Death reason discrimination method based on virtual anatomical image

Abstract

The invention discloses a death reason discrimination method based on a virtual anatomical image, and the method comprises the following steps: S1, data preprocessing: obtaining to-be-discriminated corpse CT image data, and carrying out the preprocessing; s2, feature extraction: inputting the preprocessed three-dimensional image data into a three-dimensional convolutional neural network model for feature extraction, and obtaining a feature vector through forward propagation of the three-dimensional convolutional neural network model; s3, judging a death reason: inputting the feature vector into a classifier for classification to obtain a probability value of each death reason category corresponding to the CT image data, and selecting the category with the highest probability value as a death reason; and S4, interpretability analysis: performing interpretability analysis on a death reason judgment result of the three-dimensional convolutional neural network model by using an LIME algorithm to generate a saliency map. According to the method, the features can be automatically extracted from the virtual anatomical image and the death reason is judged, so that the efficiency of forensic identification is improved.

Description

Technical Field

[0001] The present invention relates to the fields of forensic medicine and imaging, and particularly to a method for determining the cause of death based on virtual autopsy images. Background Art

[0002] In the investigation and judicial trial systems of cases, forensic examination and identification work plays a crucial role. In China, the number of cases examined and identified by forensic medicine each year is as high as millions. The work results play an irreplaceable role in key aspects such as determining the cause of death, analyzing the formation mechanism of injuries, and inferring the lethal tool, and are an essential and important part of the entire case detection and judicial trial process.

[0003] Currently, in forensic identification work, using imaging examination techniques such as CT to clarify the cause of death has become an important means. This technology has significant advantages such as non-invasiveness, rapidity, objectivity, repeatability, and the ability to display multi-layer information. In scenarios of detecting causes of death such as mechanical injury-induced death, asphyxiation death, and poisoning, CT virtual autopsy technology can effectively clarify the cause of death, reduce unnecessary autopsy work, and then quickly and properly handle cases, which is of great significance in improving the efficiency of handling cases and maintaining stability. However, since the introduction of CT detection technology into the field of forensic medicine is relatively recent, the majority of forensic technicians have limited mastery of this technology, and both the accuracy and speed of reading images need to be improved. This is mainly because CT images contain rich and complex human body structure information, which requires forensic technicians to have solid imaging knowledge and rich practical experience to accurately interpret, and this requires a large amount of time for learning and accumulation. This current situation severely restricts the comprehensive application of virtual autopsy technology in cause-of-death analysis. Summary of the Invention

[0004] The present invention mainly solves the technical problems existing in the prior art, such as the difficulty in interpreting CT virtual autopsy image data and low accuracy, and provides a method for determining the cause of death based on virtual autopsy images with high accuracy and low dependence on the individual capabilities of forensic technicians.

[0005] The present invention mainly solves the above technical problems through the following technical solutions: A method for determining the cause of death based on virtual autopsy images, comprising the following steps: S1. Data preprocessing: Obtain the data of the CT image (i.e., virtual autopsy image) of the body to be determined, and perform preprocessing to obtain the preprocessed three-dimensional image data; the preprocessing includes: (1) Stack the CT image data of each part along the Z-axis to form a unified three-dimensional image; (2) Adjust the size of the three-dimensional image to a preset size; (3) Perform intensity normalization on the three-dimensional image after adjusting the size; Data consistency and quality can be ensured through preprocessing; S2. Feature extraction: Input the preprocessed three-dimensional image data into a pre-trained three-dimensional convolutional neural network model for feature extraction. Through the forward propagation of the three-dimensional convolutional neural network model, a feature vector representing the features of the three-dimensional image data is obtained; the three-dimensional convolutional neural network model is a 3D-DenseNet121 model; S3. Cause-of-death discrimination: Input the feature vector obtained in step S2 into a Softmax classifier for classification, obtain the probability values of each cause-of-death category corresponding to the CT image data, and select the category with the highest probability value as the cause of death; the cause-of-death categories include drowning, sudden death, high fall, and mechanical asphyxia.

[0006] Preferably, the forward propagation of the three-dimensional convolutional neural network model includes: S201. Initial convolution: Use a three-dimensional convolutional layer with a convolution kernel size of 7×7×7 and a stride of 2 to perform preliminary feature extraction on the preprocessed three-dimensional image data; S202. Dense block processing: Input the data after preliminary feature extraction into several dense blocks in sequence. Each dense block contains several convolutional layers, and the output of each layer is connected to the input of all subsequent layers; S203. Transition layer processing: Use a transition layer between the dense blocks to control the number and size of the feature maps. The transition layer includes 1×1×1 convolution and average pooling; S204. Global average pooling: After the last dense block, use a global average pooling layer to convert the three-dimensional feature map into a one-dimensional feature vector; In the above process, batch normalization and ReLU activation function are used alternately.

[0007] Preferably, in step S1, the respective parts include the head, chest, and pelvis, and the CT image stacking time interval is (1.5, 1.5, 1.5).

[0008] Preferably, in step S1, adjusting the size of the three-dimensional image to a preset size specifically is: using the bilinear interpolation method to adjust the size of the three-dimensional image to (128, 128).

[0009] Preferably, the training of the three-dimensional convolutional neural network model includes the following steps: T1. Construct a training data set, and the training data set contains the CT images of corpses with known causes of death; T2. Perform data augmentation on the images in the training data set; T3. Using the training data set enhanced by the data, train the 3D-DenseNet121 model with cross entropy as the loss function, adopt the Adam optimizer, and the initial learning rate is 1e-5.

[0010] Preferably, a method for discriminating the cause of death based on virtual anatomical images further includes: S4. Interpretability analysis: Use the LIME (Local Interpretable Model-agnostic Explanations) algorithm to perform interpretability analysis on the discrimination results of the cause of death of the three-dimensional convolutional neural network model, generate a saliency map, and the saliency map is used to indicate the regions in the three-dimensional image data that have a significant impact on the discrimination results of the cause of death.

[0011] Preferably, the use of the LIME algorithm to perform interpretability analysis on the discrimination results of the cause of death of the three-dimensional convolutional neural network model includes the following steps: S401. Perturb the preprocessed three-dimensional image data X to generate a number of perturbed samples X'; S402. Input the multiple perturbed samples into the three-dimensional convolutional neural network model to obtain the discrimination results of the cause of death corresponding to each perturbed sample, denoted as y'; S403. Train a linear model based on the multiple perturbed samples and the corresponding discrimination results of the cause of death; S404. Based on the weights of the features in the linear model, determine the regions in the three-dimensional image data that have a significant impact on the discrimination results of the cause of death.

[0012] Preferably, the form of the linear model is: g(m)=β0+Σ n i=1 βᵢmᵢ where m i is the i-th feature, β0 is the intercept term, and βᵢ is the weight of the i-th feature; The training objective is to minimize the difference between g(m) and y', and the weighted least squares method is used: argmax β Σω(X’,X)(y’-g(m)) 2 where ω(X’,X) is the weight function, representing the similarity between the perturbed sample X’ and the original sample X, and its definition is: ω(X’,X)=exp(-\frac{d(X’,X) 2}{σ 2}) Where d(X’, X) is the Euclidean distance between X’ and X.

[0013] LIME quantifies the importance of each component in the linear model for the model prediction through the absolute value of the weight βᵢ of each component. The larger the absolute value, the higher the importance. This enables forensic experts to understand the basis for the 3D-DenseNet model's prediction and determine whether these bases are reasonable, thereby enhancing the credibility of the model. A positive βᵢ indicates that the presence of this component positively affects the prediction result (for example, making the model more inclined to predict a certain specific category). A negative βᵢ indicates that the presence of this component negatively affects the prediction result (for example, making the model more inclined to predict other categories). Whether βᵢ is positive or negative, as long as its absolute value is large, it indicates that this component has a significant impact on the prediction result.

[0014] The substantial effects brought by the present invention are as follows: (1) High degree of automation: The present invention can automatically extract features from virtual autopsy images and perform cause-of-death discrimination without manual intervention, improving the efficiency of forensic identification; (2) High accuracy: The present invention uses a deep learning model for feature extraction and classification, which can learn complex features in the images and improve the accuracy of cause-of-death discrimination; (3) Strong interpretability: The present invention uses the LIME algorithm to perform interpretability analysis on the discrimination results of the model, generating a saliency map, which can intuitively display the regions in the image that have an important impact on the discrimination results and enhance the credibility of the model; (4) Non-invasive: The present invention is based on the analysis of virtual autopsy images and does not require traditional autopsy, avoiding damage to the body. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flowchart of a method for discriminating the cause of death based on virtual autopsy images according to the present invention; Figure 2 is a schematic diagram of the 3D DenseNet121 network structure according to the present invention; Figure 3 is a loss curve graph of three different neural network models during the training process according to the present invention; Figure 4 is a receiver operating characteristic curve graph of the DensNet model on benign samples according to the present invention; Figure 5 is a precision-recall curve graph of the DensNet model on benign samples according to the present invention; Figure 6 is a picture of pulmonary hydrops effectively detected by LIME according to the present invention; Figure 7 is a picture of a fractured dental alveolus effectively detected by LIME according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] The technical solution of the present invention will be further specifically described below through embodiments in combination with the accompanying drawings.

[0017] Embodiment: A method for determining the cause of death based on virtual autopsy images in this embodiment is as Figure 1 shown, including the following steps: S1. Data preprocessing: Obtain the CT image data of the body to be determined and perform preprocessing to obtain the preprocessed three-dimensional image data; the preprocessing includes: (1) Stack the CT image data of each part along the Z-axis to form a unified three-dimensional image; (2) Adjust the size of the three-dimensional image to a preset size; (3) Perform intensity normalization on the three-dimensional image after adjusting the size; S2. Feature extraction: Input the preprocessed three-dimensional image data into a pre-trained three-dimensional convolutional neural network model for feature extraction, and through the forward propagation of the three-dimensional convolutional neural network model, obtain a feature vector representing the features of the three-dimensional image data; the three-dimensional convolutional neural network model is a 3D-DenseNet121 model; S3. Cause-of-death determination: Input the feature vector obtained in step S2 into a Softmax classifier for classification, obtain the probability values of each cause-of-death category corresponding to the CT image data, and select the category with the highest probability value as the cause of death; the cause-of-death categories include drowning, sudden death, high fall, and mechanical asphyxia; S4. Interpretability analysis: Use the LIME algorithm to perform interpretability analysis on the cause-of-death determination result of the three-dimensional convolutional neural network model, and generate a saliency map, which is used to indicate the region in the three-dimensional image data that has a significant impact on the cause-of-death determination result.

[0018] Since the original data may be taken from different CT imaging devices across the country, there are significant differences in the data structure and composition of each sample. Therefore, before inputting the data into the network for training, comprehensive preprocessing is required to unify the data format and structure. During the preprocessing process, for the head, chest, and pelvic data of the same individual, the bilinear interpolation method was used to stack the images on the Z-axis with a fixed spacing of (1.5, 1.5, 1.5); the number of images was adjusted according to the ratio of 25:40:30, and at the same time, the size of each image was adjusted to (128, 128) and intensity normalization was performed to construct the NIfTI format data of the complete sample, and the overall image space size was (128, 128, 95). This processing method not only helps to retain the important features of each anatomical region, but also enables the data to more comprehensively reflect the overall structural features of the human body. This comprehensive data input method will provide richer information for model training and cause of death determination, aiming to significantly improve the classification performance and accuracy of the model and provide a more reliable basis for related research.

[0019] This embodiment adopts the 3D DenseNet121 network structure as shown in Figure 2 . The input of the network is the 3D medical image data in NIfTI format constructed above, and the number of channels is set to 1, representing grayscale images, which is a typical form of forensic virtual autopsy images. After the image enters the network, it undergoes continuous convolution operations and processing by non-linear activation functions to extract meaningful features.

[0020] The adopted network is based on a densely connected convolutional neural network (CNN) and is specifically used for classifying and predicting the input 3D medical images. Different from the traditional convolutional network design, DenseNet directly passes the output of each layer to all subsequent layers through dense connections. This innovative structure significantly enhances the reuse of features and the propagation efficiency of gradients, thereby improving the learning ability of the network.

[0021] In this embodiment, the hyperparameters of the network model are appropriately adjusted to improve the performance of the model in classification tasks, and the advantages of DenseNet121 can be fully utilized to provide more accurate and reliable classification results for the research.

[0022] The core component of DenseNet121 is L dense blocks, each dense block contains multiple convolutional layers, and the output X of the u-th layer u (now the input X is a three-dimensional tensor with dimensions C×H×W×D, where D is the depth, H is the height, W is the width, and C is the number of channels) is directly connected to the input of all subsequent layers within this dense block, that is: X u+1=H u (concat(X0,X1,...,X u-1 )) This dense connection mechanism enables each layer to directly access the feature maps extracted by all previous layers, thus effectively reusing the features of early layers and avoiding information loss.

[0023] 3D inter-layer operation H u includes a series of operations for 3D data, such as 3D convolution, 3D batch normalization (3D-BN), and activation functions (such as ReLU), etc.

[0024] The size of the 3D convolution kernel is k d ×k h ×k w , the stride is S d ×S h ×S w , the padding is p d ×p h ×p w , the way of the 3D convolution operation is: Y = W’ * X + b where W’ is the 3D convolution parameter, with dimensions k d ×k h ×k w ×C in ×C out (C in is the number of input channels, C out is the number of output channels), X is the input 3D tensor, * represents the convolution operation, b is the bias term, and Y is the result after convolution. In the network, the convolution kernel size is set to 7×7×7, which has a large receptive field to extract initial global features and ensure that the network can capture important information in the image. The stride is set to 2, which can reduce the input size while retaining important feature information, balancing feature extraction and computational efficiency.

[0025] The network adopts batch normalization, which helps to stabilize the training process and accelerate the convergence speed of the model. 3D-BN normalizes each channel of the 3D tensor independently, and the formula is: 3D-BN(Y) = γ(Y - μ) / sqrt(σ 2 + ε) + β where μ and σ 2 are the mean and variance of the input 3D tensor of this layer on each channel respectively, γ and β are learnable parameters, and ε is a very small constant to prevent the denominator from being zero.

[0026] This network obtains the final output X through the activation function (ReLU) u+1 ,in X u+1 =RELU(3D-BN(Y)) 3D-DenseNet also contains transition layers to control the number of feature maps and the size of the three-dimensional space. A transition layer is used between dense blocks, which is mainly composed of 1×1×1 convolution and pooling layers. The convolution layer is used to reduce the number of channels in the feature map, thereby controlling the parameter scale of the model and reducing the computational complexity. Next, an average pooling layer with a common step size of 2 is used to further reduce the spatial size of the feature map through downsampling, compressing the data while retaining important features.

[0027] The three-dimensional 1×1×1 convolution is used to reduce the number of channels of the feature map. The number of input channels is C in , the number of output channels is C out , the three-dimensional convolution operation can be expressed as: Y 1×1×1 =W 1×1×1 *X+b 1×1×1 Among them, W 1×1×1 is the three-dimensional convolution kernel parameter, with dimensions of 1×1×1×C in ×C out , X is the input 3D tensor.

[0028] At the end of the network, a global average pooling layer is used to reduce the three-dimensional feature map to a one-dimensional vector, retaining the global average of each feature map. Compared with the traditional fully connected layer, this method can effectively reduce the number of model parameters and prevent overfitting, especially when there are relatively few 3D medical data samples. This method is particularly effective.

[0029] The pooling kernel size is k dpool ×k hpool ×k wpool , the step length is S dpool ×S hpool ×S wpool , the three-dimensional average pooling operation formula is: Y pool =1 / (k dpool ×k hpool ×k wpool )Σ kdpool-1 i=0 Σ khpool-1 i=0 Σ kwpool-1 i= 0X i+Sdpoolm,j+Shpooln,k+Swpoolo Among them, m, n, and o are the position indices at which the pooling window slides on the three-dimensional feature map.

[0030] After global pooling, the network outputs a vector with the same number of elements as the number of classes, that is, the probability distribution for each class. In this embodiment, the number of output channels of the network is set to be the same as the number of classes required for the cause-of-death classification task, and the Softmax activation function is used to generate the predicted probability for each class. p = softmax(Y pool ) = exp(Y pool ) / Σ K k=1 exp(Y poolk ) where p is a vector of length K, and p k represents the probability that the sample belongs to the k-th class, and Σ K k=1 p k=1 = 1. Through this structural design, the model can more accurately reflect the characteristics of different causes of death, providing strong support for practical applications.

[0031] When training the 3D DenseNet model, the cross-entropy loss function is used. The cross-entropy loss function is a loss function commonly used in multi-classification tasks, and its role is to evaluate the degree of inconsistency between the probability distribution output by the model and the actual class. It is proposed based on the entropy concept in information theory and aims to measure the distance between two probability distributions. For the task of cause-of-death classification, the goal is to minimize the cross-entropy between the class probability distribution predicted by the model and the true label distribution, so that the model can more accurately predict the cause-of-death category.

[0032] When training 3D DenseNet for the classification task, the cross-entropy loss function is used. Let the true label of the sample be y (one-hot encoded), and the predicted output of the model be y'. The cross-entropy loss function can be expressed as: L = -(1 / N)∑ N n=1 ∑ N k=1 [y n,k log(y' n.k )], where N is the total number of classes, which is 4 in this embodiment. The smaller the value of the cross-entropy loss, the closer the probability distribution predicted by the model is to the true class. The goal is to make the value of the loss function approach 0 as much as possible through optimization, which means that the model can accurately predict the correct class. The cross-entropy loss function combines the log-softmax operation, which can prevent numerical underflow or overflow problems during the calculation process, can well handle the classification problem of multiple cause-of-death categories in this project, helps the model converge quickly during training, and find the optimal parameter settings.

[0033] The Adam optimizer is used in the training settings. It is an efficient adaptive gradient descent algorithm that can dynamically adjust the learning rate of each parameter to adapt to the learning requirements of different features. The initial learning rate is set to 1e-5, and this learning rate is continuously adjusted during the training process according to the performance of the validation set to ensure the best training effect of the model. At the same time, in this embodiment, the cross-entropy loss function (Cross Entropy Loss) is adopted, and its specific details have been described in detail in the previous text.

[0034] During the training process, in order to effectively monitor the performance of the model, the average loss value of each epoch and the AUC (area under the curve) index of the validation set are recorded. These data can reflect the change trend of the training loss and the fluctuation of the AUC index in real time, so as to judge the convergence situation and performance improvement of the model. As the training progresses, the training loss should gradually decrease, indicating that the model is gradually learning and refining more accurate features; while the AUC value of the validation set gradually increases, reflecting the enhancement of the model's classification ability and pointing to the final training goal.

[0035] This embodiment adopts LIME (Local Interpretable Model-agnostic Explanations) to explain the 3D-DenseNet model method, helping or reminding forensic experts to understand the model decision-making process and the characteristics of the cause of death in cadaveric CT images.

[0036] LIME generates a series of new samples by perturbing the input features in the local region of the input sample X. Then, these samples and their corresponding DenseNet model prediction results are used to train a simple interpretable model (usually a linear model). Finally, this interpretable model is used to explain the prediction of the DenseNet model at the sample X.

[0037] Let Ω be the set of all possible perturbed samples. For the input sample X, perturbed samples X’∈Ω are generated by randomly masking (masking) its features. Specifically, a binary mask vector m is defined, whose length is the same as the number of features of X (i.e., n = C×H×W×D), where m i ∈{0,1} indicates whether the i-th feature is retained (m i =1) or masked (m i =0). Then the perturbed sample X’ can be expressed as: X’ i =m i X i ; where X iis the i-th feature of X. Using the DenseNet model M, the prediction result y’ = M(X’) is obtained. Then, these perturbed samples and their prediction results are used to train a linear model g, which has the form: g(m)=β0+Σ n i=1 β i m i where β0 is the intercept and β i is the weight of the i-th feature. The training objective is to minimize the difference between g(m) and y’, using weighted least squares: argmax β Σω(X’,X)(y’ - g(m)) 2 ; where ω(X’,X) is the weight function, representing the similarity between the perturbed sample X’ and the original sample X, and is defined as ω(X’,X)=exp(-\frac{d(X’,X) 2}{σ 2}); where d(X’,X) is the Euclidean distance between X’ and X.

[0038] The following table shows the performance of the existing ResNet, SENet, FullyConnectedNet models and the DenseNet of this embodiment in the head-chest-pelvis stacked model data, reflecting the classification performance differences of the models in different categories. Model type Accuracy Sensitivity Specificity Precision F1-score ResNet 0.927 0.889 0.985 0.939 0.912 SENet 0.846 0.804 0.964 0.804 0.804 FullyConnectedNet 0.590 0.372 0.801 0.442 0.400 DenseNet 0.976 0.914 1 1 0.952

[0039] It can be seen from the table data that there are differences in the prediction situations of different models in each classification (DrowningDeath (drowning), HighFall (high fall), MechanicalAsphyxia (mechanical asphyxia), SuddenDeath (sudden death)), and some models have more misjudgments in individual classifications. The overall performance of DenseNet in each category is the most excellent, with an average precision of 100% and an average recall rate of 91.4%, significantly better than other models.

[0040] Specific to the classification effects of each category, in the classification of falls from height and sudden death, the DenseNet in this embodiment performs relatively well, with an accuracy of 0.976, a sensitivity of 0.914, a specificity and precision of 1, and an F1-score of 0.952, showing obvious advantages in accurate classification and comprehensive performance, indicating its strong learning and recognition capabilities. In the classification of drowning deaths, the precision is 100% and the recall rate reaches 85.7%. This shows that the model can effectively identify most real cases, with an F1 score of 0.921. In the classification of mechanical asphyxia, the precision of the model is 100% and the recall rate is 80%.

[0041] Figure 3 The figure shows the Loss Curve graph, which demonstrates the changes in the loss values of three different neural network models (SENet154, ResNet101, DenseNet121) during the training process with respect to the number of training epochs. As can be seen from the graph, as the number of training epochs increases, the loss values of the three models generally show a downward trend, indicating that as the training progresses, the models are continuously learning and their fitting ability to the training data is gradually enhanced. However, comparatively, the loss value of the DenseNet121 model decreases relatively quickly and drops to a lower level at an earlier training epoch, suggesting that it can quickly learn the feature patterns in the data at the initial stage of training, and its convergence speed is better than that of the other two models. The loss value of DenseNet121 is the lowest, reaching a relatively stable and low level, which means that this model has a good fitting effect on the training set; the loss values of SENet154 and ResNet101 are relatively high and still show some fluctuations in the later stage of training, indicating that these two models are more sensitive to the data during the training process or are prone to falling into local optimal solutions during the optimization process, and their performance is not as stable as that of DenseNet121.

[0042] Figure 4 and Figure 5 The Receiver Operating Characteristic Curve (ROC) and Precision-Recall Curve (PR) of the DensNet model proposed in this study on benign samples are presented to show the performance of the model on each individual category. The blue, orange, green, and red curves in the figure correspond to the ROC curves and PR curves of drowning, high fall, mechanical asphyxia, and sudden death categories respectively.

[0043] Regarding the PR curve, the area under the curve (AUC) values for the high fall and mechanical asphyxiation categories both reached 1.00, demonstrating nearly perfect performance, meaning that the model could maintain a very high precision at different recall levels; the AUC values for the drowning and sudden death categories were 0.99 and 0.98 respectively, also showing excellent performance. In terms of the ROC curve, the AUC values for the high fall and mechanical asphyxiation categories were 1.00, indicating that the model had a very strong ability to distinguish positive and negative examples of these two categories; the AUC values for the drowning and sudden death categories were 0.99, also reflecting excellent performance.

[0044] Generally speaking, the above results show that the DensNet model presents good performance in accurately identifying positive examples of each classification and its generalization ability under different threshold settings, providing a valuable reference for the research and practical application of the cause-of-death classification of body CT.

[0045] In this embodiment, LIME is used as an interpretable artificial intelligence method to generate local and global explanations for the predictions of the proposed model on the test dataset. The image segmentation created by LIME highlights the key regions that contribute to classification.

[0046] Drowning deaths and high falls usually have obvious characteristic regions. For example, the lungs of a drowned body often have a large amount of accumulated water, while the skull of a high-fall body will have obvious fractures, which are key features for the classification of the cause of death of the body. Figure 6 and Figure 7 Pictures showing the accumulated water in the lungs and the fracture of the dental alveolus effectively detected by LIME are presented. It should be noted that the pink-highlighted parts accurately represent the main regions in the pictures of breast cancer, colon cancer, and lung cancer, while the blue regions indicate that there are no significant features in the pictures of normal tissues. In addition, LIME analysis confirmed that the accumulated water in the lungs and structures in the skull such as the fracture of the dental alveolus are the main factors for classification, which is consistent with clinical findings. This solution creates a visual representation of LIME values to graphically show the pixels that have the greatest impact on the model's predictions. The classification basis for mechanical asphyxiation and sudden death is not obvious in CT images, but it can still provide important reference for forensic judgment. These visual representations enable forensic practitioners to better understand the model's predictions and provide valuable insights for the decision-making process.

[0047] The method proposed in this embodiment, which combines the 3D-DenseNet network and LIME technology, shows good performance in the cause-of-death classification task of forensic virtual autopsy images, has the ability to obtain effective features for accurate classification, and the constructed dataset provides strong support for model training. In different cause-of-death category classification tasks, the model shows differences in performance.

[0048] From the loss analysis of the model training process, the 3D-DenseNet121 model can quickly learn data features in the initial stage of training. The loss value drops rapidly and finally reaches a low and stable level. Compared with the SENet154 and ResNet101 models, it has a faster convergence speed and a better fitting effect on the training set. This indicates that the 3D-DenseNet121 network structure is more suitable for processing forensic virtual autopsy image data, can more effectively learn the patterns in the data, and lays a foundation for accurate classification. The application of the LIME technique provides an intuitive explanation for the model's decision-making. By generating saliency maps, LIME can highlight key regions that have an important impact on classification, such as the water accumulation area in the lungs of a drowned body and the fracture area of the skull of a body that died from a high fall, etc., helping forensic doctors better understand the model's decision-making process and enhancing the credibility and practicality of the model in actual forensic work.

[0049] The specific embodiments described in this article are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described specific embodiments or use similar ways to substitute them, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0050] Although terms such as preprocessing and feature extraction are used more frequently in this article, the possibility of using other terms is not excluded. The use of these terms is only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional limitation is contrary to the spirit of the present invention.

Claims

1. A method for determining the cause of death based on virtual autopsy images, characterized in that It includes the following steps: S1. Data preprocessing: Obtain the CT image data of the body to be judged and perform preprocessing to obtain the preprocessed three-dimensional image data; the preprocessing includes: (1) Stack the CT image data of each part along the Z-axis to form a unified three-dimensional image; (2) Adjust the size of the three-dimensional image to a preset size; (3) Perform intensity normalization on the three-dimensional image with the adjusted size; S2. Feature extraction: Input the preprocessed three-dimensional image data into a pre-trained three-dimensional convolutional neural network model for feature extraction. Through the forward propagation of the three-dimensional convolutional neural network model, a feature vector representing the features of the three-dimensional image data is obtained; the three-dimensional convolutional neural network model is a 3D-DenseNet121 model; S3. Cause-of-death discrimination: Input the feature vector obtained in step S2 into a Softmax classifier for classification to obtain the probability values of each cause-of-death category corresponding to the CT image data, and select the category with the highest probability value as the cause of death; the cause-of-death categories include drowning, sudden death, high fall, and mechanical asphyxia.

2. The method for determining the cause of death based on virtual autopsy images according to claim 1, characterized in that, The forward propagation of the three-dimensional convolutional neural network model includes: S201. Initial convolution: Use a three-dimensional convolutional layer with a convolution kernel size of 7×7×7 and a stride of 2 to perform preliminary feature extraction on the preprocessed three-dimensional image data; S202. Dense block processing: Sequentially input the data after the preliminary feature extraction into several dense blocks. Each dense block contains several convolutional layers, and the output of each layer is connected to the input of all subsequent layers; S203. Transition layer processing: Use a transition layer between the dense blocks to control the number and size of the feature maps. The transition layer includes a 1×1×1 convolution and average pooling; S204. Global average pooling: After the last dense block, use a global average pooling layer to convert the three-dimensional feature map into a one-dimensional feature vector; In the above process, batch normalization and ReLU activation function are used alternately.

3. A method for determining the cause of death based on virtual autopsy images according to claim 1, characterized in that, In step S1, each part includes the head, chest, and pelvis, and the CT image stacking time interval is (1.5, 1.5, 1.5).

4. The method for determining the cause of death based on virtual autopsy images according to claim 3, characterized in that, In step S1, adjusting the size of the three-dimensional image to a preset size specifically means: using the bilinear interpolation method to adjust the size of the three-dimensional image to (128, 128).

5. A method for determining the cause of death based on virtual autopsy images according to claim 1, characterized in that, The training of the three-dimensional convolutional neural network model includes the following steps: T1. Construct a training data set, and the training data set contains the CT images of the body with known causes of death; T2. Perform data augmentation on the images in the training data set; T3. Use the augmented training data set to train the 3D-DenseNet121 model with cross-entropy as the loss function, and adopt the Adam optimizer with an initial learning rate of 1e-5.

6. A method for determining the cause of death based on virtual autopsy images according to any one of claims 1 to 5, characterized in that, It also includes: S4. Interpretability analysis: Use the LIME algorithm to perform interpretability analysis on the cause-of-death discrimination results of the three-dimensional convolutional neural network model to generate a saliency map, and the saliency map is used to indicate the regions in the three-dimensional image data that have a significant impact on the cause-of-death discrimination results.

7. A method for determining the cause of death based on virtual autopsy images according to claim 6, characterized in that, The interpretability analysis of the death cause discrimination result of the 3D convolutional neural network model using the LIME algorithm includes the following steps: S401. Perturb the preprocessed 3D image data X to generate a number of perturbed samples X'; S402. Input the multiple perturbed samples into the 3D convolutional neural network model to obtain the death cause discrimination result corresponding to each perturbed sample, denoted as y'; S403. Train a linear model based on the multiple perturbed samples and the corresponding death cause discrimination results; S404. Determine the regions in the 3D image data that have a significant impact on the death cause discrimination result based on the weights of the features in the linear model.

8. A method for determining the cause of death based on virtual autopsy images according to claim 7, characterized in that, The form of the linear model is: g(m)=β0+Σ n i=1 βᵢmᵢ where m i is the i-th feature, β0 is the intercept term, and βᵢ is the weight of the i-th feature; The training objective is to minimize the difference between g(m) and y', using weighted least squares: argmax β Σω(X’,X)(y’ - g(m)) 2 where ω(X',X) is the weight function, representing the similarity between the perturbed sample X' and the original sample X, and its definition is: ω(X’,X)=exp(-\frac{d(X’,X) 2}{σ 2}) where d(X',X) is the Euclidean distance between X' and X.

Images