Gastroscope image intelligent target detection method and system

Abstract

The invention relates to an intelligent target detection method and system for a gastroscope image, and the method comprises the steps: carrying out the multispectral decomposition of an obtained gastroscope image, and generating a standardized image in combination with a dynamic domain adaptation strategy; based on the standardized image, adopting a multi-scale geometric perception feature extraction network to extract fusion features; carrying out anatomical position sensing processing on the fusion features to generate position sensing features with anatomical region identification capability; on the basis of the position sensing features, morphologically-guided multi-task target detection and classification are executed, and a multi-task detection result is obtained; the multiple tasks comprise focus detection, focus boundary segmentation and focus type classification; aiming at a multi-task detection result of continuous image frames in the gastroscopy process, applying a time sequence consistency constraint, and generating a space-time optimization result; and constructing an uncertainty perception mechanism based on a space-time optimization result, and outputting a calibrated target detection result. The precision of gastroscope image detection can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method, system, electronic device, and non-transitory computer-readable storage medium for intelligent target detection in gastroscopic images. Background Art

[0002] Nowadays, gastroscopy is one of the important means widely used in clinical practice for the diagnosis and treatment of digestive system diseases. With the popularization of high-definition electronic endoscope equipment, medical staff can use images to identify gastric mucosal lesions such as erosions, ulcers, polyps and early cancer.

[0003] However, gastroscopic images themselves have characteristics such as irregular field of view, drastic changes in lighting, and large differences in lesion morphology, which lead to poor performance of conventional general models in detection accuracy and generalization ability. Summary of the Invention

[0004] In response to the technical problems existing in the prior art, the present invention provides a gastroscopic image intelligent target detection method, system, electronic device and non-transitory computer-readable storage medium that can improve the accuracy of gastroscopic image detection.

[0005] The technical solution of the present invention to solve the above technical problems is as follows: The present invention provides a method for intelligent target detection in gastroscopic images, the method comprising: Perform multispectral decomposition on the acquired gastroscopic images and combine them with dynamic domain adaptation strategy to generate standardized images; Based on the standardized image, a multi-scale geometric perception feature extraction network is used to extract fusion features; performing anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capability; Based on the position-aware features, performing morphology-guided multi-task target detection and classification to obtain a multi-task detection result; the multi-task includes lesion detection, lesion boundary segmentation, and lesion type classification; Applying temporal consistency constraints to the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; Based on the spatiotemporal optimization results, an uncertainty perception mechanism is constructed to output calibrated target detection results.

[0006] Optionally, performing multispectral decomposition on the acquired gastroscopic image and combining it with a dynamic domain adaptation strategy to generate a standardized image includes: Performing multispectral decomposition on the gastroscopic image to generate corresponding multi-channel spectral data; Based on the multi-channel spectral data, a depth information estimation algorithm is used to construct a three-dimensional geometric structure of the stomach surface; According to the three-dimensional geometric structure, an adaptive illumination equalization algorithm is used to dynamically adjust the illumination parameters of each area in the image to obtain a corrected image; Using an image enhancement network based on residual learning on the corrected image to enhance detail representation in low-contrast areas, thereby obtaining an enhanced image; The enhanced image is processed through a domain adaptation layer to eliminate image domain differences between different devices and output a standardized image.

[0007] Optionally, extracting fusion features based on the standardized image using a multi-scale geometric perception feature extraction network includes: Based on the standardized image, a dilated convolutional network is used in a main feature extraction path to extract high-resolution semantic features to obtain a first feature; Based on the three-dimensional geometric structure, a geometric constraint attention mechanism is introduced into the auxiliary feature extraction path to enhance the expression of structural information and obtain the second feature; The gated fusion unit is set to perform dynamic weighted fusion on the extracted first feature and second feature to generate a fusion feature.

[0008] Optionally, performing anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capability includes: Based on the fused features, identifying anatomical partitions in the stomach image through a position-aware module; According to the identified anatomical regions, the corresponding dedicated context encoder is called to extract regional features; Based on the regional features of each context encoder and the confidence of the position information of the anatomical partition, dynamic fusion is performed to generate the position-aware feature.

[0009] Optionally, performing morphology-guided multi-task target detection and classification based on the position-aware features to obtain a multi-task detection result includes: Based on the position-aware features, an anchor-free object detection head is used to predict the bounding box information of the lesion; Based on the bounding box information, performing fine boundary segmentation of the lesion area to generate a morphologically constrained mask; Combining the morphological constraint mask, classifying and identifying the lesion type to obtain an identification result; Based on the multi-task output of the bounding box information, the morphological constraint mask and the recognition result, a dynamic loss balancing strategy is adopted to fuse the multi-task training objectives to generate the multi-task detection result.

[0010] Optionally, applying temporal consistency constraints to the multi-task detection results of continuous image frames during gastroscopy to generate spatiotemporal optimization results includes: For the multi-task detection results of consecutive image frames, the key features of the current frame are cached through the short-term memory unit; Combined with the typical historical features stored in the long-term memory knowledge base, the spatiotemporal attention mechanism is used to model the association between the current frame and the historical frames; Based on the association modeling result, inter-frame consistency constraints are imposed on the key features, and a spatiotemporal optimized detection result is generated as the spatiotemporal optimization result.

[0011] Optionally, constructing an uncertainty perception mechanism based on the spatiotemporal optimization result and outputting a calibrated target detection result includes: Based on the spatiotemporal optimization results, a Bayesian neural network is used to estimate the uncertainty distribution of the lesion area to obtain a corresponding uncertainty distribution; Based on the uncertainty distribution, a self-calibration reanalysis process is initiated for the detected high-risk areas to obtain a detection model; By adopting the detection model enhanced by self-adversarial training, the target detection results in high-risk areas are calibrated, and the target detection results optimized by uncertainty perception are output.

[0012] Optionally, the method further includes: Based on the spatiotemporal optimization results, generating an uncertainty heat map reflecting the degree of uncertainty of each image region through the Bayesian neural network; performing a multi-scale self-calibration analysis on the uncertainty areas determined to be high risk according to the uncertainty heat map to obtain a self-calibration analysis result; Based on the self-calibration analysis results, a corresponding detection reliability evaluation report is generated, and the final calibrated target detection result is output through the human-machine collaborative decision interface.

[0013] Optionally, the method further includes: Based on the calibrated target detection results, a knowledge distillation and continuous learning process is performed to update model parameters and enhance the generalization ability of the model.

[0014] The present invention also provides a gastroscope image intelligent target detection system, the system comprising: The image generation module is used to perform multispectral decomposition on the acquired gastroscopic images and generate standardized images in combination with the dynamic domain adaptation strategy; A feature extraction module, configured to extract fusion features based on the standardized image using a multi-scale geometric perception feature extraction network; a position sensing module, configured to perform anatomical position sensing processing on the fused features to generate position sensing features with anatomical region recognition capability; An initial detection module is configured to perform morphologically guided multi-task target detection and classification based on the position-aware features to obtain a multi-task detection result; the multi-task includes lesion detection, lesion boundary segmentation, and lesion type classification; a detection optimization module, configured to impose temporal consistency constraints on the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; The target detection module is used to build an uncertainty perception mechanism based on the spatiotemporal optimization results and output calibrated target detection results.

[0015] In addition, to achieve the above-mentioned purpose, the present invention also proposes an electronic device, comprising: a memory for storing a computer software program; a processor for reading and executing the computer software program, thereby realizing a gastroscopic image intelligent target detection method as described above.

[0016] In addition, to achieve the above-mentioned purpose, the present invention also proposes a non-transitory computer-readable storage medium, in which a computer software program is stored. When the computer software program is executed by a processor, it implements a gastroscopic image intelligent target detection method as described above.

[0017] The beneficial effects of the present invention are: (1) The present invention effectively suppresses the interference caused by uneven illumination, image quality fluctuations and equipment differences by introducing multi-spectral decomposition and adaptive illumination correction mechanism. The generated standardized image has a more stable texture and light intensity distribution, providing a reliable basis for subsequent feature extraction.

[0018] (2) The present invention uses a multi-scale geometric perception network to integrate image semantic features while introducing stereo geometric constraints and attention mechanisms, significantly improving the ability to express the boundary details and spatial structure of the lesion area, which is conducive to the accurate identification of complex morphological lesions.

[0019] (3) The present invention identifies specific areas of the stomach through an anatomical position perception module and adopts a dedicated context encoder model to construct a position perception feature with regional recognition capabilities, thereby achieving precise alignment of lesion detection and classification in spatial positioning and improving diagnostic relevance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a gastroscopic image intelligent target detection method provided by the present invention; Figure 2 A schematic diagram of the structure of an intelligent target detection system for gastroscopic images provided by the present invention; Figure 3 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention; Figure 4 A schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0023] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0024] See also Figure 1 , provides a flow chart of a gastroscope image intelligent target detection method of the present invention, comprising the following steps: Step 201: Perform multispectral decomposition on the acquired gastroscopic image and generate a standardized image by combining it with a dynamic domain adaptation strategy.

[0025] In specific implementations, multispectral channel reconstruction of gastroscopic RGB images can be performed to extract richer spectral information. Combining depth estimation with illumination balancing algorithms allows for dynamic illumination adjustment of different image regions, enhancing image detail. Finally, domain adaptation techniques (e.g., eliminating image style differences between different devices) are used to output a unified, standardized image suitable for subsequent processing.

[0026] In some embodiments, step 201 may include: Perform multispectral decomposition on gastroscopic images to generate corresponding multi-channel spectral data; Based on multi-channel spectral data, the three-dimensional geometric structure of the stomach surface is constructed using a depth information estimation algorithm; According to the three-dimensional geometric structure, the adaptive illumination equalization algorithm is used to dynamically adjust the illumination parameters of each area in the image to obtain the corrected image; The image enhancement network based on residual learning is used to improve the details of low-contrast areas in the corrected image to obtain an enhanced image. The enhanced image is processed through the domain adaptation layer to eliminate the image domain differences between different devices and output a standardized image.

[0027] Specifically, acquired RGB-format gastroscopic images can be fed into a multispectral decomposition module. Based on an image reconstruction algorithm or a well-trained multispectral estimation network, this module recovers a richer set of virtual spectral channels from the limited RGB channels, generating multichannel image data containing multiple bands. By simulating the imaging characteristics of near-infrared, ultraviolet, or specific color reflectance bands, it effectively enhances the representation of areas with less pronounced color differences in the image, particularly significantly improving the separability of low-contrast lesions.

[0028] Subsequently, based on the multi-channel spectral images obtained above, a depth information estimation algorithm can be used to perform spatial modeling of the image scene. Specifically, a monocular depth estimation network based on an encoder-decoder structure can be used to extract the depth value corresponding to each pixel in the image and generate a depth map of the gastric surface. Furthermore, a point cloud or triangular mesh model can be reconstructed from the depth map to construct the three-dimensional geometric structure of the stomach surface, providing geometric prior support for illumination balance and position reasoning.

[0029] After completing the three-dimensional geometric modeling, adaptive illumination balancing is performed based on the three-dimensional structural information to address the common non-uniform illumination issues found in gastroscopic images. By estimating the incident light direction, reflection intensity, and surface normal characteristics of each region, the brightness and contrast of different locations in the image are dynamically adjusted. This illumination balancing strategy can be combined with Retinex theory, physical rendering model inversion methods, or a neural network-based illumination regression module to accurately correct for highlights, shadows, and edge occlusion, resulting in a corrected image.

[0030] For areas of low contrast, blur, or texture loss that remain after illumination correction, the image is further fed into a residual learning-based image enhancement network for detail enhancement. This enhancement network utilizes a multi-layered residual block structure with skip connections and nonlinear mapping capabilities, effectively preserving the original image structure while enhancing local edge clarity and the perceptibility of subtle lesions. The resulting output image boasts enhanced structural expressiveness and visual clarity, making it suitable for subsequent detection tasks.

[0031] To address the problem of inconsistent image style caused by differences in source devices (e.g., different brands of endoscopic imaging systems), this paper introduces a domain adaptation module to standardize the enhanced images. Specifically, by incorporating an adversarial domain discriminator, a minimum mean difference (MMD) regularization term, or a normalized domain transfer mechanism, this module achieves unsupervised alignment of image feature distributions, aligning image features from different imaging devices in the latent space. This results in a standardized image with a uniform style and clear structure.

[0032] Through the above multi-stage processing process, the present invention effectively improves the versatility, recognizability and quality stability of gastroscopic images, and provides better input data for subsequent lesion detection networks.

[0033] Step 202: Based on the standardized image, a multi-scale geometric perception feature extraction network is used to extract fusion features.

[0034] In specific implementation, a neural network combining the main path and auxiliary path structures can be used to extract high-resolution semantic features of the image in the main path, and introduce stomach geometric structure information (such as shape or surface convexity) in the auxiliary path. Through the attention mechanism and gated fusion unit, cross-scale and cross-information source feature fusion can be achieved, thereby improving the network's perception of complex lesion morphology.

[0035] In some embodiments, step 202 may include: Based on the standardized image, a dilated convolutional network is used in the main feature extraction path to extract high-resolution semantic features to obtain the first feature; Based on the three-dimensional geometric structure, a geometric constraint attention mechanism is introduced into the auxiliary feature extraction path to enhance the expression of structural information and obtain the second feature; The gated fusion unit is set to perform dynamic weighted fusion on the extracted first feature and second feature to generate a fusion feature.

[0036] First, the standardized gastroscopic images are fed into the main path feature extraction network. The main path uses a dilated convolution architecture to construct a deep neural network. This architecture expands the receptive field without increasing computational complexity, allowing it to capture a wider range of contextual semantic information while maintaining the spatial resolution of the feature map. High-resolution semantic features extracted through the multi-layer stack of dilated convolution modules are recorded as the first feature, primarily containing texture, edge, and semantic category information related to the lesion area.

[0037] Furthermore, to enhance the model's ability to model the three-dimensional structure of the stomach, a geometrically constrained attention mechanism was introduced to construct an auxiliary path feature extraction module based on the previously constructed three-dimensional stomach geometry. This module generates a spatial attention map by combining geometric priors such as the depth map, surface normal vectors, and spatial curvature. This guides the neural network to focus more on anatomically significant areas and areas of sudden curvature changes during feature extraction. The final output is a structural feature with geometric structure awareness, recorded as the second feature.

[0038] To effectively integrate semantic and geometric features, the present invention implements a gated fusion module (Gated Fusion Module) at the back end of both the primary and auxiliary paths. This module dynamically learns fusion weights and performs a channel-by-channel weighted fusion of the first and second features based on the importance and complementarity of the input features. This fusion process can be implemented using gating mechanisms (such as sigmoid gates), self-attention mechanisms, or convolutional weighting. The final output is a fusion feature that combines structural and semantic synergy, which is used in subsequent position perception processing and object detection modules.

[0039] By constructing a dual-path feature extraction network based on dilated convolution and geometric attention mechanism and introducing a gated fusion strategy, the present invention can enhance the joint expression capability of gastric wall structure, anatomical boundaries and lesion semantics while maintaining the spatial resolution of the image, significantly improving the accuracy and robustness of the target detection system.

[0040] Step 203: Perform anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capabilities.

[0041] In specific implementation, a position-aware module can be introduced to identify the anatomical region of the stomach (such as the cardia, body, antrum, etc.) in the gastroscopic image. Then, a region-specific context encoder is used to model the semantic features of each region, and feature fusion is performed based on the region confidence to generate a position-aware feature map with anatomical structure information.

[0042] In some embodiments, step 203 may include: Based on the fused features, the location-aware module identifies anatomical partitions in the stomach image. According to the identified anatomical regions, the corresponding dedicated context encoder is called to extract regional features; Based on the regional features of each context encoder and the position information confidence of the anatomical partition, dynamic fusion is performed to generate position-aware features.

[0043] First, the fused features are input into the location-aware module. Based on a convolutional neural network architecture, this module, combined with an attention mechanism or a region proposal network, automatically identifies key anatomical regions in gastroscopic images, such as the cardia, gastric body, gastric angle, and pyloric region. The module outputs the anatomical region's category labels, their corresponding location information, and confidence scores, which are used for subsequent regional feature modeling.

[0044] Based on the categories of each identified anatomical region, the system invokes a dedicated context encoder (RCE) corresponding to each region. This encoder utilizes a neural network structure tailored to the structural characteristics of the target region to mine deep features related to lesion distribution, physiological texture, and spatial context. Each RCE has a higher expressive power for the features of the region it processes, thereby improving the overall system's recognition accuracy even in complex structures.

[0045] Furthermore, the regional features output by each context encoder, along with the positional confidence of the corresponding anatomical partition, are fed into a feature fusion module. This module employs a weighted fusion strategy, automatically adjusting the contribution of each regional feature to the final result based on the confidence level. Fusion can be performed using channel-by-channel weighted fusion, attention-weighted aggregation, or information fusion controlled by a gating mechanism. The final output is a position-aware feature map with regional differentiation and structural awareness, which is then used by subsequent object detection and classification modules.

[0046] Through the above-mentioned anatomical position perception processing flow based on structural recognition and regional feature differential encoding, the present invention can effectively enhance the model's sensitivity to lesion manifestations in different anatomical regions, reduce the risk of false detection and missed detection due to structural similarity or texture confusion, and significantly improve the accuracy and clinical interpretability of lesion detection and classification results.

[0047] Step 204: Based on the location-aware features, perform morphology-guided multi-task object detection and classification to obtain a multi-task detection result.

[0048] The multi-tasks include lesion detection, lesion boundary segmentation and lesion type classification.

[0049] In its implementation, the model simultaneously performs three tasks: lesion target detection (detection bounding boxes), lesion boundary segmentation (fine contours), and lesion type classification (such as inflammation, ulcer, and tumor). This is achieved through an anchor-free detection head and a morphologically constrained masking mechanism, enabling collaborative guidance between these tasks. A dynamic loss balancing strategy is employed to optimize training results, improving the efficiency and accuracy of multi-task learning.

[0050] In some embodiments, step 204 may include: Based on position-aware features, an anchor-free object detection head is used to predict the bounding box information of the lesion; Based on the bounding box information, fine boundary segmentation of the lesion area is performed to generate a morphologically constrained mask; Combined with the morphological constraint mask, the lesion type is classified and identified to obtain the recognition result; Based on the multi-task output of bounding box information, morphological constraint mask and recognition results, a dynamic loss balance strategy is adopted to fuse multi-task training objectives and generate multi-task detection results.

[0051] First, the position-aware features are fed into the anchor-free object detection subnetwork (Anchor-Free Detection Head). This detection head uses center point estimation and boundary regression to avoid the redundant matching problem in traditional anchor frame mechanisms. It directly predicts the center position, boundary offset, and confidence score of the lesion target, thereby generating lesion bounding box information, achieving higher positioning accuracy and model convergence efficiency.

[0052] Based on the bounding box information, the detected suspected lesion area is cropped or feature-aligned and input into the boundary fine segmentation module. This module uses a multi-scale residual convolutional network or decoder structure to perform pixel-level segmentation on the lesion area edge and generate a high-precision lesion morphology constraint mask to depict the true boundary contour and shape characteristics of the lesion.

[0053] Furthermore, the generated morphologically constrained mask is used as a priori guidance and combined with regional features to enter the lesion type classification module. This module can be a fully connected classifier, an attention-guided classifier, or a graph neural network classifier. It identifies and classifies the lesion based on morphology, texture, and contextual information, and outputs a predicted lesion type, such as erosion, ulcer, hemorrhage, or mass.

[0054] To achieve effective collaboration among the subtasks, this paper further introduces a dynamic loss balancing strategy to adaptively adjust the weights of the loss functions for the three tasks of lesion detection (bounding box localization), segmentation (boundary mask generation), and classification (classification). This strategy dynamically adjusts the contribution of each loss term based on the gradient magnitude, convergence speed, or uncertainty of each task during training, thereby generating an overall multi-task detection result, achieving consistency in optimization objectives and a comprehensive improvement in final performance.

[0055] Through a multi-task collaborative detection framework guided by morphological features, this paper not only accurately locates and restores lesions' boundaries, but also improves the reliability of lesion type classification. By employing an anchor-free frame structure and a dynamic loss fusion strategy, the system effectively improves training efficiency, detection robustness, and clinical practicality. It is particularly suitable for identifying gastroscopic lesions with distinct morphological differences but fuzzy regional boundaries.

[0056] Step 205: Apply temporal consistency constraints to the multi-task detection results of continuous image frames during the gastroscopy process to generate spatiotemporal optimization results.

[0057] In specific implementation, we can target the continuous frame image data of actual gastroscopy operations, build a short-term memory module and a long-term knowledge base, and use the spatiotemporal attention mechanism to analyze the correlation between the current frame and historical frames, thereby imposing inter-frame consistency constraints on the detection results, alleviating recognition fluctuations caused by perspective jitter or image quality changes, and outputting more stable and continuous spatiotemporal optimization detection results.

[0058] In some embodiments, step 205 may include: For the multi-task detection results of consecutive image frames, the key features of the current frame are cached through the short-term memory unit; Combined with the typical historical features stored in the long-term memory knowledge base, the spatiotemporal attention mechanism is used to model the association between the current frame and the historical frames; Based on the association modeling results, inter-frame consistency constraints are imposed on key features, and the spatiotemporal optimized detection results are generated as spatiotemporal optimization results.

[0059] For each frame of standardized gastroscopic image, representative key features are extracted from the fused features, boundary information, and classification results output by the multi-task detection module and then cached in a short-term memory unit (STM). This unit uses a sliding window mechanism to maintain local contextual information across multiple consecutive frames, modeling short-term dynamic changes between images.

[0060] In parallel, a long-term memory repository (LTM) is established within the system. This repository stores typical feature representations extracted from several frames of historical images, covering different anatomical regions of the stomach, different types of lesions, and their variations. By introducing a spatio-temporal attention module (STM), the short-term memory features of the current frame are aligned with historical features in the LTM, capturing inter-frame temporal context, feature overlap, and variation trends.

[0061] This mechanism performs weighted association based on key regions, temporal position information and feature similarity, and outputs a spatiotemporal association representation between the current frame and multiple historical frames.

[0062] Based on the above spatiotemporal correlation modeling results, an inter-frame consistency constraint is introduced for the key features of the current frame. This constraint is used to limit sudden changes in the boundaries, classification, and location of the same lesion target in adjacent frames along the time axis, thereby improving the continuity and reliability of detection. Inter-frame consistency can be achieved through temporal smoothing, feature alignment loss, or attention-guided dynamic weight adjustment.

[0063] Finally, the detection output after imposing inter-frame consistency constraints is used as the spatio-temporally optimized detection result (Spatio-Temporally Optimized Output) for further uncertainty modeling or clinical output.

[0064] Through the above-mentioned processing method that combines short-term memory, long-term knowledge and spatiotemporal attention mechanism, the present invention can significantly suppress inter-frame jitter, reduce mutations and error accumulation in detection results when facing a continuous gastroscopy image sequence, effectively improve the temporal stability and continuity of lesion detection, and ensure the reliability and consistency of system output in actual clinical applications.

[0065] Step 206: Based on the spatiotemporal optimization results, an uncertainty perception mechanism is constructed to output calibrated target detection results.

[0066] In practice, techniques such as Bayesian neural networks or MC Dropout can be used to estimate the uncertainty distribution of each test result. For areas of high uncertainty, the system can perform self-calibration analysis (e.g., re-evaluating or triggering more refined reasoning), and incorporate self-adversarial enhancement mechanisms to improve model robustness. Ultimately, the system outputs more reliable, calibrated test results suitable for clinical diagnosis.

[0067] In some embodiments, step 206 may include: Based on the spatiotemporal optimization results, the Bayesian neural network is used to estimate the uncertainty distribution of the lesion area and obtain the corresponding uncertainty distribution; Based on the uncertainty distribution, a self-calibration reanalysis process is initiated for the detected high-risk areas to obtain a detection model; By adopting a detection model enhanced by self-adversarial training, the target detection results in high-risk areas are calibrated and the target detection results optimized by uncertainty perception are output.

[0068] First, the fused features generated from the multi-task spatiotemporal optimization results are input into a Bayesian Neural Network (BNN). By incorporating parameter estimation into the probability distribution, the BNN outputs an uncertainty distribution for the lesion area, accurately reflecting the confidence level and potential risk of the model's predictions. This uncertainty distribution is used to distinguish high-risk from low-risk areas, assisting in subsequent focused treatment.

[0069] Based on the uncertainty distribution, the system automatically identifies high-risk areas with elevated uncertainty in the test results. For these areas, a self-calibration reanalysis process is initiated. This process uses uncertainty feedback to adjust model parameters, re-extract features, and optimize detection strategies, thereby building or updating the corresponding self-calibration detection model. This model is specifically designed to address the complex features and fuzzy boundaries of high-risk areas, improving detection accuracy and stability.

[0070] Furthermore, to enhance the model's generalization and interference resistance in complex environments, a self-adversarial training strategy was employed to strengthen the self-calibration detection model. By simulating different interference scenarios and lesion variations, the model learned to better identify and correct misjudgments, particularly by fine-tuning detection results for targets in high-risk areas.

[0071] Ultimately, the target detection result is output after uncertainty perception and self-calibration optimization. This result not only has higher accuracy but also has risk assessment capabilities, providing reliable support for clinical diagnosis.

[0072] By introducing the uncertainty estimation and self-calibration reanalysis mechanism of the Bayesian neural network and combining it with self-adversarial training to enhance the model capabilities, the present invention effectively solves the problem of insufficient recognition of uncertain lesion areas in traditional detection methods, significantly improves the accuracy and robustness of the detection results, and enhances the practicality and credibility of the system in complex gastroscopic image environments.

[0073] In some embodiments, step 206 may further include: Based on the spatiotemporal optimization results, an uncertainty heat map reflecting the degree of uncertainty in each image region is generated through a Bayesian neural network. Based on the uncertainty heat map, a multi-scale self-calibration analysis is performed on the uncertain areas determined to be high risk to obtain the self-calibration analysis results; Based on the self-calibration analysis results, a corresponding detection reliability assessment report is generated, and the final calibrated target detection results are output through the human-computer collaborative decision-making interface.

[0074] First, the spatiotemporally optimized multi-task detection results and their corresponding features are fed into a Bayesian Neural Network (BNN). This network uses probabilistic modeling to generate an uncertainty heatmap reflecting the degree of uncertainty in each image region. This heatmap visually displays the risk level of different gastroscopic image regions through spatial distribution, providing a basis for subsequent risk focus and analysis.

[0075] Based on the prominent areas identified as high-risk in the uncertainty heat map, the system further initiates a multi-scale self-calibration analysis process. This process involves re-extracting fine-grained features, refining boundaries, and reclassifying high-risk areas at different spatial scales and resolutions. Combining prior knowledge with a model's adaptive adjustment strategy, it generates self-calibration analysis results tailored to the specific lesion morphology and location, effectively improving the accuracy and stability of high-risk area detection.

[0076] Based on the self-calibration analysis results, the system comprehensively calculates reliability assessment indicators for each detection target, generating a comprehensive detection reliability assessment report. This report includes information such as risk classification, confidence statistics, and recommended treatment plans. Finally, through the designed human-computer collaborative decision-making interface, the self-calibrated and optimized target detection results are output together with the reliability assessment information, assisting clinicians in making accurate diagnostic decisions and enhancing the practical value and reliability of the medical assistance system.

[0077] The present invention achieves comprehensive risk management and dynamic optimization of gastroscopy image detection results through intuitive risk visualization of uncertainty heat maps, precise adjustment of multi-scale self-calibration, and decision support of human-computer collaboration, significantly improving the stability, accuracy and clinical application value of the system in complex lesion identification.

[0078] See also Figure 2 , Figure 2 This is a structural diagram of an intelligent target detection system for gastroscopic images provided by the present invention.

[0079] like Figure 2 As shown, an intelligent target detection system for gastroscopic images proposed in an embodiment of the present invention includes: The image generation module 301 is used to perform multispectral decomposition on the acquired gastroscopic image and generate a standardized image in combination with a dynamic domain adaptation strategy; A feature extraction module 302 is configured to extract fusion features based on the standardized image using a multi-scale geometric perception feature extraction network; A position sensing module 303 is configured to perform anatomical position sensing processing on the fused features to generate position sensing features with anatomical region recognition capability; An initial detection module 304 is configured to perform morphology-guided multi-task target detection and classification based on the position-aware features to obtain a multi-task detection result; the multi-task includes lesion detection, lesion boundary segmentation, and lesion type classification; A detection optimization module 305 is configured to impose temporal consistency constraints on the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; The target detection module 306 is used to build an uncertainty perception mechanism based on the spatiotemporal optimization result and output a calibrated target detection result.

[0080] See also Figure 3 , Figure 3 Schematic diagram of an embodiment of an electronic device provided by an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an electronic device 400, including a memory 410, a processor 420, and a computer program 411 stored in the memory 410 and executable on the processor 420. When the processor 420 executes the computer program 411, the following steps are implemented: Perform multispectral decomposition on the acquired gastroscopic images and combine them with dynamic domain adaptation strategy to generate standardized images; Based on the standardized image, a multi-scale geometric perception feature extraction network is used to extract fusion features; performing anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capability; Based on the position-aware features, performing morphology-guided multi-task target detection and classification to obtain a multi-task detection result; the multi-task includes lesion detection, lesion boundary segmentation, and lesion type classification; Applying temporal consistency constraints to the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; Based on the spatiotemporal optimization results, an uncertainty perception mechanism is constructed to output calibrated target detection results.

[0081] See also Figure 4 , Figure 4 Schematic diagram of an embodiment of a computer-readable storage medium provided in an embodiment of the present invention. Figure 4As shown, this embodiment provides a computer-readable storage medium 500 on which a computer program 411 is stored. When the computer program 411 is executed by a processor, the following steps are implemented: Perform multispectral decomposition on the acquired gastroscopic images and combine them with dynamic domain adaptation strategy to generate standardized images; Based on the standardized image, a multi-scale geometric perception feature extraction network is used to extract fusion features; performing anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capability; Based on the position-aware features, performing morphology-guided multi-task target detection and classification to obtain a multi-task detection result; the multi-task includes lesion detection, lesion boundary segmentation, and lesion type classification; Applying temporal consistency constraints to the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; Based on the spatiotemporal optimization results, an uncertainty perception mechanism is constructed to output calibrated target detection results.

[0082] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0083] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes 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 computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A system that specifies the functions of a box or boxes.

[0085] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction system that is implemented in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0087] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0088] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for intelligent target detection in gastroscopic images, characterized in that: The method comprises: Perform multispectral decomposition on the acquired gastroscopic images and combine them with dynamic domain adaptation strategy to generate standardized images; Based on the standardized image, a multi-scale geometric perception feature extraction network is used to extract fusion features; performing anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capability; Based on the position-aware features, performing morphology-guided multi-task target detection and classification to obtain a multi-task detection result; the multi-task includes lesion detection, lesion boundary segmentation, and lesion type classification; Applying temporal consistency constraints to the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; Based on the spatiotemporal optimization results, an uncertainty perception mechanism is constructed to output calibrated target detection results.

2. The method for intelligent target detection in gastroscopic images according to claim 1, characterized in that: The multispectral decomposition of the acquired gastroscopic image and the generation of a standardized image in combination with a dynamic domain adaptation strategy include: Performing multispectral decomposition on the gastroscopic image to generate corresponding multi-channel spectral data; Based on the multi-channel spectral data, a depth information estimation algorithm is used to construct a three-dimensional geometric structure of the stomach surface; According to the three-dimensional geometric structure, an adaptive illumination equalization algorithm is used to dynamically adjust the illumination parameters of each area in the image to obtain a corrected image; Using an image enhancement network based on residual learning on the corrected image to enhance detail representation in low-contrast areas, thereby obtaining an enhanced image; The enhanced image is processed through a domain adaptation layer to eliminate image domain differences between different devices and output a standardized image.

3. The method for intelligent target detection in gastroscopic images according to claim 2, characterized in that: The step of extracting fusion features based on the standardized image using a multi-scale geometric perception feature extraction network includes: Based on the standardized image, a dilated convolutional network is used in a main feature extraction path to extract high-resolution semantic features to obtain a first feature; Based on the three-dimensional geometric structure, a geometric constraint attention mechanism is introduced into the auxiliary feature extraction path to enhance the expression of structural information and obtain the second feature; The gated fusion unit is set to perform dynamic weighted fusion on the extracted first feature and second feature to generate a fusion feature.

4. The method for intelligent target detection in gastroscopic images according to claim 3, characterized in that: The performing anatomical position-aware processing on the fused features to generate position-aware features with anatomical region recognition capability includes: Based on the fused features, identifying anatomical partitions in the stomach image through a position-aware module; According to the identified anatomical regions, the corresponding dedicated context encoder is called to extract regional features; Based on the regional features of each context encoder and the confidence of the position information of the anatomical partition, dynamic fusion is performed to generate the position-aware feature.

5. The method for intelligent target detection in gastroscopic images according to claim 4, characterized in that: The performing of morphologically guided multi-task target detection and classification based on the position-aware features to obtain a multi-task detection result includes: Based on the position-aware features, an anchor-free object detection head is used to predict the bounding box information of the lesion; Based on the bounding box information, performing fine boundary segmentation of the lesion area to generate a morphologically constrained mask; Combining the morphological constraint mask, classifying and identifying the lesion type to obtain an identification result; Based on the multi-task output of the bounding box information, the morphological constraint mask and the recognition result, a dynamic loss balancing strategy is adopted to fuse the multi-task training objectives to generate the multi-task detection result.

6. The method for intelligent target detection in gastroscopic images according to claim 5, characterized in that: The multi-task detection results of continuous image frames during gastroscopy are subjected to temporal consistency constraints to generate spatiotemporal optimization results, including: For the multi-task detection results of consecutive image frames, the key features of the current frame are cached through the short-term memory unit; Combined with the typical historical features stored in the long-term memory knowledge base, the spatiotemporal attention mechanism is used to model the association between the current frame and the historical frames; Based on the association modeling result, inter-frame consistency constraints are imposed on the key features, and a spatiotemporal optimized detection result is generated as the spatiotemporal optimization result.

7. The method for intelligent target detection in gastroscopic images according to claim 6, characterized in that: The step of constructing an uncertainty perception mechanism based on the spatiotemporal optimization result and outputting a calibrated target detection result includes: Based on the spatiotemporal optimization results, a Bayesian neural network is used to estimate the uncertainty distribution of the lesion area to obtain a corresponding uncertainty distribution; Based on the uncertainty distribution, a self-calibration reanalysis process is initiated for the detected high-risk areas to obtain a detection model; By adopting the detection model enhanced by self-adversarial training, the target detection results in high-risk areas are calibrated, and the target detection results optimized by uncertainty perception are output.

8. The method for intelligent target detection in gastroscopic images according to claim 7, characterized in that: The method further comprises: Based on the spatiotemporal optimization results, generating an uncertainty heat map reflecting the degree of uncertainty of each image region through the Bayesian neural network; performing a multi-scale self-calibration analysis on the uncertainty areas determined to be high risk according to the uncertainty heat map to obtain a self-calibration analysis result; Based on the self-calibration analysis results, a corresponding detection reliability evaluation report is generated, and the final calibrated target detection result is output through the human-machine collaborative decision interface.

9. The method for intelligent target detection in gastroscopic images according to claim 8, characterized in that: The method further comprises: Based on the calibrated target detection results, a knowledge distillation and continuous learning process is performed to update model parameters and enhance the generalization ability of the model.

10. An intelligent target detection system for gastroscopic images, characterized in that: The system comprises: The image generation module is used to perform multispectral decomposition on the acquired gastroscopic images and generate standardized images in combination with the dynamic domain adaptation strategy; A feature extraction module, configured to extract fusion features based on the standardized image using a multi-scale geometric perception feature extraction network; a position sensing module, configured to perform anatomical position sensing processing on the fused features to generate position sensing features with anatomical region recognition capability; An initial detection module is configured to perform morphologically guided multi-task target detection and classification based on the position-aware features to obtain multi-task detection results; the multi-tasks include lesion detection, lesion boundary segmentation, and lesion type classification; a detection optimization module, configured to impose temporal consistency constraints on the multi-task detection results of consecutive image frames during gastroscopy to generate spatiotemporal optimization results; The target detection module is used to build an uncertainty perception mechanism based on the spatiotemporal optimization results and output calibrated target detection results.

Images