Intelligent gastrointestinal polyp auxiliary examination system and method

Abstract

The application discloses an intelligent gastrointestinal polyp auxiliary examination system and method, adopts a "hardware architecture-software module" collaborative design, the hardware architecture provides computing power support and equipment interface, the software module realizes algorithm function and man-machine interaction, and the gastrointestinal polyp auxiliary examination is completed in linkage; the method is realized based on the system, including data preprocessing, model training, online / offline reasoning and result output whole process, aiming at the problems of complex endoscope image background, class imbalance and device difference, through attention guide UNet++, progressive enhancement, hybrid loss and TTA / post-processing process design, which is helpful to improve the segmentation robustness and usability.

Description

Technical Field

[0001] This invention belongs to the technical field of cross-application of medical image processing equipment and deep learning technology, and relates to an intelligent auxiliary examination system and method for gastrointestinal polyps. Background Technology

[0002] Gastrointestinal cancers (including gastric cancer, colorectal cancer, etc.) are among the most common types of malignant tumors. Gastrointestinal polyps, as an important form of precancerous lesions, are crucial for reducing the risk of progression through early detection and standardized treatment. Gastroscopy and colonoscopy are currently the most commonly used methods for polyp detection and treatment.

[0003] However, the traditional method of polyp localization and segmentation, which relies mainly on manual observation and labeling, is highly dependent on the operator's experience and still has the following limitations: (1) Small polyps and flat polyps are easily overlooked in complex backgrounds, leading to an increased risk of missed detection; (2) The examination images are affected by factors such as mucosal folds, residues / foam, reflections, and motion blur, making the boundary determination highly subjective; (3) Long-term examination and image reading can easily cause fatigue, affecting stability and efficiency.

[0004] To address these issues, deep learning techniques (especially the UNet series of networks) have been attempted for gastrointestinal polyp segmentation, but existing technologies still face four key bottlenecks:

[0005] ① Insufficient background suppression: Traditional UNet / UNet++ fuses features through skip connections, but it is not good at distinguishing the importance of background and polyp-related features. It is easily affected by mucosal wrinkles, debris / foam, reflection, etc., which leads to segmentation boundary shift.

[0006] ② Data augmentation is out of touch with the scene: Most studies use static augmentation (such as fixed flipping and rotation), which is difficult to cover dynamic variations such as lighting changes, motion blur, and foam / residue artifacts during the endoscopy acquisition process, thus limiting the generalization of the model;

[0007] ③ Class imbalance is not effectively addressed: Gastrointestinal polyp areas are usually much smaller than background areas, resulting in significant class imbalance. A single loss function can easily lead the model to favor background prediction, and small polyp pixels are easily ignored.

[0008] ④ Poor device compatibility: Some high-performance models have a large number of parameters and high inference time, which require high computing power and storage resources, making it difficult to run stably in resource-constrained scenarios.

[0009] Therefore, developing an intelligent auxiliary examination method and system for gastrointestinal polyps that features strong background suppression, good detection capability for small polyps, and strong equipment compatibility is significant for improving the consistency of polyp observation and the efficiency of auxiliary analysis. Summary of the Invention

[0010] The purpose of this invention is to provide an intelligent auxiliary examination system and method for gastrointestinal polyps. It adopts a collaborative design of "hardware architecture and software modules." The hardware architecture provides computing power and device interfaces, while the software modules implement algorithm functions and human-computer interaction. The two work together to complete the auxiliary examination of gastrointestinal polyps. The method is based on system implementation and includes the entire process of data preprocessing, model training, online / offline inference, and result output. Addressing the aforementioned problems in the prior art, the invention provides the following improvements and beneficial effects:

[0011] It helps enhance the ability to suppress background noise and helps distinguish gastrointestinal polyps from mucosal folds / residue / gastric foam;

[0012] It helps improve the recall rate of small gastrointestinal polyps and reduces the risk of missed detection;

[0013] It helps to adapt to the dynamic variations of different types of gastrointestinal endoscopy equipment and improves generalization.

[0014] This facilitates hardware compatibility, supports operation on general-purpose computing platforms, and improves inference efficiency to meet the auxiliary analysis needs of different application scenarios. The implementation method of this application primarily focuses on software algorithms and training / inference verification on general-purpose computing platforms, and does not limit the specific integrated hardware and software product form; subsequently, the software modules can be integrated and deployed with endoscopy workstations, edge computing terminals, etc., according to actual scenarios.

[0015] This invention is achieved through the following technical solution:

[0016] An intelligent auxiliary examination system for gastrointestinal polyps includes a hardware architecture and software modules, wherein the hardware architecture includes:

[0017] Main control unit: responsible for the overall scheduling of hardware architecture, device communication and data flow, and coordinating the working sequence of each hardware unit;

[0018] AI computing unit: performs model inference and training computations, runs attention-guided UNet++ models, and supports inference based on preset parameter configurations.

[0019] Image preprocessing unit: performs image resizing, standardization, and data augmentation, and optimizes preprocessing for common noise and artifacts in endoscopic images to provide input data for subsequent model calculations;

[0020] Device interface unit: used to connect to gastrointestinal endoscopy equipment and transmit the image / video data captured by the endoscope to the computing platform; optional data interaction with the image storage system.

[0021] Storage unit: Used to store model parameters, training dataset and segmentation results, and can be implemented using hard disk, solid-state drive or other storage media.

[0022] Human-computer interaction unit: includes display and interaction interfaces for parameter setting and result viewing;

[0023] The software module includes:

[0024] Data preprocessing module: Working in conjunction with the storage unit, it performs preprocessing operations such as data reading and partitioning (K-fold cross-validation), image size unification (default 384×384), and progressive data augmentation.

[0025] Model building module: Collaborates with the AI ​​computing unit to build attention-guided UNet++ networks and optimize network parameters;

[0026] Training optimization module: Collaborates with the main control unit and AI computing unit to realize hybrid loss function calculation, optimizer configuration and training monitoring;

[0027] Inference and analysis module: Collaborates with device interface unit and AI computing unit to perform test-time enhancement, model integration and morphological post-processing;

[0028] Results visualization module: In collaboration with the human-computer interaction unit and storage unit, it displays inspection results and supports data export;

[0029] Performance evaluation module: Works with the main control unit to calculate segmentation metrics and generate evaluation reports;

[0030] Edge deployment adaptation module: works in conjunction with the AI ​​computing unit to configure inference parameters and adapt the runtime environment on different computing platforms; model format conversion and other optional functions are available to enable application expansion under different deployment conditions.

[0031] To better implement this solution, the three-stage progressive data augmentation performed by the data preprocessing module is further divided into weak augmentation, medium augmentation and strong augmentation stages according to the training cycle and endoscope type. The augmentation process is performed by the image preprocessing unit on the processor, which performs the corresponding preprocessing and augmentation operations.

[0032] To better implement this solution, the attention-guided UNet++ network constructed by the model building module further includes an encoder, attention gate, decoder, and output layer. The encoder has a depth of 5 and the number of channels is 32, 64, 128, 256, and 512 respectively. Attention and dropout parameters can be configured to suppress irrelevant backgrounds and improve robustness.

[0033] To better implement this solution, the training optimization module further configures a hybrid loss function where DiceLoss is used to measure the overlap between the predicted mask and the labeled mask for the segmented region, BCELoss uses adaptive positive sample weights and performs range cropping so that the positive sample weights range from 1.0 to 7.0, and FocalLoss parameter γ=2.0 to improve the learning of hard-to-separate pixels; hybrid precision can be optionally enabled during training to improve training efficiency.

[0034] To better implement this solution, the inference process executed by the inference analysis module further includes loading the model according to the preset configuration, test-time enhancement, 5-fold cross-validation model integration, and morphological post-processing; the inference time is related to factors such as computing platform configuration and input resolution.

[0035] An intelligent auxiliary examination method for gastrointestinal polyps, based on the aforementioned intelligent auxiliary examination system for gastrointestinal polyps, includes the following steps:

[0036] Data preprocessing: Gastrointestinal endoscopy images are preprocessed, including reading data in pairs from the image catalog and the labeled mask catalog; splitting the dataset based on K-fold cross-validation; uniformly scaling the images and masks to a preset size (default 384×384); and employing a weak-medium-strong progressive data augmentation strategy according to the training stage. Pixel values ​​can be normalized / standardized as needed to improve training stability.

[0037] Model training: Build an attention-guided UNet++ model, configure the hybrid loss function, RAdam optimizer and learning rate scheduling strategy, train the model and save the 5-fold model parameters;

[0038] Online / offline inference: Receives gastrointestinal endoscopy images, calls the corresponding model parameters according to the preset endoscopy type / input source configuration, performs test-time enhancement, 5-fold model ensemble and morphological post-processing, and generates polyp segmentation results;

[0039] Output results: Outputs the polyp segmentation mask and its contour, area, and other derived information calculated from the mask (optional), supporting display and data export.

[0040] To better implement this solution, the three-stage progressive data augmentation step is further divided according to the training cycle and endoscope type in the data preprocessing step:

[0041] Weak augmentation phase: In the early stage of training, relatively mild data augmentation operations are used, such as random flipping of the input image and slight changes in brightness / contrast, to stabilize the initial convergence of the model;

[0042] Mid-stage augmentation: In the middle of training, moderate-intensity data augmentation operations are introduced, including random translation, scaling, and mild blurring, to enhance the model's adaptability to changes in viewpoint and imaging.

[0043] Strong augmentation phase: Strong data augmentation operations are introduced in the later stage of training, including geometric distortion, CLAHE contrast enhancement, noise perturbation, etc., to improve the robustness of the model under complex imaging conditions.

[0044] To better implement this scheme, the attention-guided UNet++ network constructed during model training further includes an encoder, an attention gate, a decoder, and an output layer; the encoder includes 5 convolutional blocks, the attention gate is used to fuse encoded features and decoded guidance signals to generate attention weights to suppress irrelevant background; the decoder includes 4 convolutional blocks, and the output layer is used to output a segmentation probability map.

[0045] The hybrid loss function is Dice-BCE-Focal loss with a weight ratio of 0.5:0.3:0.2. DiceLoss is optimized for sensitivity to small polyps, BCELoss uses adaptive positive sample weights, and FocalLoss increases the loss weight of edge pixels.

[0046] Training uses the RAdam optimizer with an initial learning rate of 3e-4 and a weight decay of 1e-5. The learning rate scheduling uses a cosine annealing hot restart strategy (T0=10, T_mult=2, eta_min=1e-6). The early stopping strategy is to terminate training if the validation set metric does not improve for 20 consecutive cycles.

[0047] To better implement this solution, the online / offline inference steps further include the following enhancements during testing: original, horizontal flip, vertical flip, and horizontal + vertical flip. The results of each transformation are then fused to obtain the segmented output.

[0048] Morphological post-processing performs 7×7 kernel closing and 3×3 kernel opening operations on the predicted mask, and retains connected regions with an area of ​​not less than 400 pixels (which can be adjusted according to the image resolution).

[0049] Inference time is related to factors such as computing platform configuration and input resolution.

[0050] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0051] By introducing attention-guided UNet++ networks and combining channel attention and spatial attention mechanisms, it helps to distinguish gastrointestinal polyps from background interference features such as mucosal folds, intestinal residue, and gastric foam, strengthens attention to key polyp regions, and suppresses irrelevant background information, thereby improving the accuracy of segmentation boundaries and reducing missegmentation and boundary offset phenomena.

[0052] To address the issue of small polyps having a low pixel ratio and being easily overlooked, a hybrid loss function (Dice-BCE-Focal) and an adaptive positive sample weighting strategy are used to enhance the model's sensitivity to small targets. Combining multi-stage enhancement and morphological post-processing helps improve the recall rate of small polyps and reduce the risk of missed detection.

[0053] A three-stage progressive data augmentation strategy is adopted to simulate dynamic variations in actual data acquisition (such as changes in illumination, motion blur, and foam artifacts), so that the model can maintain good segmentation performance under different devices and shooting conditions, and improve its adaptability and robustness to diverse inspection scenarios.

[0054] Through modular hardware design and software optimization, the system can run on general-purpose computing platforms and supports both offline and online input methods. Inference time is related to factors such as computing platform configuration and input resolution. The system output is used to assist observation and recording and does not replace physician judgment. Attached Figure Description

[0055] Figure 1 Hardware architecture flowchart for an intelligent gastrointestinal polyp auxiliary examination system;

[0056] Figure 2 Workflow diagram of the software modules for an intelligent gastrointestinal polyp auxiliary examination system;

[0057] Figure 3 Hardware-software collaborative workflow diagram for an intelligent gastrointestinal polyp auxiliary examination system;

[0058] Figure 4 A roadmap for colorectal polyp segmentation techniques in intelligent auxiliary examination methods for gastrointestinal polyps;

[0059] Figure 5 This is the fold-1 training curve;

[0060] Figure 6 This is a fold-1 sample;

[0061] Figure 7 The training curve is for fold-2.

[0062] Figure 8 This is a fold-2 sample;

[0063] Figure 9 For fold-3 training curves;

[0064] Figure 10 This is a fold-3 sample;

[0065] Figure 11 For fold-4 training curves;

[0066] Figure 12 For fold-4 samples;

[0067] Figure 13 For fold-5 training curves;

[0068] Figure 14 This is a fold-5 sample; Detailed Implementation

[0069] The following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0070] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless otherwise expressly indicated by the invention, the singular form is intended to include the plural form as well. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0071] For ease of description, the words "up," "down," "left," and "right" appearing in this invention only indicate that they are consistent with the up, down, left, and right directions of the accompanying drawings themselves, and do not limit the structure. They are merely for the purpose of facilitating the description of this invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0072] Terminology Explanation: The terms "installation," "connection," "linking," and "fixing" in this invention should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction relationship between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0073] Example 1:

[0074] This embodiment describes a preferred implementation of the present solution, an intelligent gastrointestinal polyp auxiliary examination system, including hardware architecture and software modules, such as... Figure 1 As shown, the hardware architecture includes:

[0075] Main control unit: responsible for the overall scheduling of hardware architecture, device communication and data flow, and coordinating the working sequence of each hardware unit; the main control unit can be implemented by a general-purpose CPU / SoC.

[0076] AI computing unit: used to perform model training and inference, run attention-guided UNet++ models, support inference according to preset parameter configurations, can be implemented by CPU, GPU or other general computing platforms, and support adjusting inference and training configurations according to computing power conditions.

[0077] Image preprocessing unit: Performs image resizing, standardization, and data augmentation. As a preprocessing step for model training and inference, it preprocesses and optimizes noise and artifacts in endoscopic images to provide input data for subsequent model calculations. The specific implementation can be completed by general computing resources such as CPU / GPU.

[0078] Device interface unit: used to connect to gastrointestinal endoscopy equipment and transmit image / video data to the computing platform; optional data interaction with image storage system.

[0079] Storage unit: Used to store model parameters, training datasets and segmentation results, and can be implemented using hard disks, solid-state drives or other storage media.

[0080] Human-computer interaction unit: includes display and interaction interface, used for parameter setting and result viewing;

[0081] Figure 1 The display shows the connection relationship and data flow of the core hardware components of the device. The main control unit, as the core scheduling node, coordinates the collaborative work of the AI ​​computing unit (responsible for inference training), the image preprocessing unit (responsible for image optimization), the device interface unit (responsible for external device interaction), the storage unit (responsible for data storage), and the human-computer interaction unit (responsible for operator operation and result display). The component logic and data links of the hardware architecture diagram of this solution are clearly shown.

[0082] like Figure 2 As shown, the software module includes:

[0083] Data preprocessing module: Working in conjunction with the storage unit, it performs preprocessing operations such as data reading and partitioning (K-fold cross-validation), image size unification (default 384×384), and progressive data augmentation.

[0084] Model building module: Collaborates with the AI ​​computing unit to build attention-guided UNet++ networks (including encoders, attention gates, decoders, and output layers), optimize network parameters, and support network structure visualization and parameter configuration.

[0085] Training optimization module: In collaboration with the main control unit and AI computing unit, it realizes the Dice-BCE-Focal hybrid loss calculation (weights 0.5:0.3:0.2), configures the RAdam optimizer and cosine annealing with warm restart learning rate scheduling (CosineAnnealingWarmRestarts, T0=10, T_mult=2, eta_min=1e-6), and sets an early stopping strategy (patience=20) to prevent overfitting.

[0086] Inference and analysis module: Collaborates with device interface unit and AI computing unit to perform test-time enhancement, model integration and morphological post-processing;

[0087] Results visualization module: In collaboration with the human-computer interaction unit and storage unit, it displays the original image and segmentation results, and supports saving and exporting the results.

[0088] Performance evaluation module: In collaboration with the main control unit, it calculates the segmentation metrics such as Dice, IoU, precision, recall, specificity, and accuracy, and generates an evaluation report.

[0089] Edge deployment adaptation module: works in conjunction with the AI ​​computing unit to configure inference parameters and adapt the runtime environment on different computing platforms; model format conversion and other optional functions are available to enable application expansion under different deployment conditions.

[0090] Inference time is related to factors such as computing platform configuration and input resolution.

[0091] Figure 2 Starting with the data preprocessing module, through model building, training optimization, and inference analysis, the results are finally output through the result visualization and performance evaluation module, clarifying the functional boundaries of each module and the data transfer relationship.

[0092] like Figure 3 The diagram shows the hardware-software collaborative workflow of the intelligent gastrointestinal polyp auxiliary examination system. The diagram clearly shows the technical solution of "hardware-software collaboration". It clearly shows the linkage between hardware units (device interface, image preprocessing, AI calculation, etc.) and corresponding software modules (device driver, data preprocessing, model-related modules, etc.) after the external endoscopic equipment inputs images. The main control unit realizes the synchronization of hardware resources and software functions through the software scheduling module, and fully presents the entire collaborative link of the device from data reception, processing to result output.

[0093] This approach introduces an attention-guided UNet++ network, combined with an attention mechanism, to help distinguish gastrointestinal polyps from background interference features such as mucosal folds, intestinal residue, and gastric foam. It strengthens focus on key polyp regions and suppresses irrelevant background information, thereby improving the consistency and accuracy of segmentation boundaries and reducing missegmentation and boundary shifts. Addressing the issue of small polyps having a low pixel ratio and being easily overlooked, the optimization of the hybrid loss function (Dice-BCE-Focal) and adaptive sample weighting strategy enhances the model's sensitivity to small targets. Furthermore, combining multi-stage enhancement and morphological post-processing improves the recall rate of small polyps, further reducing the risk of missed detections.

[0094] Example 2:

[0095] This embodiment of the intelligent gastrointestinal polyp auxiliary examination system is an improvement on the first embodiment. The three-stage progressive data enhancement performed by the data preprocessing module is divided into weak enhancement, medium enhancement and strong enhancement stages according to the training cycle and endoscopy type. The enhancement process is performed by the image preprocessing unit on the processor to perform the corresponding preprocessing and enhancement operations.

[0096] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0097] Example 3:

[0098] This embodiment of the intelligent gastrointestinal polyp auxiliary examination method is an improvement on the method in embodiment 1. In the attention-guided UNet++ network, the segmentation and prediction of polyp regions are achieved through a multi-layer encoder-decoder structure and attention gate mechanism. The model training process can be executed in computing environments such as CPU, GPU or DirectML, and the running process information such as training rounds, loss function and evaluation index is recorded.

[0099] ① Encoder: 5 convolutional blocks, each containing 2 3×3 convolutional layers (stride 1, padding 1), BN layer, and ReLU; number of channels 32-64-128-256-512, followed by 2×2 max pooling (stride 2).

[0100] ②Attention gate: Located in the skip connection path, it is used to fuse encoded features and decoded guiding signals to generate attention weights to suppress irrelevant background and highlight polyp-related regions.

[0101] ③ Decoder: Used for progressive upsampling and fusing multi-scale features to restore spatial resolution and output segmentation probability map; upsampling can be achieved by transposed convolution or interpolation.

[0102] ④ Output layer: 1×1 convolution reduced to 1 channel, sigmoid output probability map, and gastrointestinal polyps are also output;

[0103] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0104] Example 4:

[0105] This embodiment of the intelligent gastrointestinal polyp auxiliary examination system is an improvement on embodiment 1. The hybrid loss function configured in the training optimization module includes:

[0106] ①DiceLoss: 1-[2×|X∩Y| / (|X|+|Y|)], where X is the label mask and Y is the prediction mask, representing the DiceLoss of gastric polyp samples;

[0107] ②BCELoss: Adaptive positive sample weight ω = (total number of pixels - number of polyp pixels) / number of polyp pixels, and crop ω to a range of 1.0-7.0, which is dynamically updated in batches during the training process;

[0108] ③FocalLoss: -(1-p)^γ×y×log(p)-p^γ×(1-y)×log(1-p), where γ=2.0, used to increase attention to hard-to-distinguish pixels;

[0109] The training may optionally enable mixed precision to improve training efficiency; whether or not to enable it depends on the type of computing device.

[0110] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0111] Example 5:

[0112] This embodiment of the intelligent gastrointestinal polyp auxiliary examination system is an improvement on embodiment 1. The reasoning process executed by the reasoning analysis module includes:

[0113] ① Input configuration: Select the corresponding model parameters and processing flow according to the endoscopic image input source or the preset device type configuration;

[0114] ②TTA: Includes four transformations: original, horizontal flip, vertical flip, and horizontal + vertical flip, and the outputs of each transformation are fused.

[0115] ③ 5-fold ensemble: The ensemble result is obtained by averaging the probability plots output by the 5-fold model;

[0116] ④ Post-processing: Perform morphological closing and opening operations on the ensemble probability graph, and filter out small noise regions by combining connected region area thresholds; the morphological kernel size and area threshold can be adjusted in the configuration according to the input resolution (e.g., kernel=7, min_area=400).

[0117] ①Inference time is related to factors such as computing platform configuration and input resolution.

[0118] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0119] Example 6:

[0120] This embodiment describes a preferred implementation of the solution, such as... Figure 3 As shown, an intelligent auxiliary examination method for gastrointestinal polyps based on any of the schemes in Examples 1-5 includes the following steps:

[0121] Data preprocessing: Gastrointestinal endoscopy images are preprocessed, including reading data in pairs from the image catalog and the labeled mask catalog; splitting the dataset based on K-fold cross-validation; uniformly scaling the images and masks to a preset size (default 384×384); and employing a weak-medium-strong progressive data augmentation strategy according to the training stage. Pixel values ​​can be normalized / standardized as needed to improve training stability.

[0122] Model training: Build an attention-guided UNet++ model, configure the hybrid loss function, RAdam optimizer and learning rate scheduling strategy, train the model and save the 5-fold model parameters;

[0123] Online / offline inference: Receives gastrointestinal endoscopy images, calls the corresponding model parameters according to the preset endoscopy type / input source configuration, performs test-time enhancement, 5-fold model ensemble and morphological post-processing, and generates polyp segmentation results;

[0124] Output results: Outputs the polyp segmentation mask and its contour, area, and other derived information calculated from the mask (optional), supporting display and data export.

[0125] Example 7:

[0126] This embodiment of the intelligent auxiliary examination method for gastrointestinal polyps is an improvement on embodiment 6. In the data preprocessing step, the enhancement operation is gradually adjusted in stages according to the training process, and the specific enhancement type and intensity can be set in the configuration parameters.

[0127] Weak enhancement stage: random flipping, slight rotation, and brightness / contrast changes are used; Medium enhancement stage: random translation, scaling, and slight blurring are added; Strong enhancement stage: geometric distortion, CLAHE contrast enhancement, and noise perturbation are added to improve model robustness.

[0128] Medium-intensity augmentation stage: Based on the weak-intensity augmentation stage, medium-intensity data augmentation operations are introduced, such as random translation, scaling and mild blurring, to enhance the model's adaptability to changes in viewpoint and imaging; the specific augmentation type and parameters are set by the configuration file and can be adjusted according to the data characteristics.

[0129] Strong augmentation stage: Based on the medium augmentation stage, stronger data augmentation operations are introduced, such as geometric distortion, contrast-limited adaptive histogram equalization (CLAHE), and noise perturbation, to improve the robustness of the model under complex imaging conditions; the specific augmentation type and parameters are set by the configuration file and can be adjusted according to the data characteristics.

[0130] The other parts of this embodiment are the same as those in embodiment 6, so they will not be described again.

[0131] Example 8:

[0132] This embodiment of the intelligent gastrointestinal polyp auxiliary examination method is an improvement on embodiment 6. The attention-guided UNet++ network constructed in the model training includes an encoder, an attention gate, a decoder and an output layer. The encoder includes 5 convolutional blocks. The attention gate is used to fuse the encoded features and the decoded guidance signal and generate attention weights to suppress irrelevant background. The decoder includes 4 convolutional blocks. The output layer is used to output a segmentation probability map.

[0133] The hybrid loss function is Dice-BCE-Focal loss with a weight ratio of 0.5:0.3:0.2. DiceLoss is optimized for sensitivity to small polyps, BCELoss uses adaptive positive sample weights, and FocalLoss increases the loss weight of edge pixels.

[0134] Training uses the RAdam optimizer with an initial learning rate of 3e-4 and a weight decay of 1e-5. The learning rate scheduling uses a cosine annealing hot restart strategy (T0=10, T_mult=2, eta_min=1e-6). The early stopping strategy is to terminate training if the validation set metric does not improve for 20 consecutive cycles.

[0135] The other parts of this embodiment are the same as those in embodiment 6, so they will not be described again.

[0136] Example 9:

[0137] This embodiment of the intelligent auxiliary examination method for gastrointestinal polyps is an improvement on embodiment 6. In the online / offline inference step, the test enhancement includes four transformations: original, horizontal flip, vertical flip, and horizontal + vertical flip. The results of each transformation are fused to obtain the segmented output.

[0138] Morphological post-processing performs 7×7 kernel closing and 3×3 kernel opening operations on the predicted mask, and retains connected regions with an area of ​​not less than 400 pixels (which can be adjusted according to the image resolution).

[0139] Inference time is related to factors such as computing platform configuration and input resolution.

[0140] The other parts of this embodiment are the same as those in embodiment 6, so they will not be described again.

[0141] Example 10:

[0142] This embodiment is configured and deployed based on the intelligent gastrointestinal polyp auxiliary examination system described in any one of the embodiments 1-5, as detailed below.

[0143] Hardware component configuration:

[0144] Main control unit: Used to complete system scheduling and data flow. The main control unit can be implemented by a general-purpose CPU / SoC.

[0145] AI computing unit: Optional configuration with acceleration hardware such as GPU to improve training and inference efficiency.

[0146] Image preprocessing unit: can be implemented by CPU, GPU or other general computing resources; programmable logic devices or other optional acceleration resources can also be selected according to the application scenario.

[0147] Storage unit: can be implemented using solid-state drives, hard disk drives or other storage media, with capacity and read / write performance selected according to the application scenario.

[0148] Device interface: Used to receive endoscopic image / video data and transmit it to the computing platform;

[0149] Human-computer interaction unit: includes display and interaction interface, used for parameter setting and result viewing.

[0150] Deployment process:

[0151] Hardware assembly: The components can be connected according to the system architecture to complete the connection configuration for data input, calculation and result display;

[0152] System initialization: The device can be started to complete the necessary configuration and self-test (optional);

[0153] Inference time is related to factors such as computing platform configuration and input resolution.

[0154] Performance verification: The trained model can be loaded to run and verify the inference process; the inference time is related to factors such as computing platform configuration and input resolution.

[0155] The system can run in a general computing environment that meets relevant electrical safety requirements, and the specific hardware configuration can be selected according to the application scenario.

[0156] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. An intelligent auxiliary examination system for gastrointestinal polyps, comprising a hardware architecture and software modules, characterized in that, The hardware architecture includes: Main control unit: responsible for the overall scheduling of hardware architecture, device communication and data flow, and coordinating the working sequence of each hardware unit; AI computing unit: performs model inference and training computation, runs attention-guided UNet++ models, and supports inference based on preset parameter configurations; Image preprocessing unit: performs image resizing, standardization, and data augmentation, and preprocesses and optimizes noise and artifacts in endoscopic images to provide input data for subsequent model calculations; Equipment interface unit: connects to gastrointestinal endoscopy equipment, transmits data, and can optionally interface with image storage and transmission systems; Storage unit: Stores model parameters, dataset, and segmentation results, with model parameters stored separately; Human-computer interaction unit: includes display and interaction interfaces for parameter setting and result viewing; The software module includes: Data preprocessing module: works in conjunction with the storage unit to complete data reading and partitioning, image size unification, and progressive data augmentation; Model building module: Collaborates with the AI ​​computing unit to build attention-guided UNet++ networks and optimize network parameters; Training optimization module: Collaborates with the main control unit and AI computing unit to realize hybrid loss function calculation, optimizer configuration and training monitoring; Inference and analysis module: Collaborates with device interface unit and AI computing unit to perform test-time enhancement, model integration and morphological post-processing; Results visualization module: In collaboration with the human-computer interaction unit and storage unit, it displays inspection results and supports data export; Performance evaluation module: Works with the main control unit to calculate segmentation metrics and generate evaluation reports; Edge deployment adaptation module: works in collaboration with the AI ​​computing unit to configure inference parameters and adapt the runtime environment on different computing platforms.

2. The intelligent gastrointestinal polyp auxiliary examination system according to claim 1, characterized in that, The progressive data augmentation performed by the data preprocessing module is divided into weak augmentation, medium augmentation and strong augmentation stages according to the training rounds. The augmentation process is performed by the image preprocessing unit on the processor, which performs the corresponding preprocessing and augmentation operations.

3. The intelligent gastrointestinal polyp auxiliary examination system according to claim 1, characterized in that, The attention-guided UNet++ network constructed by the model building module includes an encoder, attention gate, decoder, and output layer. The encoder has a depth of 5 and the number of channels is 32, 64, 128, 256, and 512 respectively. The network can be configured with attention and dropout parameters.

4. The intelligent gastrointestinal polyp auxiliary examination system according to claim 1, characterized in that, In the hybrid loss function configured in the training optimization module, DiceLoss is for segmented regions, BCELoss positive sample weights range from 1.0 to 7.0, FocalLoss parameter γ=2.0, and training can optionally enable hybrid precision to improve training efficiency.

5. The intelligent gastrointestinal polyp auxiliary examination system according to claim 1, characterized in that, The reasoning process executed by the reasoning analysis module includes loading the model according to the preset configuration, test-time enhancement, 5-fold cross-validation model integration, and morphological post-processing.

6. An intelligent auxiliary examination method for gastrointestinal polyps, based on the intelligent auxiliary examination system for gastrointestinal polyps according to any one of claims 1-5, characterized in that, Includes the following steps: Data preprocessing: Preprocessing of gastrointestinal endoscopy images, including reading data in pairs from the image catalog and the annotation mask catalog; The dataset is divided based on K-fold cross-validation; the images and masks are scaled to a preset size. It employs a progressive data augmentation strategy of weak-medium-strong according to the training phase, and can choose to normalize / standardize pixel values ​​to improve training stability. Model training: Build an attention-guided UNet++ model, configure the hybrid loss function, RAdam optimizer and learning rate scheduling strategy, train the model and save the 5-fold model parameters; Online / offline inference: Receives gastrointestinal endoscopy images, calls the corresponding model parameters according to the preset endoscopy type / input source configuration, performs test-time enhancement, 5-fold model ensemble and morphological post-processing, and generates polyp segmentation results; Output results: Outputs derived information calculated from the mask, supporting display and data export.

7. The intelligent auxiliary examination method for gastrointestinal polyps according to claim 6, characterized in that, In the data preprocessing steps, the three-stage progressive data augmentation is divided according to the training cycle and endoscope type: Weak augmentation phase: In the early stage of training, relatively mild data augmentation operations are used to randomly flip the input image and make slight changes to brightness / contrast to stabilize the initial convergence of the model; Mid-stage augmentation: In the middle of training, moderate-intensity data augmentation operations are introduced, including random translation, scaling, and mild blurring, to enhance the model's adaptability to changes in viewpoint and imaging. Strong augmentation phase: Strong data augmentation operations are introduced in the later stages of training, including geometric distortion, CLAHE contrast enhancement, and noise perturbation, to improve the robustness of the model under complex imaging conditions.

8. The intelligent auxiliary examination method for gastrointestinal polyps according to claim 6, characterized in that, The attention-guided UNet++ network constructed during model training includes an encoder, an attention gate, a decoder, and an output layer. The encoder includes 5 convolutional blocks, and the attention gate is used to fuse encoded features and decoded guidance signals to generate attention weights to suppress irrelevant background. The decoder includes 4 convolutional blocks, and the output layer is used to output a segmentation probability map.

9. The intelligent auxiliary examination method for gastrointestinal polyps according to claim 6, characterized in that, The hybrid loss function used in the model training is Dice-BCE-Focal loss, with a weight ratio of 0.5:0.3:0.

2. DiceLoss is optimized for sensitivity to small polyps, BCELoss uses adaptive positive sample weights, and FocalLoss increases the weight of edge pixel loss. The model training uses the RAdam optimizer with an initial learning rate of 3e-4 and a weight decay of 1e-5. The learning rate scheduling adopts a cosine annealing hot restart strategy, where T0=10, T_mult=2, eta_min=1e-6. The early stopping strategy is to terminate training if the validation set metric does not improve for 20 consecutive cycles.

10. The intelligent auxiliary examination method for gastrointestinal polyps according to claim 6, characterized in that, In the online / offline inference steps: the test enhancement includes four transformations: original, horizontal flip, vertical flip, and horizontal + vertical flip, and the results of each transformation are fused to obtain the segmented output; Morphological post-processing performs 7×7 kernel closing operations and 3×3 kernel opening operations on the predicted mask, while retaining connected regions with an area of ​​not less than 400 pixels.

Images