Early cancer screening model based on cloud architecture and construction method

Abstract

The invention discloses an early cancer screening model based on a cloud architecture and a construction method, and particularly relates to the technical field of medical artificial intelligence. The method comprises the following steps: firstly, collecting and normalizing heterogeneous medical image and biomarker data of a plurality of geographic areas, and constructing a unified data set; then extracting image features and marker features by adopting multi-task learning, fusing equipment and region attributes, establishing a joint representation vector, and training a preliminary screening model; after the model is deployed to a cloud end, target area data is adapted through online transfer learning, and an area model is generated; further extracting an activation heat map, identifying a weak signal high response area in combination with sensitivity analysis, comparing with manual annotation, constructing a local weight adjustment strategy based on errors, optimizing the model, and outputting a final model; the method improves the recognition capability of early lesions under the condition of low signal-to-noise ratio, is suitable for remote and basic screening scenes, and has a high practical value.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to an early cancer screening model based on cloud architecture and its construction method. Background Technology

[0002] In recent years, with the increasing aging of the population and the rise in cancer incidence, early detection of cancer is crucial for improving treatment success rates and reducing mortality. However, in many primary healthcare units, especially in remote mountainous areas and resource-scarce regions, cancer screening is generally delayed or absent due to inadequate equipment and a shortage of professional radiologists, missing the optimal intervention window and placing a heavy burden on society and families.

[0003] Currently, some AI-based cancer screening models exist, but most of these models rely on locally deployed high-performance hardware, making them difficult to adapt to edge healthcare scenarios. Furthermore, existing models often use static samples for training during their construction, failing to consider dynamic factors such as different geographical regions, population genetic variations, and differences in screening equipment performance, leading to a significant decrease in screening accuracy and adaptability.

[0004] Furthermore, current models rarely possess the capability of "stepwise enhancement-regional adaptation-dynamic migration," making it difficult to respond with high sensitivity to specific weak signal regions (such as very early small lesions), thus missing suspicious signals. The challenges of early cancer screening are particularly severe in scenarios involving low-dose CT images, blurred endoscopic images, and weak biomarker expression. Summary of the Invention

[0005] The purpose of this invention is to provide an early cancer screening model and construction method based on cloud architecture to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an early cancer screening model and construction method based on cloud architecture, comprising:

[0007] Medical image data and patient history biomarker data from multiple geographical regions were collected to construct a heterogeneous raw dataset D, including image features, gene expression features and device label information;

[0008] The heterogeneous original dataset D is subjected to regional feature normalization processing to obtain the normalized dataset D′, where D′ contains the device normalization label E and the geographic variation parameter G for each data sample;

[0009] Preliminary cancer screening model based on D′ The training employs a multi-task learning method to extract image domain feature vectors separately. and biomarker feature vectors And establish a joint representation vector ;

[0010] Deploying the model to a cloud architecture-based training platform, adjusting the real-time sample data uploaded by the target region T using an online transfer learning mechanism to obtain a region-adapted model MT;

[0011] Extracting an activation heat map HT from the model MT, identifying a weak signal high response region RT through feature sensitivity analysis, which is used to indicate the image or biomarker region of the early potential lesion of cancer;

[0012] Comparing the RT region with the artificial annotation result, locally adjusting the screening weight of the model MT, and outputting a final model MF, which is used to identify the early weak feature region of cancer under low signal-to-noise ratio conditions.

[0013] Preferably, the training of the preliminary cancer screening model based on D′ includes:

[0014] The image features I′ and biomarker features B′ in D′ are respectively input into parallel convolutional neural networks and fully connected networks to extract high-dimensional representation vectors and ;

[0015] The and are fused through an attention mechanism to construct a joint representation vector ;

[0016] The is input into a classification subnetwork, supervised training is performed using a cross-entropy loss function, and a screening model is output.

[0017] The performance of the model is evaluated using a validation set.

[0018] Preferably, the multi-task learning method is used to extract image domain feature vectors and biomarker feature vectors , including:

[0019] A shared bottom layer feature extraction network is constructed to preliminarily encode image input and marker input to form a unified intermediate feature space.

[0020] An image task branch and a biomarker task branch are respectively set up above the shared layer, the image task branch uses a convolutional neural network to extract , and the biomarker task branch uses a multilayer perceptron to extract . ​

[0021] Jointly minimize image recognition error and marker anomaly detection error, optimize learning path of two branches.

[0022] Preferably, the online transfer learning mechanism is used to adjust the real-time sample data uploaded to the target region T to obtain a region-adapted model MT, including:

[0023] The model is uploaded to a cloud training platform;

[0024] Real-time sample data uploaded to the target region T is received, and normalized image features and biomarker features are extracted therefrom;

[0025] The model is parameter-adjusted based on a domain adaptation algorithm;

[0026] After the adjustment is completed, the region-adapted model MT is output.

[0027] Preferably, the activation heat map HT is extracted from the model MT, and a weak signal high response region RT is identified by a feature sensitivity analysis method, including:

[0028] The screening image of the target region T is input into the classification subnetwork of the model MT, and the feature map of the last layer of convolution output is recorded;

[0029] The Grad-CAM algorithm is used to visually process the feature map, and the activation heat map HT is generated in combination with the gradient information of the predicted category;

[0030] Based on the heat map HT, pixel-level sensitivity analysis is performed, and the influence degree of each region on the model output is evaluated by using the local perturbation method;

[0031] Image regions with high sensitivity and signal intensity lower than a set level are extracted, and are marked as weak signal high response regions RT.

[0032] Preferably, the RT region is compared with the artificial annotation result, including:

[0033] The obtained weak signal high response region RT is spatially matched with the corresponding artificial annotation mark, the intersection over union of each candidate region is calculated, and the candidate regions are divided into true positive, false positive and missed detection three categories based on the consistency of the intersection over union and the artificial annotation.

[0034] Preferably, the screening weight of the model MT is locally adjusted, and a final model MF is output, including:

[0035] For the candidate regions determined as false positive or missed detection, a local weight adjustment matrix Wloc is constructed;

[0036] The Wloc amplifies or attenuates the last several convolution channels of the model MT by region, specifically: when the intersection over union (IoU) is lower than the set threshold 0.3 and the sensitivity is higher than 0.35, the channel weight of the corresponding position is amplified by a factor of 1.6; when the IoU is higher than 0.6 and the signal strength is high, the channel weight is attenuated by a factor of 0.8.

[0037] The model MT is adjusted using a weighted loss function with region weights, taking Wloc as the prior, the loss function being a linear combination of weighted cross-entropy and focal loss, wherein the loss of the samples in the region is amplified or reduced by Wloc, small batch incremental training is used, the learning rate is 0.0003, a maximum of 30 iterations is used, and the AUC of the validation set is improved by no less than 0.01 as the early stopping criterion.

[0038] After the adjustment is completed and the preset performance threshold is reached on the low signal-to-noise ratio sample subset by the validation set, the final model MF is derived.

[0039] In the above technical solution, the technical effects and advantages provided by the present application are as follows:

[0040] 1. The present application provides a cloud architecture-based early cancer screening model construction method, which can effectively fuse heterogeneous medical data from multiple regions and multiple modalities, and construct a unified joint representation vector through multi-task learning and feature fusion mechanism, realizing comprehensive modeling of image features, biomarker features, device differences and geographical variations. By deploying on a cloud platform and combining online transfer learning and domain adaptation strategy, the model can dynamically adapt to the sample characteristics of different regions, significantly improving the accuracy and universality of early cancer screening under complex regional conditions.

[0041] 2. The present application accurately identifies weak signal high response regions by introducing activation heat map and feature sensitivity analysis, and optimizes the local weight of the model combined with artificial annotation results, effectively improving the recognition ability of the model for small lesions under low signal-to-noise ratio conditions. The final output region adaptation model not only improves the detection rate of early lesions, but also has good interpretability and deployability, especially suitable for resource-limited scenarios such as primary medical care and remote screening. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0043] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] For examples, please refer to Figure 1 As shown in this embodiment, an early cancer screening model and construction method based on cloud architecture includes:

[0046] Medical image data and patient history biomarker data from multiple geographical regions were collected to construct a heterogeneous raw dataset D, including image features, gene expression features and device label information;

[0047] The heterogeneous original dataset D is subjected to regional feature normalization processing to obtain the normalized dataset D′, where D′ contains the device normalization label E and the geographic variation parameter G for each data sample;

[0048] Preliminary cancer screening model based on D′ The training employs a multi-task learning method to extract image domain feature vectors separately. and biomarker feature vectors And establish a joint representation vector ;

[0049] model The model is deployed to a cloud-based training platform, and the real-time sample data uploaded to the target region T is adjusted using an online transfer learning mechanism to obtain the regional adaptation model MT.

[0050] Activation heatmaps (HT) are extracted from the model MT, and weak signal high response regions (RT) are identified through feature sensitivity analysis to indicate image or biomarker regions that may indicate potential early cancer lesions.

[0051] The RT region is compared with the manually labeled results, and the screening weights of the MT model are locally adjusted to output the final MF model. The MF model is used to identify weak feature regions in early cancer under low signal-to-noise ratio conditions.

[0052] In one embodiment of the present invention, in order to build a cloud-based cancer early screening model, it is first necessary to collect and construct a representative heterogeneous original dataset D as the basis for model training and optimization.

[0053] In this step, the collection objects include but are not limited to multiple medical institutions from different geographical areas, which should cover cities, suburbs, remote mountainous areas and other areas with obvious differences in medical service conditions, so as to enhance the universality and adaptability of the model in practical application. In terms of data sources, the collected raw data includes:

[0054] Medical image data: including low-dose spiral CT images, magnetic resonance (MRI) images, ultrasound images and gastroscopy images, etc., which are used for preliminary screening of different types of cancer. The image data should meet certain clarity and resolution standards, and be accompanied by corresponding timestamp and device model information.

[0055] Biomarker detection data: including cancer-related protein expression (such as CEA, AFP, CA125, etc.), gene mutation detection (such as TP53, EGFR, KRAS, etc.) and mRNA expression profile data in historical blood samples, which are used to construct biological feature information required for cross-modal joint modeling.

[0056] Device label information: due to the differences in manufacturer models and inconsistent image standardization levels of image acquisition devices used by different medical institutions, device label information E should be added to each image data, including device type, imaging resolution, manufacturer, imaging protocol and other parameters, which are used for subsequent data standardization and domain adaptation processing.

[0057] The above multi-source data from multiple medical institutions are structured and stored in the raw data set D, wherein each sample data of the data set D is represented as: ; wherein, represents the i-th image feature data; represents the corresponding biomarker feature data; represents the device label information; represents the geographical area characteristics, such as the region code, average medical resource level, population structure, etc.

[0058] In the data collection process, in order to ensure sample representativeness and feature diversity, cases with early positive diagnosis of cancer are preferably selected, and corresponding control samples (i.e. samples without cancer diagnosis records) are collected for subsequent discriminant training in the model.

[0059] In order to solve the problem of significant differences in multi-source heterogeneous medical data in image standards, detection devices, regional population characteristics, etc., regional feature normalization processing is required for the heterogeneous raw data set D to construct a standardized data set D' in a unified feature space.

[0060] ​First, to address the issue of inconsistent image distribution caused by differences in equipment, equipment labels were added to each sample. Standardization processing is performed. This process includes the following sub-steps: Unified equipment model coding: Mapping imaging devices from different manufacturers and models to a unified equipment category code, such as CT_01, MRI_03, etc.; Imaging parameter normalization: Standardizing equipment parameters such as image resolution, noise level, and gamma value to ensure they participate in model training on a uniform scale; Assigning a normalized equipment label to each sample as the processing result. And recorded in a standardized dataset.

[0061] Secondly, considering the significant differences in genetic background, lifestyle, dietary habits, and cancer incidence characteristics among populations in different geographical regions, the geographical location attributes of each sample were extracted and normalized to obtain geographical variation parameters. The geographic variation parameters G include, but are not limited to: region codes (such as country, province, and city codes); regional incidence rate statistics (such as the early detection rate of a certain cancer in the region over the past 5 years, average age of onset, etc.); population structure parameters (such as the estimated rate of a certain gene mutation carrier rate in the population of the region); and medical accessibility indicators (such as the amount of medical resources per capita, the density of imaging equipment, etc.). The dataset after equipment normalization and geographic variation modeling constitutes the normalized dataset D′, whose sample structure can be described as follows: ;in, Represents the normalized image features; This represents biomarker data that has undergone scale alignment or encoding. Standardize labels for equipment; These are geographic variation parameters.

[0062] In this invention, the training of a preliminary cancer screening model M0 based on the normalized dataset D′ includes:

[0063] First, the normalized image features I′ᵢ and the normalized biomarker features are compared. The features are extracted by feeding them into a parallel deep neural network structure to construct a high-dimensional feature vector space.

[0064] Among them, image features The input is fed into a convolutional neural network (CNN), which includes 5 convolutional layers and 2 max-pooling layers, followed by a batch normalization module and the ReLU activation function, and outputs a high-dimensional feature representation vector in the image domain. The dimension is 512.

[0065] Biomarker characteristics The multiple detection indexes such as protein expression, gene mutation site, mRNA expression abundance, etc. are standardized and input into a multi-layer perception (MLP), the MLP network is composed of three fully connected layers, each layer contains 128, 64 and 32 neuron nodes respectively, and finally outputs a feature vector , with a dimension of 128.

[0066] To enhance the expression connection between cross-modal features, an attention mechanism is introduced to weight and fuse and . The attention mechanism includes a double-channel attention module, which calculates the weight distribution of the image channel and the biomarker channel respectively, and performs weighted fusion through dot product operation. The fusion method is: after splicing and , the weight vector a is generated by inputting the attention module, which is used to adjust the importance of each dimension in and , and then perform element-level weighted summation to output the joint representation vector . The finally formed joint representation vector has a dimension of 640, which is used for subsequent classification model input.

[0067] The joint representation vector is input into the classification subnetwork, which adopts a two-layer fully connected layer structure, the first layer contains 256 nodes, and the second layer is a 2-class output node (cancer / non-cancer), and the output is the screening result probability. In the training process, the cross-entropy (Cross-Entropy) loss function is used as the objective function, and the network weight is updated through the back propagation algorithm to optimize the error between the model output and the label. In the training process, the initial learning rate is set to 0.001, the Adam optimizer is used for iterative optimization, the training batch size is 32, and the maximum training number is 100.

[0068] The performance of the preliminarily trained model M0 is evaluated using an independent validation set, and the evaluation indexes include sensitivity (Sensitivity), specificity (Specificity), F1-score and area under the ROC curve (AUC).

[0069] If the verification indexes do not reach the set threshold (for example, AUC is greater than or equal to 0.92, and sensitivity is greater than or equal to 90%), the network parameters are adjusted or the training samples are expanded for retraining.

[0070] In the present application, the multi-task learning method is used to extract and , and the sub-steps are as follows:

[0071] To improve the feature collaborative learning ability between image and biomarker data, a multi-task learning (MTL) structure is adopted, wherein a bottom network part is shared and used for simultaneously performing preliminary encoding on I' and B' to form a unified intermediate feature space.

[0072] The shared network is a double-input neural network composed of a group of shared convolution modules (for the image path) and a shared fully connected encoder (for the marker path), and the output dimension of the shared layer is unified to 256.

[0073] Two task branches are arranged above the shared layer: the image task branch adopts a residual network structure (ResNet18) to perform deep feature extraction on the image, and outputs an image domain feature vector ᵢ; the marker task branch is a two-layer fully connected network used for learning and expressing associated features, and outputs Through independent training of each branch, fine modeling of different modal data is realized.

[0074] A multi-task collaborative loss function L_total is introduced, which is defined as: ; wherein L_img is the cross-entropy loss of the image classification task, L_bio is the mean square error loss of the biomarker anomaly detection task, and are loss balancing coefficients, and the preferred values are 0.6 and 0.4. By minimizing L_total, joint optimization between the image and the biological task is realized.

[0075] After the image domain feature vector and the biomarker feature vector are extracted, to enhance the multi-modal information fusion effect and fully consider the device difference and regional population variability factors, a joint representation vector is further constructed. Specifically, the image feature , the biological feature , the device normalization label E and the geographical variation parameter G of the corresponding sample are spliced, nonlinear combination is realized through a fully connected fusion layer, and a unified joint vector is generated, that is . The fusion layer is composed of two fully connected structures, and a ReLU activation function is used to enhance the feature representation ability.

[0076] In the present application, the model is deployed to a training platform based on a cloud architecture, and an online transfer learning mechanism is used to adjust the parameters of the real-time sample data uploaded by the target region T to obtain a region-adapted model MT, including the following sub-steps:

[0077] The cancer screening model Upload to the cloud training platform through the standardized interface, the platform is a distributed architecture supporting high concurrency and elastic expansion, with the ability of remote model loading, online fine-tuning and cross-region model scheduling.

[0078] The cloud platform uses Kubernetes cluster for resource management, supports GPU accelerated training at the bottom, and provides RESTful API for model scheduling and data interaction. The model upload format is unified as ONNX (Open Neural Network Exchange) format to ensure compatibility between platforms.

[0079] The cloud platform periodically receives real-time screening sample data from the target deployment area T, which includes normalized image data and biomarker data . The data is pre-processed through the same normalization process as the training data to ensure that the device label and geographical parameters are compatible with the model input structure.

[0080] After receiving the data, the preprocessing module is called to decode and format check in batches, and the tensor format required by the model is converted. After preprocessing, the image domain features and biomarker feature vectors are automatically extracted for subsequent transfer training.

[0081] To make the model adapt to the feature distribution difference of area T, a domain adaptation algorithm is used for parameter fine-tuning. The algorithm uses an adversarial transfer learning method to align the feature distribution consistency of the source domain (training data) and the target domain (area T data) by introducing a domain classifier network D. The specific steps are as follows:

[0082] Construct a discriminator to distinguish whether the sample features come from the source domain or the target domain; define a joint loss function L_total, including task loss L_task and domain adversarial loss L_domain; where L_task is the classification loss of the source domain samples, calculated using the cross-entropy function; L_domain is the feature discrimination loss, which uses adversarial training to minimize the accuracy of the discriminator; the total loss function is defined as: where α and β are weight coefficients, and the optimal values are 0.7 and 0.3; the model is updated using the backpropagation algorithm The parameters of the shared feature extraction layer are adjusted to make the feature distribution in the target domain more consistent. The fine-tuning training adopts a mini-batch incremental update method, each batch contains 32 target domain samples, iterates 20 times, and the learning rate is 0.0005. When the model reaches the preset performance threshold (for example, AUC≥0.90, sensitivity≥88%) in the target area T validation set, it is considered that the domain adaptation is completed, and the final regional adaptation model MT is output. The MT retains the structure framework of the original model , only updates the shared feature layer and part of the high-level classification weight, to ensure that it still has good discrimination ability in the new regional environment.

[0083] The activation heat map HT is extracted from the regional adaptation model MT, and the weak signal high response region RT is identified through feature sensitivity analysis, which is used to indicate the image or biomarker region of the early potential lesion of cancer, including:

[0084] The screening image samples from the target area T are input into the classification subnetwork of the regional adaptation model MT, and the feature map output by the last layer of convolutional neural network in the inference process is recorded. The feature map reflects the response activation of the model in different spatial regions, denoted as Fmap.

[0085] To ensure spatial accuracy, the last layer of convolutional structure of the classification subnetwork is set to not have pooling operation, and the spatial information of the original image structure is maintained. The feature map size is determined according to the input image downsampling ratio, and the original Figure 1 / 16 size is preferred.

[0086] Based on the extracted feature map Fmap, the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm is used to generate the activation heat map HT. The specific steps are as follows:

[0087] The prediction result category output by the model is differentiated in reverse, and the gradient of each channel of the feature map corresponding to the category is calculated;

[0088] The global average pooling of each channel gradient is performed to obtain the weight coefficient of each channel;

[0089] After weighting and summing the weight and the corresponding channel feature map, the heat map HT is generated by processing through the ReLU function.

[0090] The heat map HT represents the response intensity of the model to different image regions, and the higher the brightness represents the more attention the model pays to the region. The value range of HT is standardized to the interval [0, 1], which is convenient for subsequent analysis and processing.

[0091] To further analyze the actual influence degree of each region in the heat map on the model decision result, a pixel-level sensitivity analysis operation is performed. The method adopted is the local perturbation method, and the steps are as follows:

[0092] The local region in the heat map HT is processed by sliding with a fixed window (for example, 8x8 pixels); in each sliding window, the pixels in the region are occluded or noise-injected, and the confidence change AC of the model output before and after perturbation is recorded; the confidence change AC is taken as the sensitivity index of the region, and a sensitivity map ST is constructed.

[0093] After obtaining the heat map HT and the sensitivity map ST, a joint discrimination rule is set for extracting the weak signal high response region RT. The extraction logic is as follows:

[0094] First, filter out the region with a heat map response value , where is the response intensity threshold, and the preferred value is 0.6; further filter the image region with a signal intensity in the filtered region, where θ2 is the lower limit threshold of image brightness or density, and the preferred value is 0.2 (representing that the region presents weak or difficult-to-detect features in the original image); at the same time, require that the corresponding sensitivity map has , is the sensitivity lower limit, and the preferred value is 0.3; the region that meets the above conditions is defined as the weak signal high response region, i.e. RT.

[0095] Compare the identified weak signal high response region RT with the manual annotation result, and locally adjust the screening weight of the model MT, output the final model MF, including the following steps:

[0096] Spatially match and classify the candidate regions

[0097] The identified weak signal high response region RT is spatially matched with the manual annotation region A of the corresponding case.

[0098] The spatial matching adopts the Intersection over Union (IoU) calculation method, which is defined as the ratio of the intersection area of the target region and the annotation region to the union area. The specific calculation method is: ; according to the calculation result, the candidate regions are divided into the following three categories:

[0099] If IoU≥0.5, it is judged as a true positive region;

[0100] If IoU<0.3, it is judged as a false positive region;

[0101] If there is no RT region corresponding to the high IoU in the marked region, it is marked as a missed detection region.

[0102] According to the classification result, a set of local weight adjustment matrix Wloc corresponding to the feature map space is constructed, which is used for the region-level weighting adjustment of the convolution channel and attention module of the model MT. The specific adjustment strategy is as follows:

[0103] If a candidate region is a false positive, and its sensitivity value (calculated by the ST graph) is greater than 0.35, and the IoU is less than 0.3, then the channel response weight of the region is multiplied by a coefficient of 1.6;

[0104] If the candidate region is a true positive but the signal strength (such as image brightness or density) is higher than 0.8, and the IoU is greater than 0.6, then its response weight is multiplied by a coefficient of 0.8 to avoid the model focusing too much on the strong signal region; the default weight of the remaining regions is 1.0, and no adjustment is made. The adjusted Wloc is consistent with the feature map dimension of the model, and is used as prior input to participate in the subsequent model training loss function calculation.

[0105] Based on Wloc, a weighted loss function is constructed to fine-tune the model MT and optimize its recognition ability for weak feature regions. The loss function used is a linear combination of weighted cross-entropy (WCE) and focal loss (Focal Loss), defined as follows: ; Where, α and β are weighting coefficients, and the optimal values are 0.6 and 0.4 respectively; the loss term of each pixel or region in WCE is amplified or reduced according to the weight of the corresponding position in Wloc. The model fine-tuning parameters are as follows: the learning rate is set to 0.0003; training uses mini-batch incremental training, with 32 samples per batch; the maximum number of iterations is set to 30 rounds; the early stopping mechanism is introduced, and the training is automatically terminated when the AUC on the validation set improves by less than 0.01 to prevent overfitting.

[0106] When the model fine-tuning is completed, independent verification is performed on the low signal-to-noise ratio sample subset. If the following performance indicators are met, it is considered that the optimization is effective: sensitivity ≥ 88%; AUC ≥ 0.90.

[0107] After meeting the above conditions, the current model is exported as the final cancer early screening model MF. Compared with the original model MT, the model MF embeds local optimization weights guided by Wloc in the attention mechanism and convolution channel response, enhancing the recognition ability of weak signal lesions and abnormal marker expression regions in the image, especially suitable for low signal-to-noise ratio screening scenarios. The MF can be deployed on a primary screening terminal to effectively identify early-stage weak signal lesions of cancer under edge image quality.

[0108] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A cloud-based early cancer screening model and its construction method, characterized in that: include: Medical image data and patient history biomarker data from multiple geographical regions were collected to construct a heterogeneous raw dataset D, including image features, gene expression features and device label information; The heterogeneous original dataset D is subjected to regional feature normalization processing to obtain the normalized dataset D′, where D′ contains the device normalization label E and the geographic variation parameter G for each data sample; Preliminary cancer screening model based on D′ The training employs a multi-task learning method to extract image domain feature vectors separately. and biomarker feature vectors And establish a joint representation vector ; model The model is deployed to a cloud-based training platform, and the real-time sample data uploaded to the target region T is adjusted using an online transfer learning mechanism to obtain the regional adaptation model MT. Activation heatmaps (HT) are extracted from the model MT, and weak signal high response regions (RT) are identified through feature sensitivity analysis to indicate image or biomarker regions that may indicate potential early cancer lesions. The RT region is compared with the manually labeled results, and the screening weights of the MT model are locally adjusted to output the final MF model. The MF model is used to identify weak feature regions in early cancer under low signal-to-noise ratio conditions.

2. The cloud-based early cancer screening model and construction method according to claim 1, characterized in that: The preliminary cancer screening model based on D′ The training includes: Image features I′ and biomarker features B′ in D′ are input into a parallel convolutional neural network and a fully connected network, respectively, to extract high-dimensional representation vectors. and ; Will and By fusing through an attention mechanism, a joint representation vector is constructed. ; Will The input is fed into a classification subnetwork, and supervised training is performed using the cross-entropy loss function to output a screening model. ; Use the validation set to test the model Conduct a performance evaluation.

3. The cloud-based early cancer screening model and construction method according to claim 2, characterized in that: The multi-task learning method is used to extract image domain feature vectors respectively. and biomarker feature vectors ,include: Construct a shared underlying feature extraction network for image input. and marker input Preliminary encoding is performed to form a unified intermediate feature space; Above the shared layer, separate branches are set up for image tasks and biomarkers. The image task branch uses a convolutional neural network to extract biomarkers. The biological task branch uses a multilayer perceptron to extract... ; We optimize the learning paths of both branches by jointly minimizing image recognition error and marker anomaly detection error.

4. The cloud-based early cancer screening model and construction method according to claim 1, characterized in that: The process of adjusting the real-time sample data uploaded to the target region T using an online transfer learning mechanism to obtain the regional adaptation model MT includes: model Uploaded to the cloud training platform; Receive real-time sample data uploaded by the target region T, and extract normalized image features and biomarker features from it; Based on domain adaptive algorithm for model Adjust the parameters; After adjustment, output the region adaptation model MT.

5. The cloud-based early cancer screening model and construction method according to claim 4, characterized in that: The process of extracting the activation heatmap HT from the model MT and identifying the weak signal high response region RT through feature sensitivity analysis includes: Input the screening image of the target region T into the classification subnetwork of model MT and record the feature map output of the last convolutional layer; The Grad-CAM algorithm is used to visualize the feature map, and the activation heatmap HT is generated by combining the gradient information of the predicted category. Pixel-level sensitivity analysis is performed based on heatmap HT, and the local perturbation method is used to evaluate the degree of influence of each region on the model output. Image regions with high sensitivity and signal strength below a set level are extracted and marked as weak signal high response regions RT.

6. The cloud-based early cancer screening model and construction method according to claim 5, characterized in that: The RT region is compared with the manually labeled results, including: The obtained weak signal high response region RT is spatially matched with the corresponding manually labeled markers. The cross-union ratio (CUI) of each candidate region is calculated, and the candidate regions are divided into three categories—true positives, false positives, and missed detections—based on the consistency between the CUI and the manually labeled markers.

7. The cloud-based early cancer screening model and construction method according to claim 6, characterized in that: The screening weights of the model MT are locally adjusted to output the final model MF, including: For candidate regions that are identified as false positives or missed detections, a local weight adjustment matrix Wloc is constructed; The Wloc method amplifies or attenuates the last few convolutional channels and attention weights of the model MT by region. Specifically, when the Intersection over Union (IoU) is lower than a set threshold of 0.3 and the sensitivity is higher than 0.35, the channel weights at the corresponding positions are amplified by a coefficient of 1.6; when the IoU is higher than 0.6 and the signal strength is high, the weights are attenuated by a coefficient of 0.

8. Using Wloc as a prior, the model MT is adjusted using a weighted loss function with region weights. The loss function is a linear combination of weighted cross-entropy and focus loss. The loss of samples within a region is amplified or reduced by Wloc. Mini-batch incremental training is used with a learning rate of 0.0003, a maximum of 30 iterations, and an early stopping criterion of AUC improvement of no less than 0.01 on the validation set. After the adjustments are completed and the preset performance threshold is achieved on the low signal-to-noise ratio sample subset through the validation set, the final model MF is exported.

Images