximum number of features. K-fold cross-validation is used in the training set to evaluate the model performance indicators under different parameter combinations. The parameter combination with the best overall performance is selected as the final model configuration, which is the optimized gradation model. The optimized grading model is evaluated on the validation set. Once the validation is successful, it is used as the final grading model.

[0017] A further improvement of this invention lies in the following method for evaluating CT image data during the follow-up period using a grading model to obtain auxiliary evaluation results: CT image data during the follow-up period were obtained and fed into the grading model to obtain efficacy indicators. Based on the changes in efficacy indicators, auxiliary assessment results are obtained.

[0018] A further improvement of this invention lies in the following method for calculating the auxiliary evaluation results: The change in rating is: ΔScore = Score_follow-up - Score_baseline Wherein, Score_follow-up is the efficacy index value, and Score_baseline is the efficacy index baseline value; If ΔScore ≤ -10, it is considered a significant improvement; If -5 ≥ ΔScore > -10, it is considered a slight improvement; If |ΔScore| < 5, the condition is considered stable; If ΔScore ≥ 5, it is considered as disease progression.

[0019] Secondly, this invention provides an auxiliary assessment system for fatty liver based on multidimensional features of the whole liver and a machine learning model, comprising: The preprocessing module is used to acquire CT image data of the target area, preprocess the CT image data, and segment out the three-dimensional mask of the liver and spleen. The multidimensional feature extraction module is used to extract clinically and pathologically relevant multidimensional features related to fatty liver grading from CT image data of the target area based on a three-dimensional mask of the liver and spleen. The grading model training module is used to integrate and standardize all extracted clinical and pathological multidimensional features related to fatty liver grading, and then construct a grading model based on the standardized multidimensional features using a random forest classifier. The model-assisted evaluation module is used to evaluate CT image data during the follow-up period using a grading model to obtain auxiliary evaluation results.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention, after obtaining a three-dimensional mask of the liver and spleen, extracts multidimensional features clinically and pathologically relevant to fatty liver grading from CT image data of the target area. This allows fatty liver assessment to move beyond a single liver-spleen ratio and comprehensively characterize the overall liver features from multiple perspectives. By utilizing multidimensional information across the entire liver, it achieves a comprehensive quantification of the degree of fatty infiltration, overcoming the problems of insufficient information utilization and difficulty in reflecting the full picture of the disease in existing methods. This invention integrates and standardizes the extracted multidimensional features, placing features of different dimensions and scales in a unified numerical space, ensuring the consistency and comparability of the model input data. Based on this, a grading model is constructed using the standardized multidimensional features based on a random forest classifier. Leveraging the random forest's ability to process high-dimensional features and its automatic learning of feature importance, a comprehensive judgment of the severity of fatty liver is achieved. This process is automatically completed by the model, reducing manual intervention and making the grading results more stable and repeatable, thus overcoming the problems of strong subjectivity and poor consistency in traditional grading methods. This invention uses a grading model to evaluate CT image data during the follow-up period, achieving quantitative analysis of the same patient at different time points. Based on a unified data processing workflow and the comparability of results output by the same grading model, the assessment results before and after evaluation are placed under the same evaluation system, thus supporting continuous dynamic monitoring of the severity of fatty liver. Compared with the traditional coarse grading method with fixed thresholds, this method can reflect the trend of disease changes based on the model output results, improving the sensitivity to treatment effects or disease progression. The entire process, from CT image data acquisition, preprocessing, 3D segmentation, multi-dimensional feature extraction, feature integration and standardization, to the construction of the grading model using a random forest classifier and automatic assessment during the follow-up period, can be completed automatically within the system, reducing manual operation steps, improving assessment efficiency, and meeting the needs of large-scale screening and long-term follow-up management. In conclusion, this invention has significant beneficial effects in improving assessment objectivity, enhancing the sufficiency of information utilization, refining grading results, and improving assessment efficiency, thereby achieving accurate, stable, and continuous dynamic auxiliary assessment of the severity of fatty liver. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention; Figure 2 This is a system diagram of the present invention. Detailed Implementation

[0022] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0023] Example 1: See Figure 1A method for auxiliary assessment of fatty liver based on multidimensional features of the whole liver and machine learning models includes the following steps: S1: Acquire CT image data of the target area, preprocess the CT image data, and segment out the three-dimensional mask of the liver and spleen.

[0024] S2 extracts clinically and pathologically relevant multidimensional features related to fatty liver grading from CT image data of the target area based on a three-dimensional mask of the liver and spleen.

[0025] S3 integrates and standardizes all extracted clinically and pathologically relevant multidimensional features related to fatty liver grading, and constructs a grading model based on the standardized multidimensional features using a random forest classifier.

[0026] S4 uses a grading model to evaluate CT image data during the follow-up period and obtains auxiliary evaluation results.

[0027] Example 2: See Figure 2 A fatty liver auxiliary assessment system based on whole liver multidimensional features and machine learning models includes: The preprocessing module is used to acquire CT image data of the target area, preprocess the CT image data, and segment out the three-dimensional mask of the liver and spleen.

[0028] The multidimensional feature extraction module is used to extract clinically and pathologically relevant multidimensional features related to fatty liver grading from CT image data of the target area based on a three-dimensional mask of the liver and spleen.

[0029] The grading model training module is used to integrate and standardize all extracted clinical and pathologically relevant multidimensional features related to fatty liver grading, and then construct a grading model based on the standardized multidimensional features using a random forest classifier.

[0030] The model-assisted evaluation module is used to evaluate CT image data during the follow-up period using a grading model to obtain auxiliary evaluation results.

[0031] Example 3: This embodiment is used to automatically obtain a three-dimensional mask of the liver and spleen from abdominal CT images of the target region, and specifically includes the following steps: Step 1: Acquire CT image data of the target area. The system reads the patient's abdominal CT image data from the image storage system. The CT image data is a DICOM format sequence containing transverse images of the entire abdomen, with a slice thickness typically of 1–5 mm.

[0032] Step two involves adjusting the window width and level of the CT image data. The raw CT values ​​are linearly mapped using a preset abdominal window width and level to enhance the contrast between the liver and surrounding tissues; the abdominal window width is set to 350–400 HU, and the window level to 40–60 HU. The purpose of linear mapping is to stretch the grayscale range of the soft tissues of interest to a more suitable intensity range for the segmentation network to learn, reducing the interference of extreme HU values ​​such as bone and gas on the network input.

[0033] The CT image data after window width and window level adjustment were normalized. The mapped image intensity was normalized to the [0,1] interval, and the normalization method was as follows:

[0034] Where I is the intensity value after adjusting the window width and window level. and The sample's minimum / maximum value or a fixed intensity range; preferably, Take -200 HU, x is set to +300 HU to make the sample intensity distribution under different scanning conditions comparable, thereby improving the stability of the segmentation model.

[0035] The normalized CT image data is resampled. All samples are resampled to a preset spatial resolution to ensure consistency of model input; preferably, the resampling resolution is 1 mm × 1 mm × 1 mm. Resampling can eliminate voxel size differences caused by different devices and slice thicknesses, ensuring that the 3D convolution has a consistent receptive field in spatial scale.

[0036] Step 3: A pre-trained 3D U-Net is used to segment the resampled CT image data, obtaining 3D masks for the liver and spleen. The 3D U-Net includes an encoder, a decoder, and an output layer: the encoder consists of 4–5 downsampling stages, each containing two 3×3×3 convolutional layers, a ReLU activation function, and BatchNormalization, with downsampling achieved through 2×2×2 max pooling; the decoder uses transposed convolution or interpolation for upsampling, and concatenates the corresponding feature maps from the encoder stages via skip connections, then performs feature recovery through two 3×3×3 convolutional layers; the output layer uses 1×1×1 convolution combined with a Softmax activation function to output probability maps of the liver and spleen.

[0037] The 3D U-Net is a pre-trained model, and its training process is as follows: 500 manually annotated 3D masks of the liver and spleen are used as the gold standard training data; the loss function uses a combination of Dice Loss and cross-entropy loss to alleviate class imbalance, and the expression is:

[0038] in To predict probabilities, The model uses the true labels; during training, data augmentation is performed by randomly applying rotation (±15°), scaling (0.9–1.1 times), translation (±10 pixels), and elastic deformation to improve the model's generalization ability; the optimizer uses Adam, with an initial learning rate of 1e-4, which is reduced to half every 50 rounds to balance convergence speed and stability.

[0039] The segmentation results are post-processed to obtain the final mask. The probability map output by the model is thresholded, with a preferred threshold of 0.5, to generate a binary mask. Subsequently, three-dimensional connected component analysis is performed on the binary mask to remove small volume noise connected components and retain the largest connected components, which are used as the final output results of the liver three-dimensional mask and the spleen three-dimensional mask, respectively.

[0040] Through the above steps, this embodiment achieves stable and objective acquisition of three-dimensional masks of the liver and spleen from abdominal CT image data through window width and level adjustment, normalization, unified resampling, and automatic segmentation based on pre-trained three-dimensional U-Net. This provides reliable input for subsequent three-dimensional quantitative analysis of the whole liver and automatic grading of fatty liver.

[0041] Example 4: In this embodiment, after obtaining the three-dimensional mask of the liver and spleen, multi-dimensional feature extraction is performed based on the original CT image data to construct a quantitative feature set closely related to the grading of fatty liver, as follows: Based on the obtained 3D masks for the liver and spleen, they are mapped back to the original CT image space. Corresponding CT value data are extracted within the voxel range defined by the masks, forming liver and spleen voxel sets respectively. If the kidney region can be automatically segmented or semi-automatically labeled, the renal cortex voxel set is obtained simultaneously.

[0042] Based on this, multidimensional features related to fatty liver grading are calculated, including the following categories: I. Ratio Characteristics: The liver-spleen ratio is calculated as the ratio of the average CT value of the entire liver to the average CT value of the entire spleen. This ratio directly reflects the degree of change in the overall density of the liver relative to the spleen. As fatty infiltration worsens, the liver CT value decreases, and the liver-spleen ratio decreases. Therefore, this indicator is a core indicator for clinical steatosis grading.

[0043] The liver-kidney ratio is calculated as the ratio of the average CT value of the entire liver to the average CT value of the renal cortex at the same level. This indicator serves as a supplementary reference when the spleen is absent or lesions in the spleen affect the CT value, and it can also reflect the overall degree of fatty infiltration in the liver.

[0044] II. Morphological characteristics: Volume is calculated by statistically analyzing the total number of liver cells using a 3D mask and multiplying this number by the voxel volume. Moderate to severe fatty liver is often accompanied by hepatomegaly; therefore, increased liver volume can serve as an auxiliary indicator of severity.

[0045] Sphericity and ellipsoidity are calculated by determining the geometric relationship between the liver's surface area and volume, resulting in a sphericity index; and by calculating the ratio of principal axis lengths, a ellipsoidity index is obtained. Fatty infiltration may cause the liver edges to become rounded, and increased sphericity or changes in ellipsoidity can quantify the degree of morphological alteration.

[0046] Discrete density, calculated by statistically analyzing the standard deviation or interquartile range of CT values ​​within a liver region, reflects the degree of dispersion in density distribution. When fat distribution is uneven, CT value dispersion increases, which helps differentiate between diffuse fatty infiltration and focal changes.

[0047] III. Global Density Features: Mean density is calculated as the average CT value of all voxels within the liver region. Increased fat content leads to an overall decrease in mean density.

[0048] Median density is the median of liver voxel CT values. This indicator is more robust to abnormally high-density voxels such as those found in blood vessels and calcifications.

[0049] Mode density, which is the intensity value that appears most frequently in the distribution of liver CT values, is used to reflect the density of the main tissue components and can more realistically reflect the density level of the main tissue.

[0050] IV. Characteristics of Key Intervals: The proportion of low-density areas (<40 HU) is calculated as the percentage of voxels with CT values ​​less than 40 HU out of the total number of voxels in the liver. This proportion is positively correlated with the extent of fatty infiltration; the higher the proportion, the more likely the grading is to be moderate to severe.

[0051] The proportion of low-density areas (<30 HU), which is the percentage of voxels with CT values ​​less than 30 HU, is used to identify areas of severe fatty infiltration and is an important indicator for distinguishing between moderate and severe fatty liver.

[0052] The proportion of high-density regions (>60 HU) is the percentage of voxels with CT values ​​greater than 60 HU, reflecting the proportion of residual liver tissue relative to normal tissue. The lower the proportion, the higher the degree of fat replacement.

[0053] Average density within a specific interval (0–40 HU): Calculate the average CT value of voxels located within the 0–40 HU interval. The lower the density within this interval, the more severe the fatty infiltration.

[0054] V. Distribution Shape Characteristics: Skewness is the skewness of the distribution of CT values ​​in the liver. In cases of fatty liver, the CT value distribution often exhibits a negative skewness (low-density tail). The lower the skewness value, the more low-density fat areas are indicated.

[0055] Kurtosis, calculated from the kurtosis of CT values, reflects the sharpness of the distribution pattern. Variations in kurtosis can reflect the uniformity of fat infiltration and tissue replacement patterns.

[0056] Through the above steps, this embodiment systematically extracts multidimensional quantitative indicators such as ratio features, morphological features, global density features, key region features, and distribution shape features from the original CT images based on the three-dimensional mask of the liver and spleen. This comprehensively depicts the range, degree, distribution pattern, and morphological changes of fat infiltration, providing a high-dimensional, objective, and clinically pathological mechanism-related input feature basis for the subsequent automatic grading model.

[0057] Example 5: After completing multidimensional feature extraction, this embodiment further performs feature integration, model training, and automatic gradation, specifically including the following steps: First, feature integration and standardization are performed. All multidimensional features clinically and pathologically relevant to fatty liver grading are acquired, including ratio features, morphological features, global density features, key interval features, and distribution shape features. These features are then concatenated in a uniform order to construct a high-dimensional feature vector. To eliminate the impact of differences in feature dimensions and numerical ranges on model training, Z-score standardization is used to normalize each feature. The standardization formula is as follows:

[0058] in, These are the original eigenvalues. This is the mean of the feature in the training set. The standard deviation is given. The standardized parameters are calculated only based on the training set and remain consistent throughout the validation and practical application phases to ensure data distribution stability.

[0059] After feature standardization, an automatic fatty liver classification model was constructed. The model type chosen was a random forest classifier. A random forest consists of multiple decision trees. Sub-samples are randomly extracted from the training set using a bootstrap method to construct each decision tree, and a subset of features is randomly selected for splitting at each node, thereby enhancing the model's generalization ability. This model can handle high-dimensional features, automatically evaluate feature importance, is insensitive to noise and outliers, and supports multi-class probability output.

[0060] The training data consisted of 444 samples labeled "mild / moderate / severe". All samples were stratified and randomly divided in an 8:2 ratio to form the training and validation sets, ensuring a relatively consistent proportion for each category. The model was trained on the training set, and hyperparameters were optimized using grid search combined with k-fold cross-validation (preferably 5-fold cross-validation). The hyperparameters traversed included combinations of decision tree number, maximum tree depth, minimum number of split samples, minimum number of leaf node samples, and maximum number of features. Performance metrics (such as accuracy and macro-average F1 score) under different parameter combinations were evaluated using cross-validation within the training set, and the optimal parameter combination was selected as the final model configuration.

[0061] To avoid overfitting, the following strategies are implemented: based on the feature importance ranking of the random forest output, the top 15 features are retained and redundant features are removed; the maximum depth of the decision tree is limited to no more than 10 layers and the minimum number of leaf node samples is no less than 5; 5-fold cross-validation is used to evaluate the model's generalization performance; and the accuracy of the validation set is monitored during training, and training is stopped early when no improvement is seen after 10 consecutive rounds.

[0062] After model optimization, the optimized grading model was evaluated on the validation set. Precision, recall, F1 score, and confusion matrix were calculated, and consistency analysis was performed with the diagnostic results of senior physicians. The results showed that, on the independent test set, the Kappa coefficient of consistency between the model's grading results and the diagnoses of senior physicians reached 0.89, significantly higher than the Kappa coefficient of 0.72 for the junior physician group, validating the clinical reliability of the model.

[0063] In practical application, for newly input cases, the standardized multidimensional feature vector is input into the final grading model. The model not only outputs fatty liver grading category labels (mild / moderate / severe) but also outputs probability estimates for each category, such as [0.85, 0.12, 0.03], to represent the grading confidence level. Simultaneously, the system can automatically generate a structured assessment report, including key feature values ​​(such as the liver-spleen ratio, the percentage of voxels below 30 HU, etc.) and visual charts such as liver density distribution histograms to assist physicians in review and clinical decision-making.

[0064] Through the above-mentioned feature integration, standardization processing, random forest model training and validation steps, this embodiment realizes the construction and clinical application of an automatic fatty liver grading model based on multidimensional features of the whole liver.

[0065] Example 6: After completing the construction of the automatic grading model for fatty liver, this embodiment further realizes dynamic auxiliary assessment of efficacy during the follow-up period based on a continuous scoring mechanism, specifically including the following steps: For patients requiring efficacy assessment, baseline (pre-treatment) CT imaging data and at least one follow-up (post-treatment) CT imaging data were obtained. The baseline and follow-up images were processed using the same data processing workflow, including window width and level adjustment, normalization, resampling, 3D organ segmentation, multidimensional feature extraction, and feature standardization. These results were then input into a trained scoring model to obtain the baseline score (Score_baseline) and the follow-up score (Score_follow-up). These scores are continuously quantified output values, capable of reflecting the severity of fat infiltration at high resolution.

[0066] Secondly, the core efficacy indicators are calculated. The system automatically calculates the score change value and score change rate, as follows: Rating change value: ΔScore = Score_follow-up Score_baseline Here, Score_follow-up is the score output by the grading model during the follow-up period, and Score_baseline is the score at the baseline period. ΔScore is used to reflect the absolute magnitude and direction of changes in the condition; a negative ΔScore indicates improvement in the condition, while a positive ΔScore indicates worsening of the condition.

[0067] Rate of change in rating: (ΔScore / |Score_baseline|) × 100% The rate of change in scores is used to represent the relative extent of improvement or progression, making patients with different baseline severity comparable.

[0068] After calculating the indicators, the system automatically determines the therapeutic effect based on preset, clinically interpretable threshold rules. The specific determination rules are as follows: If ΔScore ≤ -10, it is judged as "significant improvement", indicating that the degree of fat infiltration is significantly reduced; If -10 < ΔScore ≤ -5, it is judged as "mild improvement", indicating that there is some degree of improvement; If |ΔScore|<5, it is judged as "stable condition", indicating that the score change is within the range of physiological fluctuation or measurement error. If ΔScore ≥ 5, it is judged as "disease progression", indicating that the degree of fat infiltration has worsened.

[0069] Through the aforementioned continuous scoring calculation and segmented threshold determination mechanism, the system can not only accurately depict the trajectory of disease changes, but also output clear clinical conclusions.

[0070] After the efficacy assessment is completed, the system automatically generates a structured auxiliary evaluation report. The report includes: baseline and follow-up scores, score changes and rates of change, trend graphs, a key feature comparison table (such as the liver-spleen ratio, the proportion of low-density areas, etc.), and the final efficacy assessment result. Physicians can review and make decisions based on this report, thereby achieving objective, quantitative, and continuous dynamic monitoring of the treatment effect of fatty liver. Through the above steps, this embodiment realizes a dynamic auxiliary evaluation method for fatty liver efficacy based on multidimensional features of the whole liver and a machine learning model.

[0071] This invention achieves automatic segmentation of the entire liver and spleen using a three-dimensional deep learning model, with the algorithm automatically completing all subsequent feature extraction and calculation, completely eliminating the subjective arbitrariness of manual selection of layers and delineation regions in traditional methods. This ensures high consistency and repeatability of diagnostic results across different times and operators, providing a reliable basis for long-term follow-up and comparison of the disease. This invention abandons traditional single-level local measurements, fully utilizing the three-dimensional volume information of the entire liver. Through the multi-dimensional quantitative feature set extracted by the system, the disease can be characterized from multiple comprehensive dimensions such as the extent, severity, spatial distribution uniformity, and organ morphology of fatty infiltration. This makes the assessment results more comprehensive and accurate, closer to the comprehensive interpretation level of senior experts. This invention uses a regression model to output a continuous fatty liver severity score, replacing the traditional three-level discrete grading of "mild, moderate, and severe." This innovation significantly improves the assessment resolution, enabling sensitive capture of subtle changes in the condition. Therefore, this method not only enables rapid initial diagnosis but is also suitable for quantitative evaluation of the efficacy of treatment before and after treatment in the same patient and for long-term continuous dynamic tracking of the condition, achieving a functional leap from static diagnosis to dynamic management. Tests have shown that the grading results of the method of this invention are highly consistent with the diagnostic consensus of senior physicians. Its objectivity and stability are significantly better than manual measurement methods. In particular, it can avoid the differences caused by the lack of experience of junior physicians, which proves the reliability and technical superiority of its clinical application.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for assisting the evaluation of fatty liver based on whole liver multi-dimensional features and machine learning model, characterized in that, Includes the following steps: Acquire CT image data of the target area, preprocess the CT image data, and segment out the three-dimensional mask of the liver and spleen; Based on the three-dimensional mask of the liver and spleen, multidimensional features related to the clinical and pathological classification of fatty liver are extracted from the CT image data of the target area. All extracted clinically and pathologically relevant multidimensional features related to fatty liver grading were integrated and standardized. A grading model was then constructed based on the standardized multidimensional features using a random forest classifier. A grading model was used to evaluate CT image data during the follow-up period to obtain auxiliary evaluation results.

2. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 1, characterized in that, The specific method for acquiring CT image data of the target area, preprocessing the CT image data, and segmenting the 3D mask of the liver and spleen is as follows: Acquire CT image data of the target area; The CT image data is linearly mapped using preset abdominal window width and window level to complete the window width and window level adjustment; The CT image data after window width and window level adjustment are normalized and resampled to a uniform spatial resolution. A pre-trained 3D U-Net was used as a segmentation model to process the resampled CT image data and segment out the 3D masks of the liver and spleen.

3. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 2, characterized in that, Pre-trained 3D U-Nets as segmentation models include: The encoder consists of 4–5 downsampling stages. Each stage contains two 3×3×3 convolutional layers, a ReLU activation function, and Batch Normalization, followed by 2×2×2 max pooling for downsampling. The decoder uses transposed convolution or interpolation to upsample each upsampling stage, and concatenates the feature maps of the corresponding stages of the encoder through skip connections, and then goes through two 3×3×3 convolutional layers. The output layer uses 1×1×1 convolution and the Softmax activation function to output a binary probability map of the liver and spleen.

4. The method for auxiliary assessment of fatty liver based on multidimensional features of the whole liver and a machine learning model according to claim 2, characterized in that, The training method for the pre-trained 3D U-Net as a segmentation model is as follows: Several manually annotated 3D masks of the liver and spleen were obtained as standard data; The default loss function is a combination of Dice Loss and cross-entropy loss; The standard data U-Net model is trained by randomly applying rotation, scaling, translation, and elastic deformation during training. The Adam optimizer is used, and the initial learning rate of the U-Net model is set to 1e-4. The learning rate is reduced to half every 50 rounds until the training result meets the requirements. Training is then stopped to obtain the pre-trained 3D U-Net.

5. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 1, characterized in that, Based on the three-dimensional mask of the liver and spleen, multidimensional features related to the clinical and pathological classification of fatty liver were extracted from the CT image data of the target area. These multidimensional features related to the clinical and pathological classification of fatty liver include: The ratio characteristics are the liver-spleen ratio and the liver-kidney ratio; Morphological characteristics include volume, sphericity and ellipsoidity, and discrete packing density; Global density features include mean density, median density, and mode density; Key interval characteristics include the proportion of low-density intervals, the proportion of high-density intervals, and the average density of specific intervals; The distribution shape characteristics are skewness and kurtosis.

6. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 1, characterized in that, The following is a specific method for integrating and standardizing all extracted clinically and pathologically relevant multidimensional features related to fatty liver grading: We acquire clinically and pathologically relevant multidimensional features related to fatty liver grading, integrate all clinically and pathologically relevant multidimensional features related to fatty liver grading into a high-dimensional feature vector, and use Z-score standardization to normalize the features and eliminate the influence of dimensions.

7. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 1, characterized in that, The specific method for constructing a gradation model based on a random forest classifier using standardized multidimensional features is as follows: Pre-set several pre-labeled tag data; Divide the label data into training and validation sets according to the required proportions; A random forest classification model is constructed, which consists of several decision trees. Subsamples are randomly drawn from the training set to construct each decision tree through bootstrapping. At the same time, some features are randomly selected for division when splitting nodes. During the training phase, the random forest classification model learns the ability of different features to distinguish fatty liver severity, and outputs the final classification result through a multi-decision-tree voting mechanism, while generating probability estimates for each category. A grid search combined with cross-validation is used to perform combined traversal of parameters including the number of decision trees, maximum tree depth, minimum number of split samples, minimum number of leaf node samples, and maximum number of features. K-fold cross-validation is used in the training set to evaluate the model performance indicators under different parameter combinations. The parameter combination with the best overall performance is selected as the final model configuration, which is the optimized gradation model. The optimized grading model is evaluated on the validation set. Once the validation is successful, it is used as the final grading model.

8. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 1, characterized in that, The following is a specific method for evaluating CT image data during the follow-up period using a grading model to obtain auxiliary evaluation results: CT image data during the follow-up period were obtained and fed into the grading model to obtain efficacy indicators. Based on the changes in efficacy indicators, auxiliary assessment results are obtained.

9. The method for auxiliary assessment of fatty liver based on whole liver multidimensional features and machine learning model according to claim 8, characterized in that, The calculation method for the auxiliary evaluation results is as follows: The change in rating is: ΔScore = Score_follow-up - Score_baseline Wherein, Score_follow-up is the efficacy index value, and Score_baseline is the efficacy index baseline value; If ΔScore ≤ -10, it is considered a significant improvement; If -5 ≥ ΔScore > -10, it is considered a slight improvement; If |ΔScore| < 5, the condition is considered stable; If ΔScore ≥ 5, it is considered as disease progression.

10. A fatty liver auxiliary assessment system based on whole liver multidimensional features and machine learning model, characterized in that, include: The preprocessing module is used to acquire CT image data of the target area, preprocess the CT image data, and segment out the three-dimensional mask of the liver and spleen. The multidimensional feature extraction module is used to extract clinically and pathologically relevant multidimensional features related to fatty liver grading from CT image data of the target area based on a three-dimensional mask of the liver and spleen. The grading model training module is used to integrate and standardize all extracted clinical and pathological multidimensional features related to fatty liver grading, and then construct a grading model based on the standardized multidimensional features using a random forest classifier. The model-assisted evaluation module is used to evaluate CT image data during the follow-up period using a grading model to obtain auxiliary evaluation results.

Images