Death time inference method based on base model full-slice probability mapping graph

Abstract

The invention relates to the field of forensic medicine, in particular to a death time inference method based on a full-slice probability mapping graph of a base model, which comprises the following steps of: digitally acquiring a dead full-slice image through a high-resolution digital slice scanning system; performing preprocessing and quality control on the dead full-slice image, and then constructing a high-quality image data set capable of entering a model information propagation process; performing fine tuning training on the large-scale self-supervised pre-training base model configured with the attention mechanism on a high-quality image data set; and predicting the dead full slice image through the fine-tuned base model, drawing a predicted global in-situ category probability mapping graph, observing the brightness of different categories of mapping graphs, and making final death time category inference. According to the method, the defects of high instrument cost, high professional skill requirement, difficulty in popularization and the like are overcome, the death time output by the model after operation can be obtained only by acquiring and inputting a microscopic image of a specific tissue of a dead body by a user, and manpower and material resources are saved.

Description

Technical Field

[0001] The present invention relates to the field of forensic medicine, and specifically to a method for inferring the time of death based on the whole-slide probability mapping diagram of the base model. Background Art

[0002] The inference of the forensic time of death relies on the temporal law of postmortem changes. During the axial evolution process in the time dimension, the degree of postmortem changes that have occurred gradually deepens, and at the same time, new postmortem changes are continuously superimposed, forming a complex and dynamic interaction system between the organism and the inorganic environment. Analyzing the evolution law of this system in the time dimension is the key problem that needs to be solved urgently to achieve more accurate inference of the time of death.

[0003] Histopathological examination takes into account macroscopic visual perception at the microscopic scale and is one of the most important means for analyzing postmortem changes in forensic practice. However, at the present stage, the manual examination method of histopathological images in the process of forensic identification practice has obvious limitations such as strong subjectivity, low reading efficiency, and large errors. The rapid development of computational pathology technology has prompted the forensic community to actively explore automated whole-slide image analysis technology to assist or partially replace traditional manual histopathological examinations, reduce the labor burden of staff, and bring new opportunities and challenges for more rapid and efficient postmortem histopathological assessment.

[0004] In recent years, base models pre-trained by self-supervision on a large-scale data basis have shown powerful general feature learning capabilities, and their feature extraction and transfer generalization performance in downstream tasks are particularly prominent, and they are gradually becoming important tools in the field of computational pathology. However, the analysis methods of computational pathology do not cover enough features of postmortem tissue images. The present patent invention aims to broaden the application boundary of the visual attention base model, take the inference of the time of death as the leading scenario, and introduce advanced analysis technologies of computational pathology into the intelligent analysis of postmortem tissue images. Summary of the Invention

[0005] In order to solve the problems of obvious limitations such as strong subjectivity, low reading efficiency, and large errors in the manual examination method of histopathological images in the process of forensic identification practice at the present stage, the present invention provides a method for inferring the time of death based on the whole-slide probability mapping diagram of the base model.

[0006] The present invention is realized through the following technical solutions: A method for inferring the time of death based on the whole-slide probability mapping diagram of the base model includes the following steps:

[0007] S1. Digitally acquire the postmortem whole-slide image through a high-resolution digital slide scanning system;

[0008] S2. After preprocessing and quality control of the postmortem whole-slide image, construct a high-quality image dataset that can enter the model information propagation process;

[0009] S3. Fine-tune and train the large-scale self-supervised pre-training base model with an attention mechanism on a high-quality image dataset;

[0010] S4. Use the fine-tuned base model to predict the postmortem whole-slide images, draw the predicted global in-situ class probability mapping diagram, and make the final inference of the postmortem time category by observing the brightness of the mapping diagrams of different categories.

[0011] As a further improvement of the technical solution of the present invention, step S1 specifically includes the following steps:

[0012] (1) Extract body tissues at different postmortem time points, prepare H&E sections, and obtain postmortem whole-slide images through a high-resolution digital pathology image scanning system;

[0013] (2) The postmortem whole-slide images contain postmortem change information, from which special postmortem morphological changes can be observed, showing significant differences from the microscopic morphology of pre-mortem tissues;

[0014] (3) Divide the postmortem time into continuous time periods according to time points and use this as the basis for class annotation of postmortem whole-slide images.

[0015] As a further improvement of the technical solution of the present invention, step S2 specifically includes the following steps:

[0016] (1) Read the postmortem whole-slide images as a memory object with a pyramid structure configured with the pixel size, magnification, downsampling factor, and storage path of the whole-slide image;

[0017] (2) At the highest resolution layer without downsampling, crop the postmortem whole-slide images into Patch images and synchronously record the position information of the Patches in the postmortem whole-slide images;

[0018] (3) Evaluate the quality of the cropped Patch images, and eliminate low-quality Patches that are blurred, folded, or have too large a background;

[0019] (4) Separate the global color space of the image into an intensity space and a substrate space by non-negative matrix factorization of the color features of the Patch images, and perform directional mapping of the global color space of the image to the staining space of the reference image to achieve color normalization, so as to eliminate the influence of staining differences between different sections on the analysis results.

[0020] As a further improvement of the technical solution of the present invention, step S3 specifically includes the following steps:

[0021] (1) Label the postmortem time tags for the quality-controlled Patch images, and regard the postmortem whole-slide images as a package of Patch images containing postmortem time tag annotations;

[0022] (2) Input the preprocessed and quality-controlled Patch images into the base model for fine-tuning training;

[0023] (3) The base model is based on a large visual Transformer architecture and combines the attention mechanism to extract features from images;

[0024] (4) The base model is pre-trained in a self-supervised manner on a high-quality image dataset;

[0025] (5) In the forward propagation process of the base model, the Patch images are first divided into non-overlapping tiles of a fixed size, flattened into feature vectors, embedded into a fixed length, and then converted into a two-dimensional feature matrix through a stacking operation;

[0026] (6) An equal-length learnable [CLS] vector is stacked at the beginning of the matrix to form a new feature matrix. Self-attention mechanism operations are performed in multiple layers of Transformer encoders, and the feature matrix is updated according to the attention scores. Each feature vector of the new feature matrix is represented by weighted stacking through the remaining feature vectors to ensure the context association between each feature vector;

[0027] (7) The [CLS] embedding vector is intercepted from the updated feature matrix through matrix slicing operations and input into a classification head composed of fully connected layers to map the vector into a probability distribution representation with a length equal to the number of target categories;

[0028] (8) During the process of evaluating the performance of the model adjustment, ablation logic is used to compare the performance of ViT-Tiny, ViT-Large, and the base model to confirm the effectiveness of the base model in pre-training on large-scale histological images, which is more suitable for feature extraction and analysis of tissue images.

[0029] As a further improvement of the technical solution of the present invention, step S4 specifically includes the following steps:

[0030] (1) Create C all-zero matrices, where C is the number of predicted time-of-death categories, and the matrix size is calculated by taking the integer after dividing the original size of the whole-slide image after death by the size of the cropped Patch image;

[0031] (2) Parse the position information attached to the Patch images to convert the coordinates, and fill the corresponding positions in the all-zero matrix with the corresponding predicted probability values to obtain the probability matrix at the whole-slide image level;

[0032] (3) According to the mapping relationship between the probability level and the gray scale, generate gray scale mapping diagrams for the probability matrices of each category respectively. The probability value range of each category is 0-1. When the probability value is close to 0, the pixel gray scale at the corresponding position in the mapping diagram is darker, and when it is close to 1, the pixel drawing gray scale is brighter.

[0033] (4) Since the postmortem changes are global changes in tissue morphology, the predicted category of the postmortem time corresponding to the majority of Patch images in the postmortem whole-slide images is the final predicted category of the postmortem time for the postmortem whole-slide images.

[0034] The method for inferring the postmortem time based on the whole-slide probability mapping diagram of the base model provided by the present invention has the following advantages compared with the prior art:

[0035] (1) The present invention applies a vision Transformer model configured with an attention mechanism, especially an advanced base model that is self-supervised pre-trained on large-scale pathological image data, to postmortem pathological image analysis, and further derives a method for inferring the postmortem time range based on this. Since the base model has been pre-trained, its feature extraction ability is more powerful and can capture the subtle features of postmortem tissue changes.

[0036] (2) Compared with the methods for estimating the postmortem interval (PMI) pointed out in a large number of current studies, such as the methods based on genomics, proteomics, and metabolomics, the present invention overcomes the disadvantages of high instrument cost, high professional skill requirements, and difficulty in popularization. It only requires the user to obtain the microscopic images of specific tissues of the deceased body and input them, and then the postmortem time output by the model after calculation can be obtained, saving manpower and material resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0039] Figure 1 It shows the flow of the method for inferring the postmortem time of the vision base model probability mapping diagram.

[0040] Figure 2 It shows the microscopic histological manifestations at different postmortem times.

[0041] Figure 3 It shows the prediction confusion matrix of 16 external validation samples.

[0042] Figure 4 Indicates the predicted probability mapping diagram of the example sample. Specific implementation mode

[0043] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the solution of the present invention will be further described below. It should be noted that, without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0044] Many specific details are set forth in the following description in order to fully understand the present invention, but the present invention can also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0045] The specific embodiments of the present invention will be described in detail below.

[0046] The following will take the inference of the postmortem interval by combining the histological images of porcine liver after death with deep learning as a specific embodiment to describe in detail the technical solutions involved in the present invention. The specific implementation steps are as follows:

[0047] 1. Preparation of tissue sections and acquisition of whole slide images (whole slide images after death):

[0048] A total of 10 healthy female Bama minipigs, 6 months old and weighing 30 kg, were purchased. After being euthanized by injecting an overdose of sodium pentobarbital, they were placed in a climate chamber with a temperature of 16°C and a humidity of 45%.

[0049] One piece of liver tissue was extracted from the cadaver at 6 h, 24 h, 48 h, and 96 h after death in a continuous sampling manner from the same sample. The postmortem interval was divided into four time period intervals according to time points: 0-6 h, 6-24 h (excluding 6 h), 24-48 h (excluding 24 h), 48-96 h (excluding 48 h), and this was used as the basis for the class annotation of the whole slide images. A total of 10 pieces × 4 time points = 40 tissue sections (H&E sections) were obtained.

[0050] After being scanned by the TissueFax digital pathology slide scanning system, whole slide images in Tiff format were obtained. The microscope hardware parameters were ZEISS Plan-APOCHROMAT objective lens; Medium type: Air; Numerical Aperture: 0.95; Field of View Diameter: 160 μm; Resolution: 0.14 μm / pixel. A total of 10 pieces × 4 time points = 40 whole slide images after death were obtained.

[0051] Morphological changes such as tissue autolysis, widened tissue spaces, and changes in nuclear characteristics (such as nuclear pyknosis, fragmentation, and dissolution) can be observed postmortem, which deepen with the progression of the postmortem interval. These changes form unique characteristics that distinguish postmortem tissues from premortem tissues, constituting the theoretical basis for estimating the postmortem interval through image analysis (as Figure 2 shown).

[0052] 2. Preprocessing of postmortem whole-slide images:

[0053] (1) Cropping and quality control of postmortem whole-slide images

[0054] Read the postmortem whole-slide image as a memory object with a pyramid structure, configured with attributes such as the pixel size, magnification, downsampling factor, and storage path of the whole-slide image.

[0055] At the highest resolution layer without downsampling, crop the postmortem whole-slide image into Patch images with a fixed pixel size, and synchronously save the position information of the Patch images relative to the original whole-slide image during the cropping process.

[0056] Perform quality control on the Patch images, and screen and remove blurred (contaminated images), folded, and images with excessive backgrounds.

[0057] Contaminated images refer to images with dark contaminated areas, manifested as dark gray areas with a large proportion; while images with excessive backgrounds refer to images with a large proportion of bright gray areas.

[0058] Convert the color Patch images to grayscale images and set pixel gray thresholds. Pixels below a certain threshold are recorded as contaminated pixels, and pixels above a certain threshold are recorded as background pixels. When the proportion of contaminated pixels or background pixels exceeds their respective proportion thresholds, this image is considered a substandard image and is excluded.

[0059] In this embodiment, the images are cropped into Patch images of 512×512 size, the background gray threshold is set to 250, the contaminated pixel gray threshold is set to 50, the background proportion threshold is set to 80%, and the contamination proportion threshold is set to 25%.

[0060] (2) Image color normalization

[0061] Separate the global color space of the image into an intensity space and a substrate space by non-negative matrix factorization of the color features of the image, and perform directional mapping of the color space of the target image to the staining space of the reference image, thereby achieving color normalization.

[0062] Specific implementation steps: First, convert the image to be normalized and the reference image from the pixel space of each RGB channel through the formula to the logarithmic optical density space, where I is the pixel value, I 0 is the light source intensity, and in the RGB mode, I0 is the maximum pixel value of each color channel of the image.

[0063] Then, through the non-negative matrix factorization algorithm, the staining matrix matrix and the staining intensity matrix of the image to be normalized and the reference image are extracted respectively.

[0064] The color information of each channel of the image can be represented by the formula M = W·H, where M is the OD matrix, W is the staining matrix matrix, and H is the staining intensity matrix. The staining matrix matrix of the extracted reference image is represented by W ref is represented, and the staining intensity matrix of the image to be normalized is represented as H src , and multiplying it with the staining matrix matrix W ref can obtain the OD space representation M of the normalized image norm , which is represented by the formula as M norm = W ref ·H src . Finally, restoring M norm from the OD space back to the RGB space can obtain the normalized image of the image to be normalized in the approximate space of the reference image.

[0065] 3. Transfer fine-tuning of the vision base model on the postmortem histological image dataset

[0066] (1) Dataset division:

[0067] At the sample level, the training set and the validation set are divided. The Patch images of the same postmortem whole-slide image only appear in one part of the training set or the test set, and cannot appear in both (the training set and the test set) at the same time;

[0068] Each postmortem whole-slide image is regarded as a bag composed of Patch images, and the instance class labels are shared with the class labels of the postmortem whole-slide image;

[0069] In this embodiment, the training set and the validation set are divided according to the ratio of 8:2. As a result, the Patch images under the whole-slide images of the livers of 8 pigs after death are used for model training, and the Patch images under the whole-slide images of the livers of 2 pigs after death are used for model validation.

[0070] (3) Transfer fine-tuning training process:

[0071] The base model refers to a large-scale pre-trained model based on Vision Transformer. Its components include an image chunking and embedding layer, a Transformer encoder layer, and an output layer. The model is self-supervised pre-trained on the whole-slide image dataset of the TCGA database through the DINO algorithm and the model weights are initialized to ensure that the base model learns rich representations of histological images.

[0072] The fine-tuning training of the base model includes two data propagation processes: forward and backward.

[0073] During the forward propagation of the model, the base model first divides the Patch image into non-overlapping tiles of a fixed size. Assuming the size of the input Patch image is h×w×c, where h is the height, w is the width, and c is the number of channels (c = 3 for RGB images), if the size of each tile is, according to the formula it is calculated that each Patch image can be divided into N tiles. Then each tile is flattened into a vector of P×P×c dimensions, and through the embedding layer, the representation vector of each tile is projected into a feature vector of a fixed length D (i.e., the input dimension D of the visual base model). Then, the learnable D-dimensional [CLS] embedding vector is stacked at the beginning of the matrix. Finally, each tile will be converted into a feature matrix of shape (N + 1)×D dimensions before entering the Transformer encoding module.

[0074] In the Transformer encoder of the model, the input (N + 1)×D-dimensional feature matrix undergoes a linear transformation Q = XW Q , K = XW K , V = XW V to generate three groups of matrices: Q (query), K (key), and V (value). Among them, W Q , W K , W V are learnable weight matrices.

[0075] Through the feature matrix is updated, where is used to obtain the attention weight matrix, whose dimension is (N + 1)×(N + 1). Among them, Q is the query matrix mentioned above, K is the key matrix, and K T represents the transpose matrix of the key matrix, D k is the dimension of the feature vector corresponding to each tile, usually a preset constant, and is used to normalize the product of Q and K T for gradient stability. Multiplying the attention weight matrix by the V matrix can achieve the update of the feature matrix, and the output dimension is still (N + 1)×D. Then, through the residual connection and the normalization layer, the input and output are integrated to alleviate the vanishing gradient and improve the stability and efficiency of the fine-tuning training.

[0076] After stacking multiple Transformer encoders, an updated (N + 1)×D-dimensional feature matrix is output. The [CLS] embedding vector is extracted from the feature matrix through matrix slicing and input into a classification head composed of fully connected layers, mapping the D-dimensional feature representation to the number of target categories, and outputting the probability distribution under each category and the predicted category label of the Patch image death time.

[0077] During the backpropagation process, according to the formula calculate the loss value between the predicted probability vector and the true class, and then through the chain rule and the Adam optimizer, backpropagate the loss from the output layer to each layer of the model. Calculate the parameter gradients layer by layer and update the parameters until the set number of iteration rounds is reached. Where L is the loss function value, N is the total number of samples entering the model for prediction in each batch during the backpropagation process, C is the number of predicted time-to-death categories, y ij is the one-hot encoding of the true class label of the i-th predicted instance, is the probability that the model predicts the i-th predicted instance as the j-th class, is the logarithm of the predicted probability.

[0078] During the model fine-tuning process, dynamically monitor the decrease in validation loss and the increase in accuracy during the model fine-tuning iteration process to determine whether the model converges;

[0079] In this example, the Patch images need to be first scaled to a fixed size of 224×224 through the resize method as the model input. During the operation of the base model, the image is segmented into tiles of size 16×16. Set 24 Transformer encoder layers, set the input dimension of the model hidden layer features and the dimension D of the CLS vector to 1024, the final predicted time-to-death category C is 4, and set the number of training iterations to 50 rounds; through ablation experiments, compare the performance of the base model with the ViT-Tiny and ViT-Large models to confirm the effectiveness of the base model in pre-training on large-scale histological images, which is more suitable for feature extraction and analysis of tissue images (the ablation experiment results are listed in Table 1).

[0080] Table 1 Loss values and accuracies of training outcomes of different ViT model architectures

[0081]

[0082] 4. Probability map drawing:

[0083] For each postmortem whole-slide image, input the Patch image with position information into the fine-tuned base model, and use the softmax normalization result of the classification head as the probability value of the Patch belonging to different time-to-death categories;

[0084] Create C all-zero matrices, where C is the number of predicted time-to-death categories, and the dimension of the matrix is calculated by taking the integer after dividing the original size of the whole-slide image by the size of the cropped Patch image;

[0085] Extract the position information attached to the Patch, perform coordinate conversion, and fill the predicted probability value of the Patch image into the corresponding position in the matrix to obtain the probability matrix at the whole-slide image level;

[0086] According to the mapping relationship between the probability level and the image grayscale, a grayscale mapping graph is generated for the probability matrix of each category. The probability value range of each category is 0 - 1. When the probability value is close to 0, the pixel grayscale at the corresponding position in the mapping graph is darker; when it is close to 1, the pixel grayscale is brighter.

[0087] Since the postmortem changes are global changes in tissue morphology, it is assumed that the predicted final postmortem time category of the whole postmortem section image is the category of the postmortem time indicated by the majority of Patch images in the global whole postmortem section image. Based on this, by comparing the proportion of bright areas and dark areas in the probability mapping graphs of different categories, the postmortem time classification result of the whole postmortem section image can be obtained by observing which category has the largest proportion of bright areas in the probability mapping graph. An external validation dataset was constructed using 16 samples from different batches. A total of 15 samples were inferred to be within the correct postmortem time range. The output of the classification probability mapping graph for predicting one of the samples from the 24 - 48h postmortem time range in this specific embodiment is as Figure 3 shown, and the result indicates that the postmortem time category of this sample is 24 - 48h.

[0088] The above - mentioned are only specific implementation manners of the present invention, enabling those skilled in the art to understand or implement the present invention. Although the foregoing embodiments have been described in detail, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A method for estimating time of death based on a probability map of the full slice of a base model, characterized in that: The following steps are involved: S1. Digital acquisition of post-mortem whole-slice images using a high-resolution digital slide scanning system; S2. Preprocess and quality control the post-mortem whole-slice images to construct a high-quality image dataset that can be used in the model information propagation process; S3. Fine-tune the large-scale self-supervised pre-trained base model with attention mechanism on high-quality image datasets; S4. Predict the postmortem full-slice images using the fine-tuned base model, draw the predicted global in-situ category probability map, and make the final time of death category inference after observing the brightness of the different category maps.

2. The method for estimating time of death based on the probability map of the full slice of the base model according to claim 1 is characterized in that: Step S1 specifically includes the following steps: (1) Extract body tissue at different time points after death, prepare H&E sections, and obtain post-mortem whole-section images using a high-resolution digital pathology image scanning system; (2) Postmortem whole-slice images contain information about postmortem changes, from which special postmortem morphological changes can be observed, which are significantly different from the microscopic morphology of tissues before death; (3) Divide the time of death into successive time periods according to the time point and use this as the basis for the category labeling of the postmortem full-slice images.

3. The method for estimating time of death based on the probability map of the full slice of the base model according to claim 1 is characterized in that: Step S2 specifically includes the following steps: (1) Read the postmortem whole-slice image as a memory object with a pyramid structure configured with the pixel size, magnification, downsampling factor, and storage path of the whole-slice image; (2) At the highest resolution layer without downsampling, the postmortem whole-slice image is cropped into a patch image, and the position information of the patch in the postmortem whole-slice image is simultaneously recorded; (3) Evaluate the quality of the cropped patch images and remove low-quality patches that are blurred, folded, or have too large backgrounds; (4) The color features of the patch image are decomposed by non-negative matrix, and the global color space of the image is separated into intensity space and matrix space. The global color space of the image is then directed to the staining space of the reference image to achieve color normalization, so as to eliminate the influence of staining differences between different slices on the analysis results.

4. The method for estimating time of death based on the probability map of the full slice of the base model according to claim 3 is characterized in that: Step S3 specifically includes the following steps: (1) Label the time of death of the quality-controlled patch images, and treat the postmortem whole-slice images as a package containing patch images labeled with the time of death labels; (2) Input the preprocessed and quality-controlled patch image into the base model for fine-tuning training; (3) The base model is based on a large visual Transformer architecture and combines the attention mechanism to extract features from images; (4) The base model is self-supervised pre-trained on a high-quality image dataset; (5) The forward propagation process of the base model first divides the patch image into fixed-size tiles in a non-overlapping manner, flattens it into feature vectors, and then embeds it into a fixed length. The feature vectors are converted into a two-dimensional feature matrix through a stacking operation; (6) Equal-length learnable [CLS] vectors are stacked to the beginning of the matrix to form a new feature matrix. After the self-attention mechanism is operated in the multi-layer Transformer encoder, the feature matrix is ​​updated according to the attention score. Each feature vector of the new feature matrix is ​​represented by weighted stacking of the remaining feature vectors to ensure the contextual association between each feature vector. (7) The [CLS] embedding vector is extracted from the updated feature matrix through matrix slicing and input into a classification head consisting of a fully connected layer to map the vector into a probability distribution representation with a length equal to the number of target categories; (8) In the process of evaluating the performance of the model adjustment process, ablation logic was used to compare the performance of ViT-Tiny, ViT-Large and the base model, confirming the effectiveness of the base model in pre-training large-scale histological images and making it more suitable for feature extraction and analysis of tissue images.

5. The method for estimating time of death based on the probability map of the full slice of the base model according to claim 4 is characterized in that: Step S4 specifically includes the following steps: (1) Create C all-zero matrices, where C is the number of predicted death time categories. The matrix size is calculated by dividing the original size of the postmortem full-slice image by the size of the cropped patch image and then rounding it off; (2) Analyze the location information attached to the patch image and convert the coordinates. Fill the corresponding positions in the all-0 matrix with the corresponding predicted probability values ​​to obtain the probability matrix at the level of the whole slice image. (3) Based on the mapping relationship between probability level and grayscale, a grayscale mapping map is generated for each category’s probability matrix. The probability value range of each category is 0-1. When the probability value is close to 0, the grayscale of the pixel at the corresponding position in the mapping map is darker, and when it is close to 1, the grayscale of the pixel mapping map is brighter. (4) Since postmortem changes are global changes in tissue morphology, the time of death category pointed to by the majority of patch images in the postmortem full-slice image is the final time of death category prediction of the postmortem full-slice image.

Images