Method and apparatus for predicting risk of basal cell carcinoma recurrence based on pathological section images

Abstract

The application relates to a basal cell carcinoma recurrence risk prediction method and device based on pathological section images. The method comprises the following steps: acquiring a target digital image corresponding to a target pathological section and clinical features of a target object; performing cell distribution feature extraction processing on the target digital image according to four feature dimensions of cell composition, cell spatial topology, cell morphology and cell texture to obtain target cell distribution features; performing fusion processing on the target pathological features, the target cell distribution features and the clinical features to obtain target fusion features; and inputting the target fusion features into a risk prediction model to obtain a basal cell carcinoma recurrence risk grade corresponding to the target pathological section. The method can realize objective prediction of the basal cell carcinoma recurrence risk.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and apparatus for predicting the recurrence risk of basal cell carcinoma based on pathological slide images. Background Technology

[0002] Basal cell carcinoma (BCC) is a common non-melanocyte skin tumor. Adverse prognostic events for BCC mainly include local recurrence and metastasis, with metastasis being extremely rare. Local recurrence is a major challenge in clinical treatment management. Even with standard surgical resection techniques (e.g., 4-5 mm margins), there is still a certain risk of recurrence.

[0003] A standardized pathology report is fundamental to the standardized diagnosis and treatment of BCC and for assessing prognosis. Generally, a skin tissue pathology report should include specimen type, presence or absence of ulceration, depth of infiltration, surgical margin condition, histological type, and auxiliary diagnostic results. However, in current clinical practice, the interpretation of these indicators relies heavily on manual analysis by pathologists, which is not only subjective but also extremely time-consuming. Furthermore, even if pathologists provide a detailed list of clinical and pathological factors associated with BCC recurrence, it remains difficult to objectively and quantitatively assess a patient's recurrence risk.

[0004] Therefore, there is an urgent need for an objective method to predict the risk of recurrence of basal cell carcinoma. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and storage medium for predicting the recurrence risk of basal cell carcinoma based on pathological slide images, which can objectively predict the recurrence risk of basal cell carcinoma, in response to the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images, including: Obtain the target digital image and clinical features of the target object corresponding to the target pathological section. The target pathological section is the pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is the digital image of the pathological section excluding the background area. The target digital image is processed to extract pathological features, thereby obtaining the target pathological features; The cell distribution features of the target digital image are extracted according to four feature dimensions: cell composition, cell spatial topology, cell morphology and cell texture, to obtain the target cell distribution features. The target pathological features, target cell distribution features, and clinical features are fused to obtain the target fused features; By inputting the target fusion features into the risk prediction model, the recurrence risk level of basal cell carcinoma corresponding to the target pathological section is obtained.

[0007] Secondly, this application also provides a device for predicting the recurrence risk of basal cell carcinoma based on pathological slide images, comprising: The data acquisition module is used to acquire the target digital image and clinical features of the target object corresponding to the target pathological section. The target pathological section is a pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is a digital image of the pathological section excluding the background area. The pathological feature extraction module is used to extract pathological features from the target digital image to obtain the target pathological features. The distribution feature extraction module is used to extract cell distribution features from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology and cell texture, to obtain the target cell distribution features. The feature fusion module is used to fuse target pathological features, target cell distribution features, and clinical features to obtain target fused features; The risk prediction module is used to input the target fusion features into the risk prediction model to obtain the risk level of basal cell carcinoma recurrence corresponding to the target pathological section.

[0008] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method provided in the first aspect above.

[0009] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect above.

[0010] The aforementioned method, apparatus, device, and storage medium for predicting the recurrence risk of basal cell carcinoma based on pathological slide images acquire a target digital image corresponding to the target pathological slide and the clinical characteristics of the target object. The target pathological slide is a pathological slide of the basal cell carcinoma tissue of the target object, and the target digital image is a digital image of the pathological slide excluding the background area. Pathological features are extracted from the target digital image to obtain target pathological features. Cell distribution features are extracted from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology, and cell texture to obtain target cell distribution features. The target pathological features, target cell distribution features, and clinical features are fused to obtain target fused features. The target fused features are input into a risk prediction model to obtain the basal cell carcinoma recurrence risk level corresponding to the target pathological slide. Among them, the target pathological features characterize the intrinsic malignancy level of tumor cells, the target cell distribution features describe the tumor microenvironment of basal cell carcinoma, quantify the spatial geometric interaction between various immune cells and tumor cells in the tumor microenvironment, and the clinical features provide macroscopic information about basal cell carcinoma. In this embodiment, the target digital image corresponding to the target pathological section is processed by extracting pathological features and cell distribution features, and the target pathological features, target cell spatial distribution features and clinical features are fused. The fused target features are then input into the risk prediction model to obtain the recurrence risk of basal cell carcinoma corresponding to the target pathological section, thereby achieving an objective prediction of the recurrence risk of basal cell carcinoma. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is an application environment diagram of a basal cell carcinoma recurrence risk prediction method based on pathological slide images in one embodiment. Figure 2 This is a flowchart illustrating a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images in one embodiment. Figure 3 This is a flowchart illustrating the steps for determining target fusion features in one embodiment; Figure 4 This is a flowchart illustrating the steps for obtaining the target pathological features in one embodiment; Figure 5 This is a flowchart illustrating the steps for obtaining the target cell distribution characteristics in one embodiment; Figure 6This is a flowchart illustrating the steps for obtaining the clinical characteristics of a target object in one embodiment; Figure 7 This is a flowchart illustrating the steps for obtaining the recurrence risk level of basal cell carcinoma in one embodiment; Figure 8 This is a flowchart illustrating a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images in another embodiment. Figure 9 This is a structural block diagram of a basal cell carcinoma recurrence risk prediction device based on pathological slide images in one embodiment. Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0014] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure.

[0016] It is understood that the terms "first," "second," etc., as used herein may be used to describe various elements, but these elements are not limited by these terms. These terms are used only to distinguish one element from another. For example, without departing from the scope of this disclosure, a first resistor may be referred to as a second resistor, and similarly, a second resistor may be referred to as a first resistor. Both the first resistor and the second resistor are resistors, but they are not the same resistor.

[0017] It is understood that the term "connection" in the following embodiments should be understood as "electrical connection," "communication connection," etc., if the connected circuits, modules, units, etc., have electrical signal or data transmission with each other.

[0018] It is understandable that "at least one" refers to one or more, and "multiple" refers to two or more. "At least a part of an element" refers to part or all of an element.

[0019] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0020] The method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located on a cloud or other network server. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and medical computing devices. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0021] In one exemplary embodiment, such as Figure 2 As shown, a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images is provided, and this method is applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps 202 to 210. Wherein: Step 202: Obtain the target digital image and clinical characteristics of the target object corresponding to the target pathological section.

[0022] Among them, the target pathological section is the pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is the digital image of the pathological section excluding the background area.

[0023] In one possible implementation, the process of obtaining the target digital image corresponding to the target pathological slide includes: obtaining a full-view slice image of the target pathological slide, using a tissue segmentation algorithm to remove the background area in the full-view slice image, retaining the area containing basal cell carcinoma tissue, and obtaining the target digital image.

[0024] For example, the process of acquiring a full-field slide image of a target pathological section includes: acquiring a full-field digital slide image of a basal cell carcinoma patient, with a scan magnification of 40× and an image format of .svs.

[0025] For example, the process of using a tissue segmentation algorithm to remove the background region from a full-view slice image while retaining the region containing basal cell carcinoma tissue in the background to obtain the target digital image includes: downsampling the full-view slice image to a low resolution (e.g., 5x magnification) to obtain a low-resolution image; converting the low-resolution image to a grayscale image; calculating the optimal threshold for the grayscale image using the Otsu algorithm; binarizing the grayscale image using the maximum threshold to distinguish between the foreground tissue region and the background region; performing morphological operations (e.g., opening and closing operations) on the binarization result to remove noise and retain the complete tissue region; and saving the coordinate information of the tissue region to obtain the target digital image.

[0026] The clinical characteristics of the target subjects were extracted based on the corresponding clinical information, which provided macroscopic information on basal cell carcinoma for subsequent recurrence risk prediction.

[0027] Step 204: Extract pathological features from the target digital image to obtain the target pathological features.

[0028] The target digital image is input into a relevant deep learning network to extract pathological features, thus obtaining the target pathological features. These target pathological features are used to characterize the intrinsic malignancy of basal cell carcinoma.

[0029] Step 206: Extract cell distribution features from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology, and cell texture, to obtain the target cell distribution features.

[0030] The target cell distribution features are used to characterize the tumor microenvironment (TME) in the target pathological section. In this embodiment, data such as cell type, cell location, and cell morphology of each cell in the target digital image are first extracted. Then, features in four dimensions—cell composition, cell spatial topology, cell morphology, and cell texture—are extracted based on the cell type, cell location, and cell morphology data. The extracted features are then integrated to obtain the target cell distribution features that describe the tumor microenvironment in the target pathological section. In some embodiments, the target cell distribution features may also be referred to as TME handcrafted features.

[0031] Step 208: The target pathological features, the target cell distribution features, and the clinical features are fused to obtain the target fused features.

[0032] In this process, the target pathological features, target cell distribution features, and clinical features are fused together, so that the target fused features include feature information describing basal cell carcinoma from different modalities.

[0033] In one possible implementation, the target pathological features, target cell distribution features, and clinical features are fused as features at the same level to obtain the target fusion features; for example, the target pathological features, target cell distribution features, and clinical features are first mapped to a unified feature space, and then the mapped features are added element by element to obtain the target fusion features.

[0034] In one possible implementation, different fusion weights are set for the target pathological features, target cell distribution features, and clinical features. Based on their respective fusion weights, the target pathological features, target cell distribution features, and clinical features are weighted and fused to obtain the target fused features.

[0035] Step 210: Input the target fusion feature into the risk prediction model to obtain the basal cell carcinoma recurrence risk level corresponding to the target pathological section.

[0036] Specifically, the target fusion features are input into a pre-trained risk prediction model to obtain the risk level of basal cell carcinoma recurrence corresponding to the target pathological section.

[0037] For example, the applicant collected pathological slide images from 318 patients with basal cell carcinoma, including 98 patients with recurrence and 220 patients without recurrence, as the training dataset. A 5-fold cross-validation strategy was adopted, dividing the samples by slide level to ensure that full slide images of the same patient would not appear in both the training and validation sets simultaneously, thus avoiding data leakage. Each slide was trained using 4-fold data (approximately 254 cases) and validated using 1-fold data (approximately 64 cases) to train the risk prediction model.

[0038] For example, the risk level of basal cell carcinoma recurrence can be the recurrence probability, recurrence score, risk stratification result, or logarithmic recurrence risk.

[0039] The method provided in this application embodiment is only used for feature extraction and quantitative calculation of digital images of ex vivo pathological slides. The above feature extraction and quantitative calculation process is all completed by computer. The output risk score is only used as a technical parameter for computer processing of image information. It is an intermediate result of clinical diagnosis and cannot be directly derived or equivalent to the final diagnosis of the disease. It is not used as a direct basis for clinical medical diagnosis and treatment.

[0040] The above-described method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images involves: acquiring a target digital image corresponding to a target pathological slide and the clinical characteristics of the target object; extracting pathological features from the target digital image to obtain target pathological features; extracting cell distribution features from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology, and cell texture to obtain target cell distribution features; fusing the target pathological features, target cell distribution features, and clinical features to obtain target fused features; and inputting the target fused features into a risk prediction model to obtain the recurrence risk level of basal cell carcinoma corresponding to the target pathological slide. Among them, the target pathological features characterize the intrinsic malignancy level of tumor cells, the target cell distribution features describe the tumor microenvironment of basal cell carcinoma, quantify the spatial geometric interaction between various immune cells and tumor cells in the tumor microenvironment, and the clinical features provide macroscopic information about basal cell carcinoma. In this embodiment, the target digital image corresponding to the target pathological section is processed by extracting pathological features and cell distribution features, and the target pathological features, target cell spatial distribution features and clinical features are fused. The fused target features are then input into the risk prediction model to obtain the recurrence risk of basal cell carcinoma corresponding to the target pathological section, thereby achieving an objective prediction of the recurrence risk of basal cell carcinoma.

[0041] The above embodiments are aimed at basal cell carcinoma, which has local recurrence as the main poor prognosis time. They are different from pan-cancer survival risk prediction, cancer metastasis prediction or breast cancer recurrence and metastasis prediction schemes. They establish a dedicated prediction process based on the pathological tissue morphology of BCC, the spatial tissue relationship of cells and clinical recurrence-related factors.

[0042] In one exemplary embodiment, based on Figure 2 The illustrated embodiment provides a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images. This involves fusing target pathological features, target cell distribution features, and clinical features to obtain the target fused features, including steps 302 to 306. Figure 3 As shown.

[0043] Step 302: Use a gating mechanism to perform quality assessment on clinical features and target cell distribution features to obtain the first gating value corresponding to the clinical features and the second gating value corresponding to the target cell topological features.

[0044] In one possible implementation, if the dimensions of the clinical features and the target cell topological features are different, the clinical features and the target cell topological features are first projected onto a unified feature space of a preset dimension, and the projected clinical features are input into a first gating network to obtain a first gating value output by the first gating network; the projected target cell topological features are input into a second gating network to obtain a second gating value output by the second gating network.

[0045] The first and second gating networks are both composed of multiple fully connected layers and activation function layers. The output g value g∈[0,1] is used to characterize the effectiveness of the corresponding feature.

[0046] For example, both the first gating network and the second gating network contain two fully connected layers. The input of the first gating network is the projected clinical features, and the output is the first gating value. The input of the second gating network is the projected target cell topological features, and the output is the second gating value, as shown in the following formula: ; ; in, Indicates the first gating value. Indicates the second gating value. and The value range is [0, 1], which represents the quality score of the corresponding feature; This represents the topological features of the target cell after projection. Indicates the clinical features after projection. , , and This represents a fully connected layer, and ReLU and sigmoid represent the corresponding activation functions.

[0047] Step 304: Based on the first and second gate values, perform feature fusion processing on the clinical features and target cell distribution features to obtain auxiliary fusion features. Wherein, if the target gate value is less than a preset threshold, the fusion weight of the target feature is determined to be zero. If the target gate value is greater than or equal to the preset threshold, the fusion weight of the target feature is determined based on the target gate value. Wherein, the target gate value is the first or second gate value, and the target feature is the feature corresponding to the target gate value.

[0048] Specifically, when the first gate value is less than a preset threshold, the fusion weight corresponding to the clinical feature is set to zero; when the first gate value is greater than or equal to the preset threshold, the fusion weight corresponding to the clinical feature is determined based on the first gate value; when the second gate value is less than the preset threshold, the fusion weight corresponding to the target cell distribution feature is set to zero; when the second gate value is greater than or equal to the preset threshold, the fusion weight corresponding to the target cell distribution feature is determined based on the second gate value.

[0049] In this embodiment, clinical features and target cell distribution features are used as auxiliary modal features. For example, the preset threshold ranges from 0.2 to 0.6; optionally, the preset threshold τ ranges from 0.3 to 0.5, combining continuous gating and threshold masking to suppress low-quality auxiliary modal features. When the gating value corresponding to an auxiliary modal feature is lower than the threshold, the low-weight feature is not retained; instead, the projected auxiliary modal feature is set to zero to block its contribution to the subsequent fusion features. The process of obtaining the auxiliary fusion features is shown in the following formula: , , , in, and These are learnable fusion weights. If... The fusion weights corresponding to the target cell distribution features are: ,like If the target cell distribution features correspond to 0, then the fusion weight is 0; if The fusion weights corresponding to the clinical features are: ,like If the clinical features are 0, then the fusion weight corresponding to them is 0.

[0050] Step 306: Determine the target fusion feature based on the auxiliary fusion features and the target pathological features.

[0051] Among them, auxiliary fusion features and target pathological features are fused to obtain target fusion features.

[0052] In one possible implementation, the process of determining the target fusion feature based on auxiliary fusion features and target pathological features includes obtaining the target fusion feature using the following formula:

[0053] in, Indicates target fusion features, For balance coefficient, The value range is from 0.25 to 0.35. Indicates the target pathological features, This indicates auxiliary fusion features.

[0054] For example, if the target pathological features and the auxiliary fusion features have different dimensions, the target pathological features are first projected into the same dimensional space as the auxiliary fusion features before the fusion process is performed.

[0055] In the above embodiments, a balance coefficient is set during the fusion process to ensure the dominant role of pathological features. The balance coefficient β can reduce the influence of auxiliary fusion terms on the main pathological features, but it cannot eliminate the non-zero perturbations of low-quality auxiliary modes in the fusion vector. Therefore, the aforementioned threshold masking method can cut off the contribution of low-confidence auxiliary modes before obtaining the target fusion features.

[0056] Optionally, the balance coefficient β is set to 0.3.

[0057] In this embodiment, the spatial distribution characteristics and clinical characteristics of the target cells are used as auxiliary features. Before fusion, the quality of these two features is first assessed. If the quality assessment results of the spatial distribution characteristics and / or clinical characteristics are low, the fusion weight is reset to zero to suppress the low-quality features from entering the subsequent fusion features. This reduces the negative transfer effect of low-quality auxiliary features caused by insufficient cell detection and lack of clinical information on the fusion results, so that the risk prediction results can be interpreted along the pathological main modality and the effective auxiliary modality, avoiding the negative impact on the prediction results.

[0058] In basal cell carcinoma scenarios, small tumor nests, uneven distribution of inflammation and stroma, insufficient cell detection, or lack of clinical information can all lead to clinical features or target cell distribution features becoming low-confidence signals. In the methods provided in the above embodiments, a threshold mask is set on the basis of continuous gating to explicitly suppress low-quality auxiliary modal features. The necessity lies in the fact that a low gating value can only reduce the amplitude of auxiliary modal features, but cannot change their noise properties. In the subsequent fusion process with pathological features, low-quality auxiliary modal features may still enter the target fusion feature as non-zero perturbations and be nonlinearly amplified or fitted as spurious correlations by the risk prediction model. In the embodiments of this application, by setting a preset threshold, the fusion weight is set to zero when the gating value is less than the preset threshold, so that low-quality auxiliary modal features are cut off before entering the risk prediction model, avoiding their continuous influence on the pathological features of the main modality and improving the robustness of the risk prediction model under the condition of fluctuations in the quality of real clinical data.

[0059] In an exemplary embodiment, based on the foregoing embodiments, the provided method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images involves the process of extracting pathological features from a target digital image to obtain the target pathological features. For example... Figure 4 As shown, the process includes steps 402 to 406. Wherein: Step 402: The target digital image is divided into non-overlapping segments according to a preset size to obtain multiple image blocks.

[0060] For example, the preset size is 256×256 pixels or 512×512 pixels. For instance, non-overlapping image patches are extracted in units of 256×256 pixels with a fixed step size, and the overlap rate is set to 0.

[0061] In this embodiment, the image blocks are sampled with a fixed step size and a zero overlap rate to maintain the unique spatial coordinates of the image blocks, avoid the same cell being counted repeatedly, and reduce the computational load of full-view slice processing.

[0062] Step 404: Input each image block into the pathological feature encoder to obtain the initial pathological features corresponding to each image block.

[0063] For example, each image patch is represented using a feature encoder based on the UNI (Towards a General-Purpose Foundation Model for Computational Pathology) model to extract subcellular-level microscopic pathological high-order semantic features. The UNI model is based on the Vision Transformer (ViT) architecture and has been self-supervised pre-trained on over 100 million pathological images, demonstrating powerful pathological feature representation capabilities.

[0064] In one possible implementation, the feature dimension of the initial pathological features corresponding to the image patch is between 512 and 3072. For example, the output feature dimension is 1024, represented as a 1×1024 feature vector.

[0065] Step 406: Perform weighted fusion processing on the initial pathological features corresponding to each image block to obtain the target pathological features.

[0066] Because the digital images of pathological slides are extremely large (up to tens of thousands of pixels), they cannot be directly input into deep learning models or directly fused with features from other modalities. In this embodiment, the initial pathological features at the image patch level are aggregated into features at the sample level.

[0067] In one possible implementation, the process of weighted fusion of the initial pathological features corresponding to each image block to obtain the target pathological features includes steps A2 and A4.

[0068] Step A2: Calculate the attention weights corresponding to each initial pathological feature.

[0069] For example, the initial pathological features {h1,h2,...,h} extracted in step 404 are... n The fully connected layers map to a unified feature space, as shown in the following formula: ; in, Indicates the pathological features after mapping. This indicates a fully connected layer with an output dimension of 1×512, where n represents the number of image patches.

[0070] The importance of each image patch is quantified using an attention mechanism to obtain the attention weight of each image patch, as shown in the following formula: ; in, This represents the attention weight of the i-th image patch. and Here, T represents the matrix transpose, and n represents the number of image patches. , ,and i ≠ j .

[0071] Step A4: Based on the attention weights corresponding to each initial pathological feature, perform weighted fusion processing on each initial pathological feature to obtain the target pathological feature.

[0072] The pathological feature representation at the sample level, i.e., the target pathological feature, is obtained by weighted summation. H As shown in the following formula:

[0073] Among them, target pathological features The dimension is 1×512.

[0074] In the above embodiments, the target digital image is overlapped and segmented according to a preset size to obtain multiple image blocks. Each image block is input into a pathological feature encoder to obtain the initial pathological features corresponding to each image block. The initial pathological features corresponding to each image block are then weighted and fused to obtain the target pathological features. In this way, a smaller-scale pathological feature encoder can be used to extract pathological features, and then the initial pathological features of each image block can be weighted and fused, reducing the hardware resource consumption for pathological feature extraction.

[0075] In an exemplary embodiment, based on the foregoing embodiments, the provided method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images involves extracting cell distribution features from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology, and cell texture, to obtain the target cell distribution features. For example... Figure 5 As shown, the process includes steps 502 to 512. Wherein: Step 502: Input the target digital image into the cell instance segmentation model to determine the cell type, cell location, and cell morphology of each cell in the target digital image.

[0076] In one possible implementation, the cell instance segmentation model employs the HoVer-Net model. The HoVer-Net model is based on an encoder-decoder architecture, where the encoder is built upon ResNet and its variants, and the decoder includes kernel pixel segmentation branches, horizontal and vertical distance prediction branches, and kernel type classification branches. For example, the HoVer-Net model is used with publicly available pre-trained weights for inference, without adjusting the model weights for the specific sample, outputting tumor cells, immune cells, stromal cells, and other cell types, as well as the cell locations of each cell. In some implementations, other deep learning models with target segmentation capabilities are employed to determine the cell type, cell location, and cell morphology of each cell in the target digital image.

[0077] The encoder uses ResNet50 as the backbone network to extract image features. It employs three decoding branches: a Nuclear Pixel Branch (predicting whether each pixel belongs to a cell nucleus), a HoVer Branch (predicting the horizontal and vertical distances from each nuclear pixel to the nucleus center), and a Type Branch (predicting the type of each cell nucleus). The target digital image is segmented using a 256×256 pixel window, and inference is performed using a HoVer-Net model with pre-trained weights. Based on the horizontal and vertical distance maps, a watershed algorithm is used to separate adhered cell nucleus instances. The segmentation results are then processed with confidence filtering, area thresholding, and duplicate nucleus merging. The center coordinates (x, y), cell type (tumor cell, immune cell, or stromal cell, etc.), and morphological parameters such as area, perimeter, roundness, and contour coordinates are extracted for each cell nucleus. The cell instance segmentation results, including the location, type, and morphological information of each cell, are saved.

[0078] Step 504: Using the node cells in the target digital image as network nodes, construct a cell spatial topology network based on the cell positions of the node cells.

[0079] In this context, node cells are cells of the tumor, immune, or stromal type. In some implementations, immune cells are referred to as inflammatory cells.

[0080] Based on the cell location and cell type extracted in step 502, a cell spatial topology network is constructed to quantify the spatial organization relationship of cells in the tumor microenvironment.

[0081] In one possible implementation, the process of constructing a cellular spatial topology network based on the cell location of node cells includes: for each network node, taking the network node as a target network node, determining the neighboring nodes corresponding to the target network node, and establishing connections between the target network node and its corresponding neighboring nodes to obtain the cellular spatial topology network. The process of determining the neighboring nodes corresponding to the target network node includes steps B2 to B8.

[0082] Step B2: Based on the cell positions of each node cell, determine the distance between the target node cell and the other node cells corresponding to the target network node.

[0083] Step B4: Based on the distance between the target node cell and the other node cells, determine the candidate node cells corresponding to the target node cell.

[0084] Step B6: If the number of candidate node cells is greater than a preset value, the network nodes corresponding to the preset number of candidate node cells with the smallest distance to the target node cell are determined as the neighbor nodes of the target network node.

[0085] Step B6: If the number of candidate node cells is less than or equal to a preset value, the network node corresponding to the candidate node cell is determined as a neighbor node of the target network node.

[0086] For example, for each cell i, the Euclidean distance between it and other cells j is calculated as shown in the following formula:

[0087] Set the maximum spatial radius of action r = 25 micrometers and the number of nearest neighbors k = 5. For cell i, among the cells with a distance of less than or equal to 25 micrometers, select the k nearest neighbors as its neighbors; if the Euclidean distance d(i,j) between cell i and cell j is ≤ 25 micrometers and cell j is among the k nearest neighbors of cell i, then establish a connection edge between cell i and cell j; in this way, traverse each node cell to construct the cell spatial topology network.

[0088] In the process of constructing the cellular spatial topology network described in the above embodiments, the coordinates of the cell nucleus center are used as the node positions, and the k-nearest neighbor method and the distance threshold method are used to jointly constrain the connecting edges to construct the cellular spatial topology network. For example, the preset number k=5, and the preset distance threshold is 25 micrometers. When the distance between two cells is less than or equal to the set threshold and they are candidate nearest neighbors, they are considered to have a potential spatial interaction relationship and a connecting edge is established, thereby constructing a cellular spatial topology network that reflects the spatial organization and interaction of cells in the basal cell carcinoma tumor microenvironment.

[0089] Step 506: Count the number of tumor cells, immune cells and stromal cells respectively, and determine the cell composition characteristics corresponding to the target digital image.

[0090] The cellular composition characteristics have a dimension of 1×3, including the number of tumor cells, immune cells, and stromal cells, respectively.

[0091] Step 508: Extract the cell space topology features corresponding to the target digital image based on the cell space topology network.

[0092] In one possible implementation, one or more graph features among the following in the cellular space topology network are calculated as cellular space topology features: minimum edge length, average edge length, number of subgraphs, node degree, core degree, clustering coefficient, eccentricity, harmonic centrality, proximity centrality, betweenness centrality, and matrix barrier index.

[0093] For example, the cellular space topology features have a dimension of 1×140, including the edge length, connected subgraphs, node degree, centrality, and matrix barrier-related indicators of the cellular space topology network.

[0094] Step 510: Based on the cell morphology of each node cell, obtain the cell morphology features and cell texture features corresponding to the target digital image.

[0095] The cell morphology of each node cell includes morphological parameters such as area, perimeter, roundness, and contour coordinates of each node cell. By combining the above morphological parameters with the target digital image, the cell morphological features and cell texture features corresponding to the target digital image are calculated.

[0096] In one possible implementation, cell morphology features include the mean and variance of 14 indicators, such as area, area of ​​the circumscribed rectangle, eccentricity, roundness, elongation, occupancy, major axis length, minor axis length, perimeter, solidity, mean curvature, standard deviation of curvature, maximum curvature, and minimum curvature. The dimension of cell morphology features is 1×2. Cell texture features include the mean and variance of 9 indicators, such as second moment of angle, contrast, correlation, entropy, homogeneity, mean gray level, standard deviation of gray level, maximum gray level, and minimum gray level. The dimension of cell texture features is 1×18.

[0097] Step 512: The cell composition features, cell spatial topology features, cell morphology features and cell texture features are spliced ​​together to determine the distribution features of the target cells.

[0098] In this process, cell composition features, cell spatial topology features, cell morphology features, and cell texture features are spliced ​​together in a fixed order to obtain structured TME handcrafted features, which are then projected through a fully connected layer to obtain the target cell distribution features.

[0099] In one possible implementation, cell composition features, cell spatial topology features, cell morphology features, and cell texture features are spliced ​​together in a fixed order to obtain 1×189-dimensional TME handcrafted features, which are then projected through a fully connected layer to obtain 1×512-dimensional target cell distribution features.

[0100] In the above embodiments, cell instance segmentation processing is performed on the target digital image to obtain cell type, cell location, and cell morphology. The cell composition features, cell spatial topology features, cell morphology features, and cell texture features are further extracted. Among them, cell composition features reflect the quantity and structure of tumor cells, stromal cells, and immune cells; cell morphology features reflect the morphological heterogeneity of each cell nucleus and cell region; cell texture features reflect the gray-scale distribution and local tissue structure of cell regions; and cell spatial distribution features reflect topological relationships such as cell neighborhood, spatial connectivity, and matrix barrier. The combination of these four features can simultaneously describe the cell number, cell morphology, local texture, and spatial topology network. Compared with extracting cell spatial distribution features by only constructing a cell topology map or only using attention heatmaps to locate high-risk or high-interest areas, the target cell distribution features extracted in this embodiment have clear feature forms and are easy to reproduce and interpret.

[0101] In one exemplary embodiment, based on the foregoing embodiments, the provided method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images involves the process of acquiring the clinical characteristics of the target subject. For example... Figure 6 As shown, the process includes steps 602 to 604.

[0102] Step 602: Obtain the clinical information of the target subject.

[0103] In one possible implementation, clinical information includes one or more of the following: patient age, sex, tumor location, clinical grade, history of recurrence, surgical margin status, or immune status.

[0104] Step 604: Perform vector encoding on the clinical information to obtain clinical features.

[0105] In one possible implementation, each item of clinical information is vector-encoded to obtain clinical features.

[0106] For example, the clinical information of the target subject includes: (1) Age: a continuous variable, in years; (2) Gender: Categorical variable, coded as 0 (indicating female) or 1 (indicating male); (3) Tumor location: a categorical variable, including face, trunk, limbs, etc., encoded using one-hot encoding; (4) Clinical grading: a classification variable, which is divided into high-grade and low-grade based on tumor size, depth of invasion, etc. The above clinical information was standardized to form clinical characteristics. Dimension is 1×d c . d c The value of d is determined by the total dimension after one-hot encoding of continuous, binary, and categorical variables; for example, when age is a 1-dimensional continuous variable, gender is a 1-dimensional binary variable, tumor location is encoded as a 6-dimensional one-hot variable, and clinical grade is encoded as a 2-dimensional one-hot variable, d c =10.

[0107] In an exemplary embodiment, based on the foregoing embodiments, the provided method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images involves a process of obtaining the recurrence risk level of basal cell carcinoma corresponding to a target pathological slide based on target fusion features and a risk prediction model. For example... Figure 7 As shown, the process includes steps 702 to 708.

[0108] Step 702: Input the target fusion features into the input layer of the risk prediction model to obtain multidimensional input feature components.

[0109] Step 704: Input the input feature components of each dimension into the function layer of the risk prediction model to obtain the transformed feature components of each dimension. The function was implemented using a third-order B-spline interpolation function.

[0110] Step 706: Input the transformation feature components of each dimension into the function combination layer in the risk prediction model for nonlinear mapping processing to obtain the mapped features.

[0111] Step 708: Input the mapped features into the output layer of the risk prediction model to obtain the risk level of basal cell carcinoma recurrence.

[0112] The risk prediction model provided in this embodiment is implemented based on a spline Kolmogorov-Arnold network (KAN). For example, its network structure is [512, 256, 128, 64, 1], including 2-8 hidden layers, preferably 2 to 4 hidden layers; the KAN network structure includes: (1) Input layer, used to receive target fusion features; (2) Function layer, also known as learnable univariate function layer, is used to transform each dimension of the input features through a learnable univariate function; the univariate function is implemented using third-order B-spline interpolation, with learnable control point parameters, and each spline function is set with 8 to 16 grid nodes. In this embodiment, the function layer is implemented using a third-order B-spline interpolation function, replacing the fixed activation function of the traditional neural network (such as ReLU, Sigmoid, etc.); (3) Function combination layer, used to combine the outputs of the function layer to realize complex nonlinear mapping; (4) Output layer: The fused features are mapped to a recurrence risk score s∈[0,1] or a risk log value η through the output layer. The larger the value, the higher the recurrence risk.

[0113] In one possible implementation, the provided risk prediction model is trained using the negative log-partial likelihood loss of the Cox proportional hazards model. Specifically, the negative log-partial likelihood loss function of the Cox proportional hazards model is used to optimize the risk ranking relationship between samples by combining recurrence time and recurrence status, as shown in the following formula: , Where L represents the loss function value, This is an event indicator variable (1 indicates recurrence, 0 indicates truncation). R(t) represents the logarithmic risk predicted by the model. i (time) The risk set at time points. This loss function optimizes the risk ranking relationship between samples and is suitable for survival analysis tasks.

[0114] The Adam optimizer was used for training, with an initial learning rate of 1e-4 and a weight decay coefficient of 1e-5. Cosine annealing was used for learning rate scheduling, with a period of 50 epochs.

[0115] The training strategy includes: training for 50 epochs per fold with a batch size of 4-8; using Dropout=0.1 and gradient clipping (maximum norm 1.0) to prevent overfitting; and triggering an early stopping mechanism when the validation set C-index shows no improvement for 10 consecutive epochs.

[0116] The validation and evaluation metrics used in this embodiment include the consistency index (C-index) and the time-dependent area under the receiver operating characteristic curve (tAUC). The consistency index measures the consistency between the predicted risk score and the actual recurrence time ranking; the tAUC measures the prediction accuracy at different time points. After training on the training set, the model performance is evaluated on the test set.

[0117] In one exemplary embodiment, such as Figure 8 As shown, a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images is provided, and this method is applied to... Figure 1 Taking terminal 102 as an example, the explanation includes the following steps 802 to 214. Wherein: Step 802: Obtain the target digital image corresponding to the target pathological section; wherein, the target pathological section is a pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is a digital image of the pathological section excluding the background area.

[0118] Step 804: The target digital image is segmented into multiple image blocks of a preset size without overlap. The target digital image is divided into multiple image blocks by non-overlapping segments of a preset size.

[0119] Step 806: Input each image block into the pathological feature encoder to obtain the initial pathological features corresponding to each image block.

[0120] Step 808: Perform weighted fusion processing on the initial pathological features corresponding to each image block to obtain the target pathological features.

[0121] The process of weighted fusion of the initial pathological features corresponding to each image patch to obtain the target pathological features includes: calculating the attention weights corresponding to each initial pathological feature; and performing weighted fusion of each initial pathological feature based on the attention weights corresponding to each initial pathological feature to obtain the target pathological features.

[0122] Step 810: Input the target digital image into the cell instance segmentation model to determine the cell type, cell location, and cell morphology of each cell in the target digital image.

[0123] Step 812: Using the node cells in the target digital image as network nodes, construct a cell spatial topology network based on the cell location of the node cells. The node cells are cells of the type of tumor cells, immune cells, or stromal cells.

[0124] Step 814: Count the number of tumor cells, immune cells and stromal cells respectively, and determine the cell composition characteristics corresponding to the target digital image.

[0125] Step 816: Extract the cell space topology features corresponding to the target digital image based on the cell space topology network.

[0126] Step 818: Based on the cell morphology of each node cell, obtain the cell morphology features and cell texture features corresponding to the target digital image.

[0127] Step 820: The cell composition features, cell spatial topology features, cell morphology features and cell texture features are spliced ​​together to determine the distribution features of the target cells.

[0128] Step 822: Obtain the clinical information of the target object, perform vector encoding on the clinical information, and obtain clinical features.

[0129] Step 824: Use the gating mechanism to perform quality assessment on clinical features and target cell distribution features to obtain the first gating value corresponding to the clinical features and the second gating value corresponding to the target cell distribution features.

[0130] Step 826: Based on the first gate value and the second gate value, perform feature fusion processing on the clinical features and the target cell distribution features to obtain auxiliary fusion features. Wherein, if the target gate value is less than a preset threshold, the fusion weight of the target feature is determined to be zero. If the target gate value is greater than or equal to the preset threshold, the fusion weight of the target feature is determined based on the target gate value. The target gate value is the first gate value or the second gate value, and the target feature is the feature corresponding to the target gate value.

[0131] Step 828: Determine the target fusion feature based on the auxiliary fusion features and the target pathological features.

[0132] The target fusion features are obtained using the following formula: , in, Indicates target fusion features, For balance coefficient, The value range is from 0.25 to 0.35. Indicates the target pathological features, This indicates auxiliary fusion features.

[0133] Step 830: Input the target fusion features into the risk prediction model to obtain the basal cell carcinoma recurrence risk level corresponding to the target pathological section.

[0134] The process of inputting the target fusion features into the risk prediction model to obtain the recurrence risk level of basal cell carcinoma corresponding to the target pathological section includes: inputting the target fusion features into the input layer of the risk prediction model to obtain multi-dimensional input feature components; inputting each dimension of input feature components into the function layer of the risk prediction model to obtain each dimension of transformed feature components, the function layer being implemented using a third-order B-spline interpolation function; inputting each dimension of transformed feature components into the function combination layer of the risk prediction model for nonlinear mapping processing to obtain mapped features; and inputting the mapped features into the output layer of the risk prediction model to obtain the recurrence risk level of basal cell carcinoma.

[0135] For example, the applicant trains the extracted features and the model on the training set based on the method and risk prediction model provided in the above embodiments, and then evaluates the performance of the model on the test set.

[0136] The applicant conducted an ablation experiment on the method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images provided in the embodiments of this application. The experimental results are as follows: Using only pathological features: C-index = 0.786 ± 0.092, tAUC = 0.756 ± 0.111; Using only cell distribution characteristics: C-index = 0.740 ± 0.040, tAUC = 0.707 ± 0.045; Using only clinical characteristics: C-index = 0.571 ± 0.131, tAUC = 0.534 ± 0.142; When using the method provided in this embodiment: C-index=0.808±0.041, tAUC=0.759±0.041.

[0137] The experimental results show that pathological features contribute the most to the prediction of recurrence risk, but there is room for improvement when used alone; cell distribution features and clinical features have limited effects when used alone, but their performance can be significantly improved when combined with pathological features.

[0138] In other experiments, the applicant also verified that the proposed hierarchical quality-aware fusion strategy based on quality assessment and gating values, which performs feature suppression, can more effectively integrate multimodal information and improve the reliability of prediction results compared to simple feature splicing.

[0139] The methods provided in the above embodiments can realize a fully automated pathological image analysis and risk prediction process, provide objective and repeatable recurrence risk assessment results, reduce the workload of pathologists, and assist clinicians in developing personalized follow-up and treatment plans.

[0140] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0141] It is understood that the term "based on" as used in this application is used to describe one or more factors that influence the determination, but does not exclude other factors that may influence the determination. For example, the phrase "determine A based on B" means that the determination of A can be based entirely or at least partially on factor B. That is, B is a factor that influences the determination of A, but does not exclude the fact that the determination of A is also based on C.

[0142] Based on the same inventive concept, this application also provides a device for predicting the recurrence risk of basal cell carcinoma based on pathological slide images, which is used to implement the above-mentioned method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the device for predicting the recurrence risk of basal cell carcinoma based on pathological slide images provided below can be found in the limitations of the method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images described above, and will not be repeated here.

[0143] In one exemplary embodiment, such as Figure 9 As shown, a device for predicting the recurrence risk of basal cell carcinoma based on pathological slide images is provided, including: a data acquisition module 902, a pathological feature extraction module 904, a distribution feature extraction module 906, a feature fusion module 908, and a risk prediction module 910. Wherein: The data acquisition module 902 is used to acquire the target digital image corresponding to the target pathological section and the clinical characteristics of the target object. The target pathological section is a pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is a digital image of the pathological section excluding the background area.

[0144] The pathological feature extraction module 904 is used to extract pathological features from the target digital image to obtain the target pathological features.

[0145] The distribution feature extraction module 906 is used to extract cell distribution features from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology and cell texture, so as to obtain the target cell distribution features.

[0146] The quality assessment module 908 is used to perform quality assessment on clinical features and target cell distribution features using a gating mechanism to obtain a first gating value corresponding to the clinical features and a second gating value corresponding to the target cell distribution features.

[0147] The feature fusion module 910 is used to fuse target pathological features, target cell distribution features and clinical features to obtain target fused features.

[0148] The risk prediction module 914 is used to input the target fusion features into the risk prediction model to obtain the risk level of basal cell carcinoma recurrence corresponding to the target pathological section.

[0149] In an exemplary embodiment, the feature fusion module 910 includes a quality assessment unit, a first fusion unit, and a second fusion unit. The quality assessment unit is used to perform quality assessment on clinical features and target cell distribution features using a gating mechanism to obtain a first gating value corresponding to the clinical features and a second gating value corresponding to the target cell distribution features. The first fusion unit is used to perform feature fusion processing on the clinical features and target cell distribution features based on the first gating value and the second gating value to obtain auxiliary fusion features. Specifically, if the target gating value is less than a preset threshold, the fusion weight of the target feature is determined to be zero. If the target gating value is greater than or equal to the preset threshold, the fusion weight of the target feature is determined based on the target gating value. The target gating value is either the first gating value or the second gating value, and the target feature is the feature corresponding to the target gating value. The second fusion unit is used to determine the target fusion feature based on the auxiliary fusion feature and the target pathological feature.

[0150] In an exemplary embodiment, the second fusion unit is used to obtain the target fusion feature using the following formula: , in, Indicates target fusion features, For balance coefficient, The value range is from 0.25 to 0.35. Indicates the target pathological features, This indicates auxiliary fusion features.

[0151] In an exemplary embodiment, the pathological feature extraction module 904 is used to perform non-overlapping segmentation of the target digital image according to a preset size to obtain multiple image blocks; input each image block into the pathological feature encoder to obtain the initial pathological features corresponding to each image block; and perform weighted fusion processing on the initial pathological features corresponding to each image block to obtain the target pathological features.

[0152] In an exemplary embodiment, the pathological feature extraction module 904 is used to calculate the attention weight corresponding to each initial pathological feature; based on the attention weight corresponding to each initial pathological feature, the initial pathological features are weighted and fused to obtain the target pathological feature.

[0153] In an exemplary embodiment, the distribution feature extraction module 906 is used to input the target digital image into a cell instance segmentation model to determine the cell type, cell location, and cell morphology of each cell in the target digital image; to use the node cells in the target digital image as network nodes, and to construct a cell spatial topology network based on the cell location of the node cells, wherein the node cells are cells of cell type tumor cells, immune cells, or stromal cells; to count the number of tumor cells, immune cells, and stromal cells respectively, and to determine the cell composition features corresponding to the target digital image; to extract the cell spatial topology features corresponding to the target digital image based on the cell spatial topology network; to obtain the cell morphology features and cell texture features corresponding to the target digital image based on the cell morphology of each node cell; and to perform a concatenation process on the cell composition features, cell spatial topology features, cell morphology features, and cell texture features to determine the distribution features of the target cells.

[0154] In an exemplary embodiment, the distribution feature extraction module 906 is used to, for each network node, take the network node as the target network node, determine the neighboring nodes corresponding to the target network node, and establish connection edges between the target network node and the corresponding neighboring nodes to obtain a cell space topology network; wherein, the process of determining the neighboring nodes corresponding to the target network node includes: determining the distance between the target node cell and the other node cells based on the cell position of each node cell; determining the candidate node cells corresponding to the target node cell based on the distance between the target node cell and the other node cells; if the number of candidate node cells is greater than a preset number, determining the network nodes corresponding to the preset number of candidate node cells with the smallest distance to the target node cell as the neighboring nodes of the target network node; if the number of candidate node cells is less than or equal to a preset value, determining the network nodes corresponding to the candidate node cells as the neighboring nodes of the target network node.

[0155] In an exemplary embodiment, the risk prediction module 914 is used to input the target fusion features into the input layer of the risk prediction model to obtain multi-dimensional input feature components; input each dimension of the input feature components into the function layer of the risk prediction model to obtain each dimension of transformed feature components, wherein the function layer is implemented using a third-order B-spline interpolation function; input each dimension of the transformed feature components into the function combination layer of the risk prediction model for nonlinear mapping processing to obtain mapped features; and input the mapped features into the output layer of the risk prediction model to obtain the basal cell carcinoma recurrence risk level.

[0156] In an exemplary embodiment, the data acquisition module 902 is used to acquire clinical information of the target object; and to perform vector encoding processing on the clinical information to obtain clinical features.

[0157] Each module in the aforementioned basal cell carcinoma recurrence risk prediction device based on pathological slide images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0158] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0159] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0160] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0161] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0162] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0163] It should be noted that the user information (including but not limited to user device information, user personal information, and patient clinical information) and data (including but not limited to pathological slide images, and other data used for analysis, storage, and display) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with relevant regulations.

[0164] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0165] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0166] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting the recurrence risk of basal cell carcinoma based on pathological slide images, characterized in that, The method includes: Obtain the target digital image and clinical features of the target object corresponding to the target pathological section, wherein the target pathological section is a pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is a digital image of the pathological section excluding the background area; The target digital image is subjected to pathological feature extraction processing to obtain the target pathological features; The target digital image is processed to extract cell distribution features according to four feature dimensions: cell composition, cell spatial topology, cell morphology and cell texture, to obtain the target cell distribution features. The target pathological features, the target cell distribution features, and the clinical features are fused to obtain the target fused features; The target fusion features are input into the risk prediction model to obtain the basal cell carcinoma recurrence risk level corresponding to the target pathological section.

2. The method according to claim 1, characterized in that, The process of fusing the target pathological features, the target cell distribution features, and the clinical features to obtain the target fused features includes: The quality of the clinical feature and the target cell distribution feature is assessed using a gating mechanism to obtain a first gating value corresponding to the clinical feature and a second gating value corresponding to the target cell distribution feature. Based on the first gate value and the second gate value, feature fusion processing is performed on the clinical features and the target cell distribution features to obtain auxiliary fusion features. Specifically, when the target gate value is less than a preset threshold, the fusion weight of the target feature is determined to be zero. When the target gate value is greater than or equal to the preset threshold, the fusion weight of the target feature is determined based on the target gate value. The target gate value is either the first gate value or the second gate value, and the target feature is the feature corresponding to the target gate value. Based on the auxiliary fusion features and the target pathological features, the target fusion features are determined.

3. The method according to claim 2, characterized in that, The determination of the target fusion feature based on the auxiliary fusion feature and the target pathological feature includes: The target fusion feature is obtained using the following formula: ， in, This represents the target fusion feature. For balance coefficient, The value range is from 0.25 to 0.

35. Indicates the target pathological features, This refers to the auxiliary fusion feature.

4. The method according to any one of claims 1-3, characterized in that, The process of extracting pathological features from the target digital image to obtain target pathological features includes: The target digital image is divided into multiple image blocks by a preset size without overlap. Each of the image blocks is input into the pathological feature encoder to obtain the initial pathological features corresponding to each of the image blocks. The initial pathological features corresponding to each image block are weighted and fused to obtain the target pathological features.

5. The method according to claim 4, characterized in that, The step of performing weighted fusion processing on the initial pathological features corresponding to each of the image blocks to obtain the target pathological features includes: Calculate the attention weights corresponding to each of the initial pathological features; Based on the attention weights corresponding to each initial pathological feature, the initial pathological features are weighted and fused to obtain the target pathological feature.

6. The method according to any one of claims 1-3, characterized in that, The target digital image is processed by extracting cell distribution features according to four feature dimensions: cell composition, cell spatial topology, cell morphology, and cell texture, to obtain target cell distribution features, including: The target digital image is input into a cell instance segmentation model to determine the cell type, cell location, and cell morphology of each cell in the target digital image. Using the node cells in the target digital image as network nodes, a cell spatial topology network is constructed based on the cell positions of the node cells. The node cells are cells of the type of tumor cells, immune cells, or stromal cells. The number of tumor cells, immune cells, and stromal cells is counted to determine the cellular composition characteristics corresponding to the target digital image; Based on the cell space topology network, extract the cell space topology features corresponding to the target digital image; Based on the cell morphology of each node cell, the cell morphology features and cell texture features corresponding to the target digital image are obtained; The cell composition features, cell spatial topology features, cell morphology features, and cell texture features are spliced ​​together to determine the distribution features of the target cells.

7. The method according to claim 6, characterized in that, The construction of a cell spatial topology network based on the cell positions of the node cells includes: For each network node, the network node is taken as the target network node, the neighboring nodes corresponding to the target network node are determined, and connection edges are established between the target network node and the corresponding neighboring nodes to obtain the cell space topology network; The process of determining the neighboring nodes corresponding to the target network node includes: Based on the cell position of each node cell, the distance between the target node cell corresponding to the target network node and the other node cells is determined; Based on the distance between the target node cell and the other node cells, candidate node cells corresponding to the target node cell are determined; If the number of candidate node cells is greater than a preset number, the network nodes corresponding to the preset number of candidate node cells that have the smallest distance from the target node cell are determined as the neighbor nodes of the target network node. If the number of candidate node cells is less than or equal to a preset value, the network node corresponding to the candidate node cell is determined as a neighbor node of the target network node.

8. The method according to any one of claims 1-3, characterized in that, The step of inputting the target fusion features into the risk prediction model to obtain the basal cell carcinoma recurrence risk level corresponding to the target pathological section includes: The target fusion features are input into the input layer of the risk prediction model to obtain multidimensional input feature components; Each dimension of the input feature components is input into the function layer of the risk prediction model to obtain the transformation feature components of each dimension. The function layer is implemented using a third-order B-spline interpolation function. The transformation feature components of each dimension are input into the function combination layer of the risk prediction model for nonlinear mapping processing to obtain the mapped features. The mapped features are input into the output layer of the risk prediction model to obtain the risk level of basal cell carcinoma recurrence.

9. The method according to any one of claims 1-3, characterized in that, The process of obtaining the clinical characteristics of the target subject includes: Obtain the clinical information of the target subjects; The clinical information is processed by vector encoding to obtain the clinical features.

10. A device for predicting the recurrence risk of basal cell carcinoma based on pathological slide images, characterized in that, The device includes: The data acquisition module is used to acquire the target digital image and clinical features of the target object corresponding to the target pathological section, wherein the target pathological section is a pathological section of the basal cell carcinoma tissue of the target object, and the target digital image is a digital image of the pathological section excluding the background area. The pathological feature extraction module is used to perform pathological feature extraction processing on the target digital image to obtain the target pathological features; The distribution feature extraction module is used to extract cell distribution features from the target digital image according to four feature dimensions: cell composition, cell spatial topology, cell morphology and cell texture, to obtain the target cell distribution features. The feature fusion module is used to fuse the target pathological features, the target cell distribution features, and the clinical features to obtain the target fused features; The risk prediction module is used to input the target fusion features into the risk prediction model to obtain the basal cell carcinoma recurrence risk level corresponding to the target pathological section.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images