A method and system for detecting endometrial cancer cells

Abstract

The method and system for detecting endometrial cancer cells are proposed in the method, image blocks are matched with pathological report texts, unified semantic features are obtained by using multi-positive sample contrast learning, semantic entropy driven dynamic masking-random shuffling self-supervised training of a visual neural network is performed, then the nucleus is positioned, a weighted cell graph is constructed by taking cells as nodes and adjacent relationships as edges, topological features are obtained, then multi-modal features are generated by cross-scale attention and gate fusion of the visual features and the semantic features, a detection head jointly outputs categories, a bounding box and a mask, and supervision or weak supervision training is performed by comprehensively considering cross-modal consistency loss, a "cancer-normal" preference pair is automatically generated by using high and low probability areas output by an initial model, a reward model is trained, and PPO policy iteration fine tuning is adopted, so that the cell level cancer detection can be accurately and robustly completed without a large number of pixel level labels, and the pathological screening efficiency and accuracy are significantly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cancer detection, and particularly relates to an endometrial carcinoma cell detection method and system.

BACKGROUND

[0002] Endometrial carcinoma is one of the most common malignant tumors in gynecology, and its incidence and mortality have been rising for nearly ten years, and it shows a clear trend of youth. Large sample clinical studies show that the 5-year survival rate of patients who receive standard treatment in the early stage can be as high as 85%; however, once missed or misdiagnosed into the middle and late stages, the prognosis will deteriorate sharply. Therefore, rapid and accurate pathological screening of endometrial cytology sections is of great significance to improve the survival rate of patients.

[0003] Currently, pathologists still need to browse each layer, zoom in on the region, and finely interpret on the panoramic digital section, and complete reading is very time-consuming. In addition, the consistency of the same patient's section between different doctors is usually low, and is easily affected by reading experience, fatigue, distraction and other factors. Furthermore, the benign and malignant transformation zone often presents a low-contrast, fuzzy boundary transition zone, and traditional segmentation or classification models based on fixed scales are difficult to balance.

[0004] Therefore, there is an urgent need for an endometrial carcinoma cell detection method and system with high precision, high efficiency and high robustness to assist doctors in rapid and accurate diagnosis of the disease.

SUMMARY

[0005] In view of this, the embodiments of the present application provide an endometrial carcinoma cell detection method and system.

[0006] In a first aspect, the embodiments of the present application provide an endometrial carcinoma cell detection method, which comprises:

[0007] S1, pairing the image blocks with the pathological report texts, obtaining a visual encoder, a text encoder and a unified semantic feature by training a multi-positive sample contrast loss function;

[0008] S2, dividing the image blocks into image small blocks, dynamically adjusting the masking proportion based on semantic entropy, performing masking processing according to the masking proportion, and optimizing the visual encoder by training a comprehensive loss function;

[0009] S3, locating the cell nucleus center through the optimized visual encoder, and then constructing a cell graph by taking the cells as nodes and the adjacent relationship as edges;

[0010] S4, constructing a detection model based on the unified semantic feature, the image feature and the cell graph feature;

[0011] S5, according to the current prediction result of the detection model, obtaining the region with high probability of cancer and the region with high probability of cancer and forming a preferred pair, then constructing a region contrast loss function according to the preference to train the reward model, and then adjusting the detection model through reinforcement learning.

[0012] As described above, the aspect and any possible implementation, further provided an implementation, the S1 before also includes:

[0013] Obtain panoramic digital pathology sections and pathology report texts, cut each panoramic digital pathology section into a plurality of image blocks x with a size of H0xW0, and extract the corresponding pathology report text t.

[0014] As described above, the aspect and any possible implementation, further provided an implementation, the S1 specifically includes:

[0015] S11, generating two perspectives {x (1) ,x (2)} based on the image block x, and generating two sentence rewriting {t (1) ,t (2)} based on the corresponding pathology report text t to form a four-dimensional vector {x (i) ,t (j)}, wherein i,j∈{1,2};

[0016] S12, extract the visual vector v=f θ (x (k) ) of the image block through the visual neural network, extract the text vector u=g φ (t (k) ) of the text through the text neural network, and obtain the fusion vector z=γv+(1-γ)u by adding a gating coefficient through gating fusion, γ is the gating coefficient, γ=σ(W g [v‖u]),σ is the Sigmoid activation function, W g is the weight matrix of the gating, f θ is the visual neural network processing, x (k) is the image feature of the kth perspective, g φ is the text neural network processing, t (k) is the kth text perspective;

[0017] S13, construct a multi-positive sample contrast loss function for training, and minimize the loss function:

[0018]

[0019] Wherein, P is a set of positive sample pairs, T is a temperature coefficient, is the embedding vector of the i th image, is the embedding vector of the jth text, and <·, ·> is the dot product operation;

[0020] S14, output the trained visual encoder, text encoder and unified semantic feature.

[0021] Aspects and any possible implementation manners described above, further provide an implementation manner, and the S2 specifically comprises:

[0022] S21, divide the image block into a patch sequence by the visual encoder Cut the image block of P × P pixels into a patch sequence Wherein, N = (H0 / P) × (W0 / P);

[0023] S22, call the unified semantic feature v * for each image block, and obtain the category probability distribution p = softmax(W c v * ) after projection, and calculate the normalized semantic entropy:

[0024]

[0025] Wherein, W c is the weight matrix of category prediction, p j is the prediction probability of the jth category, and C is the number of classification categories;

[0026] S23, dynamically determine the masking proportion according to the semantic entropy:

[0027]

[0028] Wherein, σ is a Sigmoid function, and α and τ are hyperparameters;

[0029] S24, according to the masking proportion p mask Randomly sample from the patch sequence to form a set Mask the patches in the set , and then randomly permute the Token vectors of the patches that are not sampled to obtain a disordered sequence Wherein, P perm is a random permutation matrix, and h vis is the Token sequence of the patch that is not masked;

[0030] S25, input to the visual encoder to obtain the representation H enc , predict through the global branch, and minimize the global branch loss function Through the local branch, predict for each masked position and minimizing the local branch loss function where CLS is a neural network feature pooling vector, t mask is a token representation of the occlusion position, is a local reconstruction feature of the i-th position, is a real feature of the i-th position;

[0031] S26, by synthesizing the loss function jointly optimizing the visual encoder; wherein λ g is a weight coefficient of the global loss, λ l is a weight coefficient of the local loss.

[0032] Aspects and any possible implementation manners described above, further provide an implementation manner, the S3 specifically comprises:

[0033] S31, obtaining the center coordinates c i = (x i ,y i ), i = 1...N c of each nucleus node by the optimized visual encoder define the graph node v i as (c i ,f i (0) ), form a node set V = {v where r i is the morphological radius of the i-th nucleus, μ i is the average chroma of the i-th nucleus, s i is the class logits of the i-th nucleus, e i is the visual embedding vector of the i-th nucleus, c i = (x i ,y i ) is the center coordinates of the i-th nucleus, d0 is the initial feature dimension, N c is the total number of cell nodes;

[0034] S32, performing Delaunay triangulation on the center coordinates of all nucleus nodes to obtain an initial candidate adjacency set E raw , pruning the edges with a length exceeding the maximum connection distance threshold r max , to obtain a candidate adjacency set E = {(i,j)|||c i -c j ||≤r max}; Wherein ||c i -c j || is the Euclidean distance between the i-th and j-th nodes;

[0035] S33, calculate the geometric weight of each edge (i, j) e E and the morphological weight The geometric weight and the morphological weight are multiplied to obtain the comprehensive weight a ij , and a sparse weighted adjacency matrix A = [a ij ] is constructed.

[0036] S34, output the cell graph containing the node set V, the candidate adjacency set E and the weighted adjacency matrix A

[0037] G = (V, E, A).

[0038] Aspects and any possible implementation manners described above, further provide an implementation manner, and S4 specifically comprises:

[0039] S41, perform ROI feature extraction on the Patch sequence :

[0040]

[0041] Wherein, f roi is the ROI feature vector, F img is the image feature map, and bbox is the region of interest coordinates.

[0042] S42, update the cell graph G = (V, E, A) using a 2-layer GraphSAGE network:

[0043]

[0044] Wherein, is the node feature of the l-th layer GNN, W (l) is the l-th layer GNN parameter, is the neighbor node set of node i, and sigma is the activation function.

[0045] S43, perform graph-level pooling to obtain cell graph features:

[0046]

[0047] Wherein, g tile is the graph-level pooling feature vector, f i 2 is the final layer feature of the node.

[0048] S44, take the ROI feature as a query: Q = W q f roi , The cell graph feature g tile and the unified semantic feature u are linearly mapped into key-value pairs respectively:

[0049]

[0050] wherein W q is the query projection weight, W g and W g ' are the weights for linearly mapping g tile to K / V space, W u and W u ' are the weights for linearly mapping semantic feature vector u to K / V space, and d is the attention hidden space dimension;

[0051] S45, calculating the attention weight obtaining the attention fused feature according to the attention weight

[0052] S46, calculating the dynamic gating coefficient γ = σ(W c [f roi ||z tri ]), W c ∈ R 1×2d , and performing fusion to obtain the unified ROI feature

[0053] S47, constructing a detection model, wherein the detection head of the detection model comprises a classification head, a bounding box regression head and a mask head, wherein the classification head is represented by the following formula: p = softmax(W cls f * +b cls ), the bounding box regression head is represented by the following formula: Δ bbox = W reg f * +b reg , and the mask head is represented by the following formula:

[0054] The comprehensive training loss function of the detection model is represented by the following formula:

[0055]

[0056] wherein the cross-modal consistency loss function is the region proposal network loss, is the ROI level category classification loss, is the ROI level bounding box regression loss, is the ROI level segmentation mask loss, λ box is the bounding box regression loss weight, λ mask is the mask loss weight, and λ alignFor consistency loss weight, m represents hinge loss margin, d(·,·) represents distance metric between features, W cls represents classification head weight, b cls represents classification head bias, W reg represents bounding box regression weight, b reg represents bounding box regression bias, Deconv (4) represents up-sampling 4x deconvolution layer, Conv 1×1 represents 1x1 convolution to generate mask channel.

[0057] Aspects and any possible implementation manners described above, further provide an implementation manner, the S5 specifically comprises:

[0058] S51, calculate the probability that each ROI is determined to be cancerous by the detection model:

[0059] p i = Softmax(W cls · f roi +b cls ), i = 1, 2,..., N;

[0060] Wherein, W cls and b cls are the weight and bias of the classification head, f roi is the feature extracted from the image feature by ROIAlign;

[0061] S52, according to the probability p i output by the model, select the region A that meets the following probability range as the region with high probability of cancer: A = {ROI i | p i ≥ τ high , p i ∈ [0, 1]}; select the region B that meets the following probability range as the region with high probability of cancer: B = {ROI j | p j ≤ τ low , p i ∈ [0, 1]}; and form a preference pair (A, B) by the high probability region A and the low probability region B;

[0062] S53, for a pair of preference pairs (A, B), train the reward model by constructing a region contrast loss function, which is represented as: σ is the Sigmoid function, r(A) and r(B) are the output scores of the reward model for A and B respectively;

[0063] S54, minimize the contrast loss function update the model parameters of the reward model by back propagation;

[0064] S55, the PPO loss function is represented as: ρ t is the ratio of the current policy to the old policy, A t is the advantage estimate, and ∈ is a smoothing factor.

[0065] Aspects and any possible implementation manners described above, further provide an implementation manner, the two image perspectives are selected from data augmentation categories and / or geometric transformation categories, the data augmentation categories are divided into five categories of geometric transformation, intensity transformation, affine and nonlinear transformation, noise injection and filtering, the geometric transformation includes rotation, scaling, flipping and cropping, the intensity transformation includes brightness, contrast, saturation and hue, the noise injection is adding random noise, the affine and nonlinear transformation includes simulating deformation of organs and different morphologies of lesions, and the filtering includes blur filter and sharpening filter.

[0066] Aspects and any possible implementation manners described above, further provide an implementation manner, the method further comprises:

[0067] The image blocks are respectively cut into image patches of different sizes at multiple scales to obtain a corresponding multi-scale image patch sequence;

[0068] The multi-scale image patch sequence is respectively generated into multi-scale visual features by a visual neural network;

[0069] The cell graph nodes are clustered according to the sizes of the cell nuclei, and a plurality of scale cell subgraphs are respectively constructed;

[0070] The multi-scale visual features and the corresponding scale cell subgraph features are fused to generate multi-scale fusion features;

[0071] The multi-scale fusion features are fused from top to bottom to form a unified multi-scale pyramid feature representation, and the unified multi-scale pyramid feature representation is used as an input feature of the detection model.

[0072] In a second aspect, an embodiment of the present application provides an endometrial cancer cell detection system, the system comprising:

[0073] A preprocessing module is configured to pair the image blocks with the pathological report texts, obtain a visual encoder, a text encoder and a unified semantic feature by training a multi-positive sample contrast loss function;

[0074] A masking module is configured to cut the image blocks into image patches, dynamically adjust a masking ratio based on semantic entropy, perform masking processing according to the masking ratio, and optimize the visual encoder by training a comprehensive loss function.

[0075] A construction module is configured to construct a cell graph by taking cells as nodes and adjacent relationships as edges after locating the nucleus center by the optimized visual encoder;

[0076] A processing module is configured to construct a detection model based on the unified semantic feature, the image feature and the cell graph feature;

[0077] A detection module is configured to obtain a high-probability cancer region and a high-probability cancer region to form a preferred pair according to the current prediction result of the detection model, train a reward model according to the preferred pair, and adjust the detection model through reinforcement learning.

[0078] One of the above technical solutions has the following beneficial effects:

[0079] The method and system for detecting endometrial cancer cells proposed in the method of the embodiment of the application first pair the image blocks with the pathological report texts, obtain unified semantic features by using multi-positive sample contrast learning, then perform semantic entropy-driven dynamic masking-random shuffling self-supervised training of a visual neural network, then locate the nucleus, construct a weighted cell graph by taking cells as nodes and adjacent relationships as edges, and obtain topological features, then generate multi-modal features by cross-scale attention and gate fusion of the visual features and the semantic features, and finally the detection head jointly outputs the category, the bounding box and the mask, and is supervised or weakly supervised by the cross-modal consistency loss. Further, the high-probability and low-probability regions output by the initial model are automatically generated to form a “cancer-normal” preferred pair, a reward model is trained, and PPO policy iteration fine-tuning is adopted to realize precise and interpretable detection of endometrial cancer cells. The technical solution can automatically learn cell-level lesion features under the condition of only providing slice-level or weak annotation, output category, bounding box and mask information, and significantly reduce the burden of manual reading of high-resolution pathological sections. At the same time, through continuous iteration of the adaptive preferred pair and the reward model, the detection strategy can be continuously improved by the data, and the robust generalization to different staining batches and scanning devices is maintained, and the pathological screening efficiency and accuracy are improved as a whole. BRIEF DESCRIPTION OF DRAWINGS

[0080] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0081] Figure 1 The flowchart of the method for detecting endometrial cancer cells provided by the embodiment of the application is shown in the figure.

[0082] Figure 2A schematic block diagram of an endometrial carcinoma cell detection system provided by an embodiment of the present application is shown in the figure;

[0083] Figure 3 A schematic diagram of the hardware structure of an endometrial carcinoma cell detection system provided by an embodiment of the present application is shown in the figure.

DETAILED DESCRIPTION

[0084] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in detail with specific embodiments and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0085] Please refer to Figure 1 , which is a flowchart of an endometrial carcinoma cell detection method provided by an embodiment of the present application, as shown in the figure, the method comprises the following steps:

[0086] S1, pair the image blocks with the pathology report text, obtain the visual encoder, the text encoder and the unified semantic feature by training the multi-positive sample contrast loss function;

[0087] S2, divide the image blocks into image small blocks, dynamically adjust the masking proportion based on the semantic entropy, perform masking processing according to the masking proportion, and optimize the visual encoder by training the comprehensive loss function;

[0088] S3, locate the nucleus center through the optimized visual encoder, then take the cells as nodes and the adjacent relationship as edges to construct a cell graph;

[0089] S4, construct a detection model based on the unified semantic feature, the image feature and the cell graph feature;

[0090] S5, according to the current prediction result of the detection model, obtain the area with high probability of cancer and the area with high probability of cancer, form a preference pair, then train the reward model according to the preference, and adjust the detection model through reinforcement learning.

[0091] Specifically, S1 further comprises the following steps:

[0092] Obtain panoramic digital pathology sections and pathology report texts, cut each panoramic digital pathology section into multiple image blocks x with a size of H0xW0, and extract the corresponding pathology report text t.

[0093] It should be noted that the WSI (Whole-Slide Image, panoramic digital pathology section) is a large image file generated by scanning the entire tissue section on the slide in ultra-high resolution, which saves all the spatial information of the lesion, and is convenient for subsequent algorithm processing and doctor zooming. The report corresponding to each WSI is given by a pathologist, usually containing semantic information such as sampling site, microscopic description, diagnostic conclusion, grading / type, etc. Due to the large size of WSI, GPU cannot process it all at once, and needs to be cut according to a fixed window.

[0094] Specifically, the S1 specifically includes:

[0095] S11, generating two kinds of perspectives {x (1) ,x (2)} based on the image blocks x, and generating two kinds of sentence rewriting {t (1) ,t (2)} based on the corresponding pathological report text t to form a four-dimensional vector {x (i) ,t (j)}, wherein i, j ∈ {1, 2};

[0096] S12, extracting a visual vector v = f θ (x (k) ) of the image block by a visual neural network, extracting a text vector u = g φ (t (k) ) of the text by a text neural network, and obtaining a fusion vector z = γv + (1-γ)u by gate fusion with an additional gate coefficient, wherein γ is the gate coefficient, γ = σ(W g [v||u]), σ is a Sigmoid activation function, W g is a weight matrix of the gate, f θ is a visual neural network processing, x (k) is an image feature of the kth perspective, and g φ is a text neural network processing, t (k) is the kth text perspective;

[0097] S13, constructing a multi-positive sample contrast loss function for training, and minimizing the loss function:

[0098]

[0099] wherein P is a positive sample pair set, T is a temperature coefficient, is an embedding vector of the i th image, is an embedding vector of the j th text, and <·, ·> is a dot product operation;

[0100] S14, outputting the visual encoder, the text encoder and the unified semantic feature obtained by training.

[0101] It should be noted that the positive samples can be expanded by the dual-view text-image to alleviate the small sample, noise and bias writing problems, and encourage the model to learn the representation corresponding to the medical semantic robustness instead of relying on single view details. The two image views are selected from data augmentation categories and / or geometric transformation categories, which are divided into five categories: geometric transformation, intensity transformation, affine and nonlinear transformation, noise injection and filtering. The geometric transformation includes rotation, scaling, flipping and cropping. The intensity transformation includes brightness, contrast, saturation and hue. The noise injection is to add random noise. The affine and nonlinear transformation includes simulating the deformation of organs and different morphologies of lesions. The filtering includes blur filter and sharpening filter.

[0102] In addition, the gating coefficient γ allows the model to dynamically determine the reliability of image evidence and text evidence according to the sample, overcoming the interference of missing / redundant information in the report text, and obtaining a cross-modal shared semantic space that provides a "soft label" prior for the subsequent detection network.

[0103] Compared with the prior art processing method, each anchor has multiple positive samples at the same time, which fully utilizes the four-view combination, and the visual-textual representation is aggregated on the same case and separated between different cases, eliminating the early training of image-text alignment. This step establishes a unified semantic coordinate system for images and texts, so that all Patch, ROI and cell nodes can be consistent with the pathological semantic measurement, which significantly improves the robustness and interpretability of cancer cell detection.

[0104] Specifically, the S2 specifically includes:

[0105] S21, obtaining a Patch sequence by dividing the image block into P*P pixel image blocks through a visual encoder wherein N=(H0 / P) * (W0 / P);

[0106] S22, calling a unified semantic feature v * for each image block, obtaining a class probability distribution p=softmax(W c v * ) after projection, and calculating a normalized semantic entropy:

[0107]

[0108] wherein W c is a weight matrix for class prediction, p j is a prediction probability of the jth class, and C is the number of classification categories;

[0109] S23, dynamically determining a shielding proportion according to the semantic entropy:

[0110] ​

[0111] Where σ is the Sigmoid function, and α and τ are hyperparameters;

[0112] S24. According to the shading ratio p mask A set is formed by randomly sampling from the Patch sequence. set The inner patch is masked, and then the token vectors of the unsampled patches are randomly permuted to obtain a disordered sequence. Among them, P perm Let h be a random permutation matrix. vis The token sequence of the unmasked patch;

[0113] S25, will The input visual encoder yields the representation H enc Global branch prediction And minimize the global branch loss function For each shaded location, local branches are used. predict And minimize the local branch loss function Where CLS is the neural network feature aggregation vector, t mask The token representation of the obscured location. For the local reconstruction feature at position i, The true feature of the i-th position;

[0114] S26. By integrating the loss function Jointly optimize the visual encoder; where λ g λ represents the weighting coefficients of the global loss. l These are the weighting coefficients for local losses.

[0115] It should be noted that the above steps discretize the large blocks of WSI into learnable tokens using the ViT-MAE method, preparing for subsequent self-supervised reconstruction. The self-supervised strategy of "semantic-guided dynamic occlusion + random permutation" allows the visual encoder to adaptively focus on regions rich in diagnostic value; combined with global / local dual reconstruction and a unified semantic objective, it can significantly improve the encoder's robust representation ability of morphological differences, color variations, and tissue structures, providing a more accurate and interpretable feature basis for subsequent cell map modeling and detection networks, thereby improving the accuracy and reliability of endometrial cancer cell screening.

[0116] Specifically, S3 includes:

[0117] S31. Obtain the center coordinates c of each cell nucleus node using the optimized visual encoder. i =(x i ,yi ), i = 1...N c , and extract the initial feature vector of the cell v i is defined as (c i , f i ) (0) , forming a node set V where r i is the morphological radius of the i-th nucleus, μ i is the average chromaticity of the i-th nucleus, s i is the class logits of the i-th nucleus, e i is the visual embedding vector of the i-th nucleus, c i = (x i , y i ) is the center coordinate of the i-th nucleus, d0 is the initial feature dimension, N c is the total number of cell nodes;

[0118] S32, perform Delaunay triangulation on the center coordinates of all nucleus nodes to obtain an initial candidate adjacency set E raw , prune edges with a length exceeding the maximum connection distance threshold r max , and obtain the candidate adjacency set E = {(i, j) || |c i - c j | |≤ r max}; where | |c i - c j | | is the Euclidean distance between the i-th and j-th nodes;

[0119] S33, calculate the geometric weight and the morphological weight of each edge (i, j) ∈ E, multiply the geometric weight and the morphological weight to obtain the comprehensive weight a ij , and construct a sparse weighted adjacency matrix A = [a ij ];

[0120] S34, output the cell graph containing the node set V, the candidate adjacency set E, and the weighted adjacency matrix A

[0121] G = (V, E, A).

[0122] This process first locates and quantifies the morphological-semantic features of each nucleus, then generates candidate edges that conform to the tissue structure using Delaunay + distance threshold, and finally fuses geometric and morphological similarity to construct a sparse weighted adjacency matrix. The cell graph obtained in this way is not only concise but also biologically reasonable, and can accurately model the interaction between cells in the graph neural network, significantly improving the ability to distinguish endometrial cancer cell clusters, infiltration boundaries, and microenvironment heterogeneity.

[0123] Aspects and any possible implementation mentioned above, further provide an implementation, the S4 specifically comprises:

[0124] S41, performing ROI feature extraction on the Patch sequence

[0125]

[0126] Wherein, f roi is the ROI feature vector, F img is the image feature map, and bbox is the region of interest coordinates;

[0127] S42, updating the cell graph G=(V, E, A) using a 2-layer GraphSAGE network:

[0128]

[0129] Wherein, is the node feature of the l-th layer GNN, W (l) is the l-th layer GNN parameter, is the neighbor node set of node i, and σ is the activation function;

[0130] S43, performing graph-level pooling to obtain cell graph features:

[0131]

[0132] Wherein, g tile is the graph-level pooling feature vector, f i 2 is the final layer feature of the node;

[0133] S44, taking the ROI feature as a query: The cell graph feature g tile and the unified semantic feature u are linearly mapped into key-value pairs respectively:

[0134]

[0135] Wherein, W q is the query projection weight, W g and W g ' are weights for linearly mapping g tile to K / V space, W u and W u ' are weights for linearly mapping the semantic feature vector u to K / V space, and d is the attention hidden space dimension;

[0136] S45, calculating the attention weight ​The attention fusion feature is obtained according to the attention weight

[0137] S46, calculate the dynamic gating coefficient γ = σ(W c f roi ||z tri ), W c ∈R 1×2d , and perform fusion to obtain the unified ROI feature

[0138] S47, construct a detection model, and the detection head of the detection model comprises a classification head, a bounding box regression head and a mask head, wherein the classification head is represented by the following formula: p = softmax(W cls f * +b cls ), the bounding box regression head is represented by the following formula: Δ bbox = W reg f * +b reg , and the mask head is represented by the following formula:

[0139] The comprehensive training loss function of the detection model is represented by the following formula:

[0140]

[0141] Wherein, the cross-modal consistency loss function is the region proposal network loss, is the ROI level category classification loss, is the ROI level bounding box regression loss, is the ROI level segmentation mask loss, λ box is the bounding box regression loss weight, λ mask is the mask loss weight, λ align is the consistency loss weight, m represents the hinge loss margin, d(·,·) represents the distance measurement between features, W cls represents the classification head weight, b cls represents the classification head bias, W reg represents the bounding box regression weight, b reg represents the bounding box regression bias, Deconv (4) represents the 4x up-sampling deconvolution layer, Conv 1×1 represents the 1x1 convolution to generate the mask channel.

[0142] According to the above-mentioned aspects and any possible implementation, further provided is an implementation, and the S5 specifically comprises:

[0143] S51, calculate the probability of each ROI being determined as cancerous by detecting the model:

[0144] p i i=1, 2,..., N; cls roi cls ), i=1, 2,..., N;

[0145] where W cls and b cls are the weights and biases of the classification head, f roi is the feature extracted from the image feature by ROIAlign;

[0146] S52, according to the probability p i output by the model, select the region A that meets the following probability range as the region with high probability of cancer: A={ROI i |p i ≥τ high , p i ∈[0, 1]}; select the region B that meets the following probability range as the region with high probability of cancer: B={ROI j |p j ≤τ low , p i ∈[0, 1]}; and form a preference pair (A, B) by the high probability region A and the low probability region B;

[0147] S53, for a pair of preference pairs (A, B), train the reward model by constructing a region contrast loss function, which is expressed as: σ is the Sigmoid function, r(A) and r(B) are the output scores of the reward model for A and B, respectively;

[0148] S54, minimize the contrast loss function update the model parameters of the reward model by backpropagation;

[0149] S55, fine-tune the reward model by constructing a reinforcement learning loss function, which is expressed as: ρ t is the ratio of the current policy to the old policy, A t is the advantage estimate, and ∈ is the smoothing factor.

[0150] Aspects and any possible implementation manners described above, further provide an implementation manner, the method further comprises:

[0151] cut the image blocks into image small blocks of different sizes at multiple scales respectively, to obtain a corresponding multi-scale image small block Patch sequence;​​

[0152] respectively generate multi-scale visual features through a visual neural network;

[0153] According to the size of the cell nucleus, the cell graph nodes are clustered, and a plurality of scale cell subgraphs are constructed;

[0154] The multi-scale visual features and the corresponding scale cell subgraph features are fused to generate multi-scale fusion features;

[0155] The multi-scale fusion features are fused from top to bottom to form a unified multi-scale pyramid feature representation, which is used as the input feature of the detection model.

[0156] Through the above steps, the following technical effects are achieved:

[0157] 1. Improve the segmentation accuracy and robustness:

[0158] Fusion of shape prior: By introducing shape prior information in the graph convolution network, the model can better capture the shape features of the target object, especially in the ultrasound images with fuzzy boundaries and more noise, effectively improving the accuracy of segmentation.

[0159] Application of graph convolution network: GCN can handle irregular graph structure data, suitable for combining pixel and shape key point information, realize modeling of complex topological structure, and improve the recognition ability of target region.

[0160] 2. Enhance the generalization ability of the model:

[0161] Self-supervised pre-training: Use a large number of unlabeled ultrasound images for self-supervised learning to pre-train the feature extraction model, obtain a general image feature representation, reduce the dependence on a large amount of labeled data, and improve the generalization performance of the model in small sample cases.

[0162] Multi-task learning framework: Through denoising and feature extraction at the same time, the model learns more rich and robust feature representation under the condition of sharing parameters, complements the information between different tasks, and improves the overall performance.

[0163] 3. Improve the accuracy of detection and classification:

[0164] Accurate positioning of candidate regions: Through fine segmentation results, candidate regions are extracted to ensure high coincidence between candidate regions and actual target regions, reduce background interference, and improve the accuracy of subsequent classification.

[0165] Effective use of high-level features: Region features are extracted from high-level features, fully utilizing deep semantic information, and enhancing the recognition ability of the classification model for lesion features.

[0166] Probability output: The classification model outputs the probability of candidate regions containing targets and the probability distribution of lesion categories, providing quantitative basis for clinical diagnosis, facilitating doctors to make risk assessment and decision.

[0167] 4. Real-time requirements are met and processing efficiency is improved:

[0168] Lightweight network architecture: Lightweight neural network structures such as MobileNet are used to reduce the computational complexity of the model and meet real-time processing requirements.

[0169] Model optimization and acceleration: Through model pruning, quantization and hardware acceleration, the inference speed and resource utilization efficiency of the model are further improved, suitable for deployment in actual medical devices.

[0170] 5. Improved model interpretability and controllability:

[0171] Introduction of shape prior: The model's use of shape information is clarified, making the segmentation results more interpretable and helping doctors understand the model's judgment basis.

[0172] Adjustable shape constraint: Through the weight adjustment of shape prior loss, the model's dependence on shape prior can be controlled, adapting to the needs of different application scenarios.

[0173] 6. Reduced dependence on labeled data, saving labor costs:

[0174] Self-supervised learning and multi-task learning: Pre-training with unlabeled data reduces the need for large amounts of labeled data, and in the multi-task learning framework, the learning process of denoising and feature extraction is shared, saving data labeling and model training costs.

[0175] In summary, the method of the embodiment of the present application proposes an endometrial cancer cell detection method and system, which first pairs the image blocks with the pathological report text, uses multi-positive sample contrast learning to obtain unified semantic features, then performs semantic entropy driven dynamic masking-random shuffling self-supervised training of the visual neural network, then locates the cell nucleus, constructs a weighted cell graph with cells as nodes and adjacent relationships as edges, and obtains topological features, and then generates multi-modal features through cross-scale attention and gate fusion with visual features and semantic features, the detection head jointly outputs the category, the bounding box and the mask, and is supervised or weakly supervised by the cross-modal consistency loss, and further uses the high and low probability regions output by the initial model to automatically generate the 'cancer-normal' preference pair, trains the reward model, and uses the PPO policy iteration fine-tuning to realize the precise and interpretable detection of the endometrial cancer cells. The technical scheme of the present application can automatically learn the cell-level lesion features under the condition of only providing slice-level or weak annotation, output the category, the bounding box and the mask information, and significantly reduce the artificial reading burden of high-resolution pathological sections; at the same time, through the continuous iteration of the adaptive preference pair and the reward model, the detection strategy can be continuously improved by the data, and the robust generalization to different staining batches and scanning devices is maintained, and the pathological screening efficiency and accuracy are improved as a whole.

[0176] The embodiment of the present application further provides a device for implementing each step and method in the above method embodiment.

[0177] Please refer to Figure 2 which is a schematic block diagram of an endometrial cancer cell detection system provided by the embodiment of the present application. The system comprises:

[0178] The preprocessing module 210 is configured to pair the image blocks with the pathological report text, and obtain a visual encoder, a text encoder and unified semantic features by training a multi-positive sample contrast loss function;

[0179] The masking module 220 is configured to split the image blocks into image patches, dynamically adjust the masking ratio based on the semantic entropy, perform masking processing according to the masking ratio, and optimize the visual encoder by training a comprehensive loss function;

[0180] The construction module 230 is configured to locate the cell nucleus center through the optimized visual encoder, and then construct a cell graph with cells as nodes and adjacent relationships as edges;

[0181] The processing module 240 is configured to construct a detection model based on the unified semantic features, the image features and the cell graph features;

[0182] The detection module 250 is used to obtain regions with high probability of cancer and regions with high probability of cancer based on the current prediction results of the detection model and form a preference pair. Then, a region contrast loss function is constructed based on the preference to train the reward model, and the detection model is adjusted through reinforcement learning.

[0183] Because each unit module in this embodiment can execute Figure 1 For the methods shown, parts not described in detail in this embodiment, please refer to the [examples / descriptions]. Figure 1 Related explanations.

[0184] Please refer to Figure 3 This is a schematic diagram of the hardware structure of the ultrasound image segmentation system for adenomyosis provided in an embodiment of the present invention. The data prediction device includes at least one processor and a memory. The at least one processor is coupled to the memory and is used to read and execute instructions in the memory to perform the endometrial cancer cell detection method provided in an embodiment of the present invention.

[0185] Thirdly, embodiments of the present invention provide a computer-readable medium. The computer-readable medium stores program code that, when executed on a computer, causes the computer to perform the endometrial cancer cell detection method provided in the embodiments of the present invention.

[0186] At the hardware level, the device may include a processor, and optionally an internal bus, network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the device may also include other hardware required for other business operations.

[0187] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be categorized into address bus, data bus, control bus, etc.

[0188] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0189] The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware code processing executed by a hardware code processor, or executed by a combination of hardware and software modules in the code processor. The software modules can be located in storage media in the art such as random access memories (RAMs), flash memories, read only memories (ROMs), programmable read only memories (PROMs), electrically programmable read only memories (EPROMs), electrically erasable and programmable read only memories (EEPROMs), registers, or the like. The storage media is located in the storage, and the processor reads information in the storage media to execute the steps of the above-described methods in combination with hardware.

[0190] The systems, apparatuses, modules or units illustrated by the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0191] For the convenience of description, the above apparatuses are described in functions as various units or modules respectively described. Of course, the functions of the units or modules can be implemented in the same or multiple software and / or hardware in the implementation of the present application.

[0192] Those skilled in the art should understand that embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0193] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The apparatuses specified in a flow or multiple flows and / or blocks

[0194] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0195] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0196] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0197] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the computer stores information about an operating system, application software, and / or the like. Memory is an example of computer readable media.

[0198] Computer readable media includes permanent and non-permanent, moveable and non- moveable media that can be implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disks (DVDs) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that is accessible to a computing device. According to the definition provided herein, a computer readable medium excludes transitory computer readable media, such as modulated data signals and carrier waves.

[0199] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0200] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0201] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0202] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0203] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for detecting endometrial cancer cells, characterized in that, The method includes: S1. Pair large image blocks with pathology report text, and obtain visual encoder, text encoder and unified semantic features by training a multi-positive-sample contrast loss function; S2. Divide large image blocks into smaller image blocks, dynamically adjust the occlusion ratio based on semantic entropy, perform occlusion processing according to the occlusion ratio, and optimize by training a visual encoder with a comprehensive loss function. S3. After locating the center of the cell nucleus using the optimized visual encoder, a cell graph is constructed using cells as nodes and neighbor relationships as edges. S4. Construct a detection model based on unified semantic features, image features, and cell diagram features; S5. Based on the current prediction results of the detection model, obtain regions with high probability of cancer and regions with high probability of cancer and form preference pairs. Then, construct a region contrast loss function based on the preference to train the reward model, and then adjust the detection model through reinforcement learning. S2 specifically includes: S21. Large blocks of image are processed using a visual encoder. Cut into Image patches of pixels are used to obtain a Patch sequence. ,in, ; S22. Apply unified semantic features to each large image block. After projection, the category probability distribution is obtained. And calculate the normalized semantic entropy: ； in, This is the weight matrix for class prediction. Let be the predicted probability of class j. The number of categories; S23. Dynamically determine the masking ratio based on semantic entropy: ； in, It is the Sigmoid function. and It's a hyperparameter; S24. According to the shading ratio A set is formed by randomly sampling from the Patch sequence. , will set The inner patch is masked, and then the token vectors of the unsampled patches are randomly permuted to obtain a disordered sequence. ;in, It is a random permutation matrix. The token sequence of the unmasked patch; S25, will The input visual encoder yields the representation Global branch prediction And minimize the global branch loss function For each shaded location, local branches are used. predict And minimize the local branch loss function ;in, This is a summary vector of neural network features. The token representation of the obscured location. For the local reconstruction feature at position i, The true feature of the i-th position; S26. By integrating the loss function Jointly optimize the visual encoder; among which, These are the weighting coefficients for the global loss. These are the weighting coefficients for local losses.

2. The method for detecting endometrial cancer cells according to claim 1, characterized in that, Before S1, the following also applies: Obtain panoramic digital pathology slides and pathology report text, and crop each panoramic digital pathology slide to a size of [size missing]. Multiple large image blocks Extract the corresponding pathology report text. .

3. The method for detecting endometrial cancer cells according to claim 2, characterized in that, S1 specifically includes: S11, Based on large image blocks Generate two perspectives Meanwhile, based on the corresponding pathology report text Generate two sentence rewrites Forming a four-dimensional vector ,in, ; S12. Extract visual vectors of large blocks of the image using a visual neural network. The text vector of the text is extracted using a text neural network. The fusion vector is obtained by adding gating coefficients through gating fusion. , The gating coefficient, , The Sigmoid activation function is used. Here is the weight matrix for gating. For visual neural network processing, For the image features of the k-th viewpoint, For text neural network processing, The k-th text perspective; S13. Construct a multi-positive-sample contrast loss function for training, and minimize this loss function: ； in, For the set of positive sample pairs, For temperature coefficient, Let be the embedding vector of the i-th image. Let be the embedding vector of the j-th text. This is a dot product operation; S14. Output the trained visual encoder, text encoder, and unified semantic features.

4. The method for detecting endometrial cancer cells according to claim 3, characterized in that, S3 specifically includes: S31. Obtain the center coordinates of each cell nucleus node using the optimized visual encoder. , And extract the initial feature vector of the cell. ; Graph nodes Defined as , forming a node set ;in, Let be the morphological radius of the i-th cell nucleus. Let i be the average chromaticity of the i-th cell nucleus. Let logits be the category of the i-th cell nucleus. Let be the visual embedding vector of the i-th cell nucleus. Let be the center coordinates of the i-th cell nucleus. As the initial feature dimension, This represents the total number of cell nodes; S32. Perform Delaunay triangulation on the center coordinates of all cell nucleus nodes to obtain the initial candidate adjacency set. Pruning a side whose length exceeds the maximum connection distance threshold Find the edges and obtain the candidate adjacency set. ;in, Let be the Euclidean distance between the i-th and j-th nodes; S33. Calculate each edge geometric weights and shape weight The geometric weight and the morphological weight are multiplied together to obtain the comprehensive weight. Construct a sparse weighted adjacency matrix ; S34, Output includes forming a node set Candidate Adjacency Set and weighted adjacency matrix Cell diagram 。 5. The method for detecting endometrial cancer cells according to claim 4, characterized in that, S4 specifically includes: S41. For the Patch sequence Perform ROI feature extraction: ； in, For the ROI feature vector, For image feature maps, Coordinates of the region of interest; S42. Use a 2-layer GraphSAGE network to visualize cell graphs. Update: ； in, The node features of the l-th layer GNN, For the parameters of the l-th layer GNN, Let be the set of neighboring nodes of node i. For activation functions; S43. Perform graph-level pooling to obtain cell map features: ； in, For graph-level pooling feature vectors, The final layer features of the node; S44. Use ROI features as the query: Cellular map features and unified semantic features Linearly mapped to key-value pairs respectively: ； in, To query the projected weights, and To be Weights that are linearly mapped to the K / V space. and The weights are used to linearly map the semantic feature vector u to the K / V space, where d is the dimension of the attention latent space; S45. Calculate attention weights Features of attention fusion are obtained based on attention weights. ; S46. Calculate the dynamic gating coefficient. The fusion process yields a unified ROI feature. ; S47. Construct a detection model whose detection head includes a classification head, a bounding box regression head, and a mask head. The classification head is represented by the following formula: The bounding box regression head is represented by the following formula: The mask head is represented by the following formula: ; The overall training loss function of the detection model is expressed by the following formula: ； Among them, the cross-modal consistency loss function , For regional network loss, For ROI level category classification loss, For ROI-level bounding box regression loss, For ROI-level segmentation mask loss, For bounding box regression loss weights, For mask loss weights, For consistency loss weights, This indicates the hinge loss margin. Represents the distance metric between features. Indicates the category head weight. This indicates the category header bias. Indicates the bounding box regression weights. This indicates the bounding box regression bias. This indicates a deconvolution layer with an upsampling of 4×. This indicates that a 1×1 convolution generates a mask channel.

6. The method for detecting endometrial cancer cells according to claim 5, characterized in that, S5 specifically includes: S51. Calculate the probability that each ROI is identified as cancerous using the detection model: ； in, and The weights and biases of the classification heads. Features extracted from image features using ROIAlign; S52. Based on the probability output by the model Region A, which falls within the following probability range, is selected as a region with a high probability of cancer: Region B, which falls within the following probability range, is selected as the region with a high probability of cancer: ; and by constructing preference pairs through high-probability region A and low-probability region B. ; S53, For a pair of preference pairs The reward model is trained by constructing a region contrast loss function, which is expressed as follows: , It is the Sigmoid function. and These are the output scores of the reward model for A and B, respectively. S54. Minimize the contrastive loss function The model parameters of the reward model are updated through backpropagation; S55. Fine-tune the reward model by constructing a reinforcement learning loss function. The PPO loss function is expressed as: , It is the ratio of the current strategy to the old strategy. It is an advantage estimate. It is a smoothing factor.

7. The method for detecting endometrial cancer cells according to claim 6, characterized in that, The two image perspectives are selected from data augmentation types and / or geometric transformation types. The data augmentation types are divided into five types: geometric transformation, intensity transformation, affine and nonlinear transformation, noise injection, and filtering. Geometric transformation includes rotation, scaling, flipping, and cropping. Intensity transformation includes brightness, contrast, saturation, and hue. Noise injection adds random noise. Affine and nonlinear transformations simulate organ deformation and different morphologies of lesions. Filtering includes blur filters and sharpening filters.

8. The method for detecting endometrial cancer cells according to claim 7, characterized in that, The method further includes: The large image block is divided into smaller image blocks of different sizes at multiple scales to obtain the corresponding multi-scale image patch sequence. The multi-scale image patch sequences are processed by a visual neural network to generate multi-scale visual features. The cell graph nodes are clustered according to the cell nucleus size to construct cell sub-graphs of multiple scales. Multi-scale visual features are fused with corresponding scale cell sub-image features to generate multi-scale fused features. The multi-scale fusion features are fused from top to bottom using a feature pyramid to form a unified multi-scale pyramid feature representation, which is then used as the input feature of the detection model.

9. A system for detecting endometrial cancer cells using the method according to any one of claims 1-8, characterized in that, The system includes: The preprocessing module is used to pair large image blocks with pathology report text, and obtains a visual encoder, a text encoder and unified semantic features by training a multi-positive-sample contrast loss function; The masking module is used to divide large image blocks into smaller image blocks, dynamically adjust the masking ratio based on semantic entropy, perform masking processing according to the masking ratio, and optimize by training a visual encoder with a comprehensive loss function. The module is used to locate the center of the cell nucleus using an optimized visual encoder, and then construct a cell graph with cells as nodes and neighbor relationships as edges. The processing module is used to construct a detection model based on unified semantic features, image features, and cell graph features; The detection module is used to obtain regions with a high probability of cancer and regions with a high probability of cancer based on the current prediction results of the detection model and form a preference pair. Then, a region contrast loss function is constructed based on the preference to train the reward model, and the detection model is adjusted through reinforcement learning.

Images