Visual intelligent detection method for accurately identifying stomach polyp

Abstract

The invention discloses a visual intelligent detection method for accurate recognition of stomach polyps, and relates to the technical field of artificial intelligence and medical image processing, and the method comprises the steps: collecting a gastroscope image, and carrying out the preprocessing; inputting the preprocessed gastroscope image into a YOLO11n backbone network for feature extraction to obtain a multi-scale feature map; the feature map pairs of adjacent scales are input to a C2PSAEDFN module for feature enhancement; after the enhanced multi-scale feature maps are up-sampled to the same spatial scale, splicing is performed in a channel dimension, and information is fused through convolution operation to generate fusion features; and inputting the fused features into a detection head module for classification and regression prediction, outputting category confidence and bounding box coordinates of the polyp, and filtering redundant boxes by adopting a non-maximum suppression algorithm. According to the method, the YOLO11n target detection network architecture is improved, so that the detection precision and the positioning stability of the multi-scale polyp target in the gastroscope image are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical image processing technology, and in particular to a visual intelligent detection method for accurate identification of gastric polyps. Background Technology

[0002] Gastric polyps are common lesions on the surface of the gastric mucosa, and some types have the potential to develop into cancer, thus requiring early detection and treatment. Currently, clinical diagnosis mainly relies on manual observation of gastroscopy images by doctors. However, manual detection suffers from problems such as strong reliance on experience, low efficiency, and a high risk of missed detections, especially when dealing with small polyps or those with indistinct borders, making accurate identification even more difficult.

[0003] With the development of artificial intelligence, deep learning-based image recognition technology has been gradually applied to medical image analysis tasks. Among them, the YOLO series models have attracted widespread attention due to their fast detection speed and simple structure. However, the traditional YOLO structure has certain limitations in handling complex backgrounds, recognizing fine-grained small targets, and targets with blurred edges. For example, it suffers from insufficient feature fusion and insufficient semantic information expression capabilities, making it difficult to meet the needs of refined medical detection. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a visual intelligent detection method for accurate identification of gastric polyps, which solves the problems of insufficient feature fusion and insufficient semantic information expression ability of the traditional YOLO structure when dealing with complex backgrounds, identifying fine-grained small targets and targets with blurred edges.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a visual intelligent detection method for accurate identification of gastric polyps, comprising: acquiring gastroscopy images and preprocessing them; inputting the preprocessed gastroscopy images into a YOLO11n backbone network for feature extraction to obtain multi-scale feature maps; inputting adjacent scale feature map pairs into a dual-domain collaborative enhancement module (C2PSA_EDFN module) for feature enhancement to generate enhanced multi-scale feature maps; upsampling the enhanced multi-scale feature maps to the same spatial scale and then concatenating them along the channel dimension, further fusing information through convolution operations to generate unified fused features rich in multi-scale context; inputting the fused features into a detection head module for classification and regression prediction, outputting the polyp category confidence and precise bounding box coordinates, and using a non-maximum suppression algorithm to filter redundant boxes to obtain the final detection result; superimposing the final detection result back onto the original gastroscopy image to form a visualization result, and simultaneously generating a detection report.

[0007] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the preprocessing includes normalization, noise suppression, color enhancement, and size uniformity.

[0008] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the YOLO11n backbone network includes multiple channels compression bottleneck modules and spatial pyramid pooling modules.

[0009] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the dual-domain collaborative enhancement module (C2PSA_EDFN module) includes a boundary-sensitive attention mechanism and a frequency domain enhanced feedforward network.

[0010] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the boundary-sensitive attention mechanism includes calculating the spatial and channel correlation between features through a multi-head self-attention mechanism; The PSA module is used to implement multi-head attention mapping in both horizontal and vertical directions through depthwise separable convolution, which enhances the expression of boundary and detail features.

[0011] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the frequency domain enhanced feedforward network includes: expanding the feature dimension using 1×1 convolution, extracting local spatial information using depthwise separable convolution, and simultaneously dividing the input features into patches of fixed size. For each patch, a two-dimensional fast Fourier transform is performed to convert the features to the frequency domain. The spectrum is modulated by a learnable frequency domain filter and then inverse FFT is performed to transform it back to the time domain, thus achieving effective fusion of frequency domain and spatial information.

[0012] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the generation of unified fusion features rich in multi-scale context includes upsampling the multi-scale feature maps enhanced by the dual-domain collaborative enhancement module (C2PSA_EDFN module) to the same spatial resolution in the spatial dimension, splicing them in the channel dimension, and further fusing them through 3×3 convolution to generate a unified multi-scale fusion feature representation.

[0013] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the detection head module includes three layers of output, corresponding to the fusion features of three scales, P3, P4 and P5, respectively, to cover polyp detection tasks of small, medium and large sizes.

[0014] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the method for obtaining the final detection result includes performing bounding box regression and category classification on the fused feature map using a detection head module, and outputting the location coordinates, confidence level, and category probability of the polyp. By combining multi-scale prediction with a non-maximum suppression strategy, duplicate detection boxes are effectively removed, and the final detection results are obtained.

[0015] As a preferred embodiment of the visual intelligent detection method for accurate identification of gastric polyps described in this invention, the detection report records in detail the location, size, and detection confidence level of the polyp.

[0016] The beneficial effects of this invention are as follows: By introducing a dual-domain collaborative enhancement module (C2PSA_EDFN module), joint modeling of boundaries and fine textures is achieved, and adaptive recalibration of channels and space is completed by combining Sigmoid, which improves the discrimination power and overall recall rate for small and blurred polyps; by upsampling the enhanced multi-scale feature maps to the same spatial scale, splicing them in the channel dimension and performing convolutional deep fusion, the unification and semantic alignment of multi-scale context are achieved, which improves the more stable localization performance and overall detection accuracy of polyps of different scales; by outputting bounding boxes and class confidence by a three-branch detection head and combining non-maximum suppression, redundancy suppression and result deduplication of multi-scale prediction are achieved, and finally, unique and accurate detection results are achieved, which are convenient for overlay visualization and report generation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a visual intelligent detection method for the accurate identification of gastric polyps. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Reference Figure 1 This is one embodiment of the present invention, which provides a visual intelligent detection method for accurate identification of gastric polyps, comprising the following steps: This invention employs an improved PrecisePolypNet network based on the YOLO11n network. This network includes a backbone network with multi-scale perception capabilities and a detection head module. The backbone network adopts a deep and shallow feature fusion architecture and introduces a dual-domain collaborative enhancement module (C2PSA_EDFN module) to enhance feature representation capabilities. The PrecisePolypNet network outputs feature maps based on three scales: P3, P4, and P5. Through feature fusion paths, it performs upsampling and connection operations to achieve joint detection of small, medium, and large targets, ultimately outputting the location and category of polyps, thus achieving high-precision identification and localization of gastric polyps.

[0023] Furthermore, the training and optimization steps for the PrecisePolypNet network are as follows: We used a publicly available dataset of gastric endoscopic images, supplemented with manually added samples, to cover gastric polyps of different shapes, sizes, and locations. To improve the model's generalization ability in complex environments, we implemented several enhancement strategies on the training data, including: horizontal and vertical image flipping, HSV color gamut adjustment (brightness, saturation, and contrast changes), random scaling and cropping, and blurring and noise perturbation. These steps expanded the dataset, significantly broadening the sample distribution while maintaining the validity of the bounding boxes, thus improving the model's robustness to polyps under different poses, lighting conditions, and scales.

[0024] It should be noted that the publicly available gastrointestinal image dataset used for training the PrecisePolypNet network contains 758 images. Based on this, the present invention uses data augmentation techniques, such as adjusting brightness and adding Gaussian noise, to expand the dataset to 3032 images.

[0025] Furthermore, the gastric endoscopy image dataset is divided into a training set and a test set. The training set consists of 900 images from Kvasir-SEG and 550 images from CVC-ClinicDB, totaling 1450 samples. The test set has five parts: the remaining 100 images from Kvasir-SEG; the remaining 62 images from CVC-ClinicDB; the complete CVC-ColonDB dataset (380 samples); the complete ETIS dataset (196 samples); and 60 images from the CVC-300 dataset (CVC-300 refers to 912 images from EndoScene minus 612 images from ColonDB).

[0026] The original stomach images and their annotations are organized in Pascal VOC format (.xml), containing target categories, bounding box coordinates, and image attributes. Since this invention is an improvement on the YOLO11 object detection model, and the network input format is YOLO, the annotation files of the open-source dataset need to be converted from XML format to the .txt format required by YOLO, with the category set to Polpy. During the conversion process, coordinate normalization is performed on each target box, and it is re-encoded according to the category number. The LabelImg tool is used to accurately annotate the remaining images in the expanded dataset, ensuring clear polyp targets and accurate categories.

[0027] After the dataset was labeled, the images were fed into the PrecisePolypNet network. During model training, the improved YOLO11n architecture was used, its default configuration file was loaded, and the backbone network and detection head were jointly trained. The AdamW optimizer was employed, which effectively suppresses overfitting while maintaining training stability, making it particularly suitable for training scenarios with high-resolution images and complex target backgrounds. For the loss function, YOLO's default multi-task joint loss was used, including CIoU bounding box regression loss, target confidence loss, and classification loss, improving the model's prediction accuracy for target location and category.

[0028] The input image size for training was set to 640×640, the total number of training epochs was 200, and the batch size was 64. To improve the model's generalization ability, a Mosaic data augmentation strategy was introduced during training to enhance sample diversity and improve robustness in complex scenarios. The learning rate scheduling strategy adopted cosine annealing for dynamic decay, and a warmup mechanism was enabled in the first 3 epochs to smooth the gradient update process in the early stages of training. For model evaluation, mAP@0.5 and mAP@0.5:0.95 were used as the main performance indicators. Precision and Recall were also used to comprehensively evaluate the model's performance on the validation set to ensure its detection accuracy and generalization ability in small target detection tasks such as gastric polyps.

[0029] S1. Acquire gastroscopy images and perform preprocessing.

[0030] The raw stomach images are acquired during the endoscopic examination. Preprocessing operations such as normalization, noise suppression, color enhancement, and size unification are performed on the raw stomach images to ensure the quality of the input data.

[0031] Specifically, an endoscopy device is used to acquire images of the patient's stomach interior. The images are typically in RGB color format. The image acquisition system supports continuous video frame reading and real-time keyframe capture for static target detection processing. The acquired images are input to an image analysis system via USB or network interface as input for subsequent processing. Size unification refers to scaling images of different sizes to the same dimensions. For example, if one image is 1024*1680 and another is 1200*960, both images are resized to 640*640. The original gastroscopy images are enhanced to improve image features; the image brightness distribution is normalized to avoid loss of polyp features due to insufficient local exposure; the contrast of red and brown areas is enhanced by histogram equalization to make the lesion area more obvious; filtering algorithms (such as median filtering or bilateral filtering) are used to suppress gastric fluid reflection and blurred edges; finally, image sharpening is performed to obtain high-quality polyp edge information.

[0032] S2. Input the preprocessed gastroscopy image into the YOLO11n backbone network for feature extraction to obtain multi-scale feature maps.

[0033] The preprocessed image is input into the YOLO11n backbone network (structured with multiple channel compression bottleneck modules and spatial pyramid pooling modules). The backbone network outputs multi-scale feature maps, including feature representations at three resolutions: P3 (1 / 8), P4 (1 / 16), and P5 (1 / 32).

[0034] S3. For feature map pairs at adjacent scales, input them into the dual-domain collaborative enhancement module (C2PSA_EDFN module) for feature enhancement to generate enhanced multi-scale feature maps.

[0035] Specifically, in the backbone of YOLO11n, a new dual-domain collaborative enhancement module (C2PSA_EDFN module) has been introduced. The dual-domain collaborative enhancement module (C2PSA_EDFN module) is embedded in the high-level feature path of the backbone network, replacing the original convolution stacking structure to enhance the representation ability of high-frequency domain information and improve the accuracy of gastric polyp detection.

[0036] For feature map pairs at adjacent scales (e.g., P5 and P4, P4 and P3), the input is fed into the dual-domain collaborative enhancement module (C2PSA_EDFN module). The dual-domain collaborative enhancement module (C2PSA_EDFN module) combines a boundary-sensitive attention mechanism and a frequency-domain enhanced feedforward network (EDFN). First, it calculates the spatial and channel correlations between features through a multi-head self-attention mechanism. Then, it utilizes the PSA module to achieve horizontal and vertical multi-head attention mapping through depthwise separable convolutions, enhancing the expression of boundary and detail features. This module effectively improves the boundary sensitivity and expressive power of the input features through mechanisms such as channel expansion, attention weighting, frequency-domain enhanced feedforward networks, and residual connections, outputting an enhanced fused feature map. The specific process of the dual-domain collaborative enhancement module (C2PSA_EDFN module) is as follows: The Frequency Domain Enhanced Feedforward Network (EDFN) first uses 1×1 convolutions to double the number of input feature channels to twice the number of hidden channels, then applies depthwise convolutions to enhance local spatial awareness. Next, a two-dimensional Fast Fourier Transform (FFT) is used to transform the segmented feature map to the frequency domain, multiplying it by learnable frequency domain weight parameters to enhance and weight the frequency domain features. Finally, an inverse FFT is used to transform it back to the time domain, and a 1×1 convolution restores the number of channels to the input dimension. This design effectively combines the spatial awareness of time-domain convolutions with the global feature representation capabilities of frequency-domain transformations, enhancing the model's sensitivity to boundary and detailed features.

[0037] It should be noted that the segmented feature maps are segmented using a rearrange operation, dividing the image into multiple patches. Before performing the Fourier transform, it is necessary to ensure that the image size is an integer multiple of the patch size. If not, padding is used to ensure that each patch is complete.

[0038] 2) The multi-head position-sensitive attention mechanism (PSABlock_EDFN) replaces the built-in feedforward neural network with the aforementioned frequency-domain enhanced feedforward network (EDFN). This attention mechanism adaptively focuses on the spatial location and channel information of the input features. Combined with the frequency-domain enhancement mechanism, it enables feature interactions to capture both spatial local details and global spectral information in the frequency domain, greatly improving the richness and discriminative power of feature representation. Simultaneously, a residual connection structure is employed to ensure stable gradient propagation and promote effective training of deep networks.

[0039] 3) A deep feature extraction module is formed by stacking multi-head position-sensitive attention mechanisms (PSABlock_EDFN) layer by layer. The input feature map is first expanded by 1×1 convolution, then input into the multi-layer multi-head position-sensitive attention mechanism (PSABlock_EDFN) for multi-level joint feature enhancement in the time and frequency domains, and finally restored by 1×1 convolution to achieve efficient encoding of input features. It is particularly suitable for processing multi-scale feature maps (such as P3, P4, P5), and can effectively capture and fuse boundary information at different spatial resolutions, thereby improving the performance of the lower gastric polyp detection task.

[0040] It should be noted that the core of the dual-domain collaborative enhancement module (C2PSA_EDFN module) consists of multiple PSABlock_EDFN modules, forming a stacked structure for in-depth feature processing. The PSABlock_EDFN module is inherited from PSABlock and adds an improved frequency domain enhanced feedforward network (EDFN) to the original self-attention mechanism. This frequency domain enhanced feedforward network (EDFN) further enhances the model's expressive power through nonlinear transformations.

[0041] The dual-domain collaborative enhancement module (C2PSA_EDFN module) uses a parameterized frequency domain filter self.fft as a trainable frequency domain weight, enabling the model to autonomously adjust different frequency components to better capture frequency domain feature patterns. At the same time, the multi-head position-sensitive attention mechanism (PSABlock_EDFN), combined with the frequency domain and local processing capabilities of the frequency domain enhancement feedforward network (EDFN), effectively enhances the joint modeling and expression capabilities of local details and global frequency information.

[0042] The dual-domain collaborative enhancement module (C2PSA_EDFN module) includes attention mapping channels in both the horizontal and vertical directions. It utilizes depthwise convolution for spatially selective enhancement and combines this with a sigmoid function to achieve response modulation, thereby effectively allocating attention to the spatial dimensions of the feature map and enhancing the representation of polyp boundaries and morphology. Through these features, the dual-domain collaborative enhancement module (C2PSA_EDFN module) aims to improve detection accuracy and reduce computational complexity.

[0043] S4. After the enhanced multi-scale feature maps are upsampled to the same spatial scale, they are spliced ​​in the channel dimension and further fused through convolution operation to generate unified fused features rich in multi-scale context.

[0044] After upsampling (nn.Upsample) all the enhanced multi-scale feature maps (P3′, P4′, P5′) output by the dual-domain collaborative enhancement module (C2PSA_EDFN module) to the same spatial resolution, torch.cat is used to concatenate the multiple feature maps in the channel dimension (i.e., dim=1). 3x3 convolution is then used on the concatenated feature maps to further fuse information from different scales and generate the final feature representation.

[0045] S5. Input the fused feature map into the detection head module for classification and regression prediction, output the polyp category confidence and precise bounding box coordinates respectively, and use the non-maximum suppression algorithm to filter redundant boxes to obtain the final detection result. Specifically, the detection head module performs bounding box regression and class classification on the fused feature map, and outputs the location coordinates, confidence score and class probability of the polyp; By combining multi-scale prediction with a non-maximum suppression strategy, duplicate detection boxes are effectively removed, and the final detection results are obtained.

[0046] S6. The final test results are overlaid back onto the original gastroscopy image to form a visualization result, and a test report is generated at the same time.

[0047] The detection bounding box and category labels are visually drawn on the gastroscopy image, supporting the export of static images such as JPEG and PNG, and can also be integrated into the endoscopic real-time annotation and feedback system. The final detection results can be stored and generated as video or PDF reports, detailing the location, size, and detection confidence level of the polyps, facilitating subsequent review and diagnosis, and improving the early detection rate and treatment efficiency of gastric polyps. Combined with a cloud platform, the detection data can be uploaded and managed, supporting remote medical services, case comparison, and continuous iteration and optimization training of the model.

[0048] This embodiment also provides a computer device applicable to a visual intelligent detection method for accurate identification of gastric polyps, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the visual intelligent detection method for accurate identification of gastric polyps as proposed in the above embodiment.

[0049] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0050] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the visual intelligent detection method for accurate identification of gastric polyps as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0051] In summary, this invention achieves its goals through: It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A visual intelligent detection method for accurate identification of gastric polyps, characterized in that: The application relates to a method for detecting polyps in a gastroscope image. The gastroscope image is collected and preprocessed; The preprocessed gastroscope image is input into a YOLO11n backbone network for feature extraction to obtain multi-scale feature maps; The feature maps of adjacent scales are input into a dual-domain collaborative enhancement module for feature enhancement to generate enhanced multi-scale feature maps; The enhanced multi-scale feature maps are up-sampled to the same spatial scale, spliced in the channel dimension, and further fused through convolution operation to generate unified and multi-scale context-rich fusion features; The fusion features are input into a detection head module for classification and regression prediction to output the class confidence and accurate bounding box coordinates of the polyps, and a non-maximum suppression algorithm is used to filter redundant boxes to obtain the final detection result; The final detection result is superimposed back to the original gastroscope image to form a visual result, and a detection report is generated.

2. The visual intelligent detection method for precise identification of gastric polyps according to claim 1, characterized in that: The preprocessing includes normalization, noise suppression, color enhancement and size unification. The YOLO11n backbone network comprises multiple groups of channel compression bottleneck modules and spatial pyramid pooling modules.

3. The visual intelligent detection method for precise identification of gastric polyps according to claim 2, characterized in that: The dual-domain collaborative enhancement module comprises a boundary-sensitive attention mechanism and a frequency domain enhancement feedforward network.

4. The visual intelligent detection method for precise identification of gastric polyps according to claim 3, characterized in that: The boundary-sensitive attention mechanism comprises calculating the spatial and channel correlation between features through a multi-head self-attention mechanism.

5. The visual intelligent detection method for precise identification of gastric polyps as claimed in claim 4, wherein: The PSA module realizes multi-head attention mapping in the horizontal and vertical directions through a depth separable convolution to strengthen the expression of boundary and detail features. The frequency domain enhancement feedforward network comprises expanding the feature dimension through a 1x1 convolution, extracting local spatial information through a depth separable convolution, and dividing the input features into fixed-size patches; 6. The visual intelligent detection method for precise identification of gastric polyps as claimed in claim 4, wherein: A two-dimensional fast Fourier transform is performed on each patch to convert the features into the frequency domain, the spectrum is modulated through a learnable frequency domain filter, and an inverse FFT transform is performed to return to the time domain, realizing effective fusion of frequency domain and spatial information. The generation of unified and multi-scale context-rich fusion features comprises splicing the multi-scale feature maps enhanced by the dual-domain collaborative enhancement module in the channel dimension after up-sampling to the same spatial resolution in the spatial dimension, and further fusing through a 3x3 convolution to generate a unified multi-scale fusion feature representation.

7. The visual intelligent detection method for precise identification of gastric polyps as claimed in claim 6, wherein: The detection head module comprises three layers of output corresponding to the P3, P4 and P5 three-scale fusion features for covering small, medium and large polyp detection tasks.

8. The visual intelligent detection method for precise identification of gastric polyps as claimed in claim 7, wherein: The final detection result is obtained by performing bounding box regression and class classification on the fusion feature maps through the detection head module to output the position coordinates, confidence and class probability of the polyps.

9. The visual intelligent detection method for precise identification of gastric polyps according to claim 8, characterized in that: Through multi-scale prediction combined with a non-maximum suppression strategy, repeated detection boxes are effectively removed to obtain the final detection result. The detection report records the position, size and detection confidence of the polyps in detail.

10. The visual intelligent detection method for precise identification of gastric polyps according to claim 9, characterized in that: ​

Images