TWI789932B - Imaging detection method and device for colorectal polyps and system thereof - Google Patents

Abstract

The invention provides an imaging detection method and device for colorectal polyps and system thereof. The imaging detection device for colorectal polyps inputs an image received from an endoscopic device to an inference module to obtain a plurality of location block information and their corresponding probability information, and marks a location block on the image based on the filtered location block information. At the same time, the control pedal can be used to control the switch of the inference module. This convenience imaging detection system for colorectal polyps helps doctors reduce the rate of misdiagnosis.

Description

Colorectal polyp image detection method, device and system

The present invention relates to a colorectal polyp image detection method, device and system, in particular to a colorectal polyp image detection method, device and system using a convolution deep learning model.

In the clinical practice of colonoscopy, the colonoscope is used to check whether there are lesions in the large intestine. At the beginning of the examination, the colonoscope will go deep into the ileocecal valve along the direction of the large intestine. When the colonoscope is pulled out, the doctor will carefully check the intestinal wall. If there are suspicious lesions, the colonoscope will be removed according to the size of the polyp and the degree of cancerization. Alternatively, it can be surgically removed and sent to the pathology department for a more detailed diagnosis. Effective detection and removal of polyps can reduce the risk of colorectal cancer.

At present, the detection of polyps is performed by human eyes by gastroenterologists using a colon endoscope. Traditional endoscopes use xenon gas (Xenon) as a light source, allowing endoscopists to see images similar to white daylight light sources. However, intestinal polyps are not only small and flat when they first grow out, but even have the same color as the surrounding normal tissue. Just like a mung bean falling on the grass, it is difficult to find, and it also makes it easy for beginners in endoscopy to miss the existence of polyps during the inspection process. Studies have shown that 17%-28% of polyps are missed during colonoscopy, which also increases the risk of undetected colorectal cancer (Kim, N. H., Jung, Y. S., Jeong, W. S., Yang, H. J., Park, S. K., Choi, K., & Park, D. I. (2017). Miss rate of colorectal neoplastic polyps and risk factors for missed polyps in consecutive colonoscopies. Intestinal research, 15(3), 411.).

Therefore, the medical field urgently needs a real-time colorectal polyp artificial intelligence image detection system that can be easily carried and used. It can be flexibly matched with various brands of colonoscopy devices, automatically mark the lesion area, and assist physicians or technicians to detect the existence of polyps. Send out a warning sound to reduce the misdiagnosis rate.

The present invention provides an image detection system for large intestine polyps, which includes: an image detection device used in conjunction with an endoscope device and a display device, which includes: an image reading module, which is used to read and receive An image from the endoscope device; an inference module, which reduces the resolution of the image when an inference switch is turned on, and then inputs the image into an inference model to obtain a plurality of location block information and their respective correspondences Probability information, the inference model uses a supervised convolutional deep learning model, which is trained using a Resnet50 classification model with polyp images marked by the method of category activation mapping; and a display module, the probability information is greater than a filter The location block information corresponding to the threshold is subjected to a repeated detection screening, and a location block is marked on the image according to the filtered location block information; and a deduction switch pedal is connected with the image detection device , which sends a first control signal to switch the inference switch.

The present invention also provides a large intestine polyp image detection device, which is used with an endoscope device and a display device, which includes: an image reading module, which is used to read the image received from the endoscope device An image; an inference module, which reduces the resolution of the image and then inputs the image into an inference model to obtain a plurality of location block information and their corresponding probability information when an inference switch is turned on, the inference model A supervised convolutional deep learning model is used, which is trained by using a Resnet50 classification model with class activation mapping method to mark polyp images; and a display module, which corresponds to the probability information greater than a filtering threshold The location block information is subjected to a repeated detection filter, and a location block is marked on the image according to the filtered location block information.

The present invention also provides an image detection method for large intestine polyps, comprising: reading an image received from an endoscope device; when an inference switch is turned on, reducing the resolution of the image and then inputting the image into an inference model In order to obtain a plurality of location block information and their corresponding probability information, the inference model adopts a supervised convolutional deep learning model, which is trained using a Resnet50 classification model and a class activation mapping method to mark polyp images; and The location block information corresponding to the probability information greater than a filtering threshold is subjected to a repetitive detection screening, and a location block is marked on the image according to the filtered location block information.

In some specific embodiments, it further includes: a tone switch pedal, connected to the image detection device, which sends a second control signal to switch a tone switch; and a tone volume pedal, connected to the image The detection device is connected, and it sends a third control signal to switch the volume of a prompt sound; wherein the image detection device further includes a notification module, when the prompt sound switch is on, when the display module marks the location area block, a prompting sound device is triggered to emit a prompting sound according to the volume of the prompting sound.

In some embodiments, the inference model is executed on a GPU, the inference model adopts a Yolov4 architecture, and runs on a TensorRT framework.

In some embodiments, the resolution of the image is reduced to 672 x 352.

In some embodiments, the TensorRT framework uses 16 floating point precision.

In some specific embodiments, the filtering threshold value is 0.15, and the repetition detection screening is to sort the probability information greater than the filtering threshold value in order of size, and successively calculate the corresponding numbers of every two pieces of the probability information according to the sorting. One piece of overlapping information among the location block information, when the overlapping information exceeds a certain value, the location block information corresponding to the smaller probability information among the two location block information is deleted.

In some embodiments, the image reading module reads the image received from the endoscope device by a buffer space capture card.

In some embodiments, the image detection device further includes an input detection module for continuously confirming the first control signal, the second control signal, or the third control signal.

In some embodiments, the image detection device further includes an initial module for setting the prompt sound device, setting a pedal message channel and determining the existence of a program file.

The colorectal polyp image detection system provided by the present invention is not only convenient to carry and move, but also can be flexibly matched with various brands of colonoscope devices. Among them, the colorectal polyp image detection method developed by artificial intelligence can immediately mark the lesion area and assist doctors Or the technician discovers the existence of polyps and sends out a warning sound to reduce the rate of misdiagnosis. At the same time, when the doctor's hands are busy with the colonoscopy scan, he can easily use the foot pedal to adjust whether to use artificial intelligence to assist in the detection of polyps at any time, or to adjust whether to turn on the prompt sound and adjust the volume of the prompt sound.

Other technical contents, features and effects of the present invention will be clearly presented in the following detailed description of preferred embodiments with reference to the drawings.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

The term "installation" disclosed in the description of the present invention is used to describe the combination of structures. It generally refers to multiple structures that cannot be easily separated or dropped after being combined. They can be fixed connections or detachable connections. , It can also be integrally connected; it can be mechanically connected or electrically connected; it can be directly physically connected or indirectly connected through an intermediary, and it can be internal communication between two parts, for example: using threads, Tenon, buckle, nail, adhesive or any combination of high frequency.

The terms "connection", "connection" or "electrical connection" disclosed in the description of the present invention are used to describe the structural combination relationship, which refers to the use of electric wires, circuit boards, connecting wires, network cables, switching devices, and conversion devices. , bluetooth or wireless network for power or network communication.

As used herein, the articles "a", "an" and "any" refer to one or more than one (ie, at least one) of the grammatical items of the item. For example, "an element" means one element or more than one element.

The term "about", "approximately" or "approximately" as used herein essentially means that the stated value or range of values is within 20%, preferably within 10%, and more preferably within 5% float within. Numerical quantities provided herein are approximations and are intended to be inferred if the terms "about", "approximately" or "approximately" are not used.

As shown in Figure 1, a large intestine polyp image detection system 1 according to an embodiment of the present invention includes an image detection device 10 and an inference switch pedal 21, and the image detection device 10 includes: an image reading module 14, a Inference module 15 and a display module 16 . The image detection device 10 may further include a notification module 17 , besides, it may further include an initialization module 11 , an input detection module 12 and an execution module 13 . The large intestine polyp image detection system 1 may further include a plurality of control pedals, for example, the inference switch pedal 21 forms a control pedal group 20 with a prompt tone switch pedal 22 and a prompt volume pedal 33 .

The image detection device 10 adopts a colorectal polyp image detection method to detect the polyp feature in the colonoscope image. The image detection method for large intestine polyps includes: reading an image received from the endoscope device 40; when an inference switch is turned on, reducing the resolution of the image and then inputting the image into an inference model to obtain a plurality of positions Block information and their corresponding probability information, the inference model adopts a supervised convolutional deep learning model, which uses a Resnet50 classification model with a class activation mapping method to mark polyp images for training; and the probability information is greater than The location block information corresponding to a filtering threshold is subjected to a repeated detection filter, and a location block is marked on the image according to the filtered location block information.

Generally, the image of the endoscope device 40 for detecting the large intestine is about 30 frames per second (frame per second, fps). In order to mark the position of the large intestine polyp on the colonoscopy image in real time, the image detection device 10 can be a computing host, which has a main board including a central processing unit and a graphics processing unit, such as using Nvidia's edge computing device Jetson Xavier, It can minimize the size of the hardware, and at the same time, the computing efficiency can also meet the above image processing requirements. Combined with the following model computing and system hardware and software integration, it can achieve the performance of each image processing speed within 33ms. Examples of hardware specifications are shown in Table 1. Table 1 Hardware specification table Hardware Specification graphics processor 512-core Volta GPU with Tensor Cores CPU 8-core ARM v8.2 64-bit CPU, 8MB L2+4MB L3 Memory 32GB 256-bit LPDDR4x | 137GB/s storage device 32GB eMMC 5.1 Deep Learning Accelerator [2x] NVIDIA Engines visual accelerator 7-way VLIW Vision Processor encoder/decoder [2x] 4Kp60 | HEVC/[2x] 4Kp60 | 12-Bit Support

Please also refer to FIG. 2 , the image detection device 10 can be designed as a portable device that can be carried by one hand, for example, the size is about 118.2mm x 135mm x 93.1mm. In order to facilitate the connection to other devices, the appearance is equipped with USB, HDMI, SDI and other connection ports, so that the image detection device 10 can be used with different brands of endoscope devices 40 and display devices 30, and can also be connected to control pedals to control Image detection function. For example, the image detection device 10 is connected to the control pedal set 20 through a USB connection port, connected to a display device 30 through an HDMI connection port, and connected to an endoscope device 40 through an SDI connection port of an image capture card.

For example, the image detection device 10 can be connected to the inference switch pedal 21 via a USB port to control the artificial intelligence colorectal polyp image detection function. When the inference switch pedal 21 is triggered, a first control signal can be sent, and the first control signal can switch the inference switch in turn. The prompt sound switch pedal 22 and the prompt volume pedal 23 can also be connected to the image detection device 10 via the USB port, and the prompt sound switch pedal 22 can trigger a second control signal to switch the prompt sound switch. The prompt volume pedal 23 can trigger a third control signal to switch the volume of the prompt sound (for example: large/medium/small).

As shown in Figure 3, the image detection device 10 can further include an initial module 11, the program of the initial program checks whether the configuration file exists (step S10), if the configuration file exists, then set the prompting sound device (step S11), set The pedal message pipeline (step S12) is used to receive the first control signal to the third control signal, and confirm the existence of the program file (step S13). If it exists, start the full power mode of the system (step S14 ), so that the system runs at the highest clock speed.

The image detection device 10 may further include an input detection module 12 . After the setting of the pedal signal channel is completed, the input detection module 12 will continue to confirm the pedal input (step S15), and if there is a first control signal, a second control signal or a third control signal output from the pedal being triggered, it will be sent to the program . Each pedal control signal is to switch an option. If it is a control pedal of a switch, each trigger is an interactive switch between on and off. If it is a control pedal that controls three options (such as three different volumes), Each trigger is a sequential switch between the three options.

The image detection device 10 may further include an execution module 13, which will initialize the image reading module 14, the inference module 15 and the display module 16, and load the trained image detection inference model of large intestine polyps into the graphics processor.

As shown in FIG. 4 , the image capture card of the image detection device 10 will continuously capture scanning images from the endoscope device 40 at approximately 30 images per second. In order to reduce the delay and achieve a speed that can match the image speed transmitted from the colonoscope device, the system of the embodiment of the present invention selects a capture card with a line buffer space (line buffer). The entire image information is processed in a single line in the capture card. In the common frame buffer space (frame buffer) capture card, it can avoid the delay (about 33ms) due to the need to wait for the entire image capture to complete before starting the subsequent processing.

The image reading module 14 is used to read an image received from the endoscope device 40 , and the example image resolution is about 1920×1080 (image aspect ratio is 16:9). The image reading module 14 will continue to wait for the input image file (step S20), and then confirm the inference on/off (step S21). If the state of the inference switch is on, the image file is sent to the inference module 15 (step S22) for subsequent inference processing. If the state of the inference switch is off, directly add system information to the foreground of the image (step S23), and then output the image (step S24). The system information can be a trademark or system name or pattern, system version information, inference switch mark, prompt tone switch mark or volume mark.

As shown in Figure 5, when the inference switch is turned on, after the inference module 15 receives the image sent by the image reading module 14, it first reduces the image resolution (step S30), in order to satisfy the model of the embodiment of the present invention And the goal of software inference time must be less than 33ms. With the integration of software and hardware in the embodiment of the present invention, at least the resolution needs to be reduced to 704 x 384 or 704x 352, and it is best to reduce it to 672 x 352 to achieve the most suitable inference Time and inference accuracy. In addition, the RGB value of each pixel can also be normalized, that is, each value is divided by 255. Afterwards, the image is input into an inference model (step S31).

TensorRT is a high-performance deep learning inference framework launched by Nvidia. The Yolov4 architecture is optimized by TensoRT to make inferences faster. It improves computing efficiency by simplifying convolutional layers, maintains accuracy and reduces GPU memory usage. Therefore, the trained Yolov4 inference model will be converted into the form of TensorRT, and the floating point number 32 will be reduced to the precision of floating point number (floating point, FP) 16, and then deployed to the hardware. The above-mentioned software-hardware integrated architecture reduces floating-point numbers at the same time, and can achieve the fastest inference ability with a relatively small accuracy loss.

The inference model is trained with approximately 400,000 colonoscopy images, using supervised deep learning. The source of the training image can be a colonoscopy image in DICOM format, or a colonoscopy image converted from a colonoscopy video in mp4 format one per second. Each image is marked with the position of the polyp, and the polyp image is combined with the information of the polyp position to input the inference model for training. Through multiple iterations and parameter optimization, the loss (Loss) is reduced to the minimum value in the area. , to get a trained mature inference model.

Next, explain how to obtain polyp images and polyp location information. There are two ways to mark polyps. One is to use a doctor trained in gastroenterology to mark the location of polyps according to the results of the pathology report. . Another alternative is to only mark whether there are polyps. If there are polyps in the image, the doctor will mark this image as an image with polyps (but there is no need to circle the location of the polyps). The images are divided into polyps After the polyp is removed, it will enter a Resnet50 classification model. The Resnet50 classification model contains 49 layers of convolutional layers, 1 layer of average pooling layer and 1 layer of fully connected layer, with a total of 32 times of dimensionality reduction. Each layer of the model will learn different features. , and finally output the classification probability value of each image with polyp and without polyp. Such alternatives can drastically reduce development time.

As shown in FIG. 6A , it is the original image of the colonoscope. After the classification model, after the classification probability value is output, you can use the method of model visualization Class Activation Mapping (CAM) to find out which range in the picture is judged as a polyp by the classification model, that is, the last convolutional layer Then use the global average pooling (Global Average Pooling, GAP), and then according to different output nodes, multiply all the feature maps before GAP by a set of fully connected layers after GAP, as multiplying different proportions for different feature maps, And superimpose all its outputs together, you can get a CAM result map similar to a heat map (Heatmap).

As shown in FIG. 6B , the color of the heat map detected by the colonoscopy image represents the probability value of the colorectal polyps in the block where the color is located. Afterwards, as shown in Figure 6C, set the area in the heat map with a probability value greater than 0.5, 0.6 or 0.7, more preferably, the area with a probability greater than or equal to 0.7 is set to 1, and other areas are set to 0, so that the edge detection algorithm can be used (find_contour of opencv) Find the contour around the polyp. If the area inside the contour is less than 1/400 of the original image, remove the contour. If it is greater than or equal to 1/400, take out the lower left and upper right positions on the contour (Xmin, Ymin, Xmax, Ymax) form a rectangular frame to frame the detected colorectal polyp area B to output the detection result of the colonoscopy scan. This result can replace the position image of the polyp marked by an expert to save training time.

Before putting the above image and colorectal polyp area information into the inference model training using the Yolov4 object detection framework, the above-mentioned processed multiple images and the marked multiple colorectal polyp area are analyzed by K-means clustering algorithm , to obtain nine sets of anchor box (anchor box) information, as the anchor box parameters used by the subsequent Yolov4 object detection framework for polyp images of the large intestine, nine sets of information are more suitable calculation quantities. Each marked large intestine polyp region contains at least block area (H, W) information. First, nine groups of large intestine polyp regions are randomly selected as group centers, and then the large intestine polyp regions of all images are overlapped with individual group centers ( Intersection over Union, IoU) calculation. IoU=(intersection area of large intestine polyp region and individual cluster centers/union area of large intestine polyp area and individual cluster centers). Determine the large intestine polyp area to the group center with large overlapping information, the new group can calculate the new group center according to the average value of the group in each group, and repeatedly divide the new group according to the overlapping information, and calculate The average value of the group in the new group, until the group hearts do not change, the final nine group hearts are nine sets of anchor box (anchor box) information.

The inference model is a convolutional deep learning model (Convolutional Neural Network, CNN). This model is formed by stacking multiple convolutional layers. It adopts the Yolov4 object detection architecture and uses the Darknet53 backbone (backbone). Darknet53 is composed of 52 convolutional layers stacked, and the fully connected layer is omitted at the end, and the dimensionality reduction (down-sampling) is performed 32 times during the entire convolution process. After neck (neck) processing, using 3 layers of maximum pooling layers, after different times of dimensionality enhancement (up to 2 times), three grid (grid) feature maps of different sizes are generated on the head (head), They are the feature maps of 8 times of dimensionality reduction, 16 times of dimensionality reduction and 32 times of dimensionality reduction. The feature map with dimensionality reduction 8 times is more beneficial for the detection of small objects, and the feature map with dimensionality reduction 32 times is more beneficial for the detection of large objects. Among them, the feature map with dimensionality reduction 8 times will use the smallest three sets of anchor boxes and perform regression prediction at the center point of each grid, and the feature map with dimensionality reduction 16 times will use three sets of anchor boxes of the middle size in each grid. The center point of each grid is used for regression prediction, and the feature map with dimensionality reduction 32 times will use the largest three groups of anchor boxes to perform regression prediction at the center point of each grid.

After each real-time colorectal image is input into the trained and mature inference model, multiple location block information (x, y, w, h) and their corresponding probability information (p) can be obtained through the above calculation. The probability information is Chance that the detected object is a polyp. It can be seen that each image will have at least (number of grids*3)*3 grid group information, and each group of information will have at least five items of data (x, y, w, h, p). Finally, the inference module 15 outputs the image file and a plurality of location block information and their corresponding probability information (step S32 ) to the display module 16 .

As shown in FIG. 7 , the display module 16 can wait for the vertical blanking period (step S41 ) before uploading the image file, so as to avoid misdiagnosis caused by different images appearing on the upper and lower half of the screen at the same time. It is judged whether a polyp is detected or not according to the plurality of location block information transmitted from the inference module 15 and the corresponding probability information (step S42). If no polyp is detected, the system information is added to the foreground of the image (step S43) and then the image is output (step S44).

If a polyp is detected, the location block information corresponding to the probability information greater than a filtering threshold is subjected to a repeated detection screening (step S45 ). In order to find the most suitable frame, for example, a filter threshold value of 0.15 can be set, and all location block information corresponding to probability information greater than 0.15 will be filtered for duplicate detection. Each location block information includes location center position (X, Y) and block area (H, W) information, and the overlap information (IoU) between every two location blocks can be calculated by using the location block information. IoU=(intersection area between two location blocks/union area between two location blocks).

The probability information whose probability information is greater than 0.15 is first sorted from large to small, and the location block corresponding to the maximum probability information is firstly calculated to overlap with the location blocks corresponding to other probability information respectively. If the overlapping information of the two location blocks exceeds 0.5, among the two location blocks, the probability information with the smaller probability information and the corresponding location block information are deleted. Then continue to calculate the overlapping information of the position block corresponding to the second largest probability information in the remaining probability information, and delete it according to the above conditions, and then continue to leave The same operation is performed for the third largest probability information among the probability information, and so on until all remaining location blocks are compared with other location blocks. The last remaining location block information is the filtered location block information. According to the filtered location block information, one of the corresponding location blocks is marked on the original image (step S46), and the system information ( Step S47), finally output the image and the detected polyp position.

The image detection device 10 can further include a notification module 17, and the notification module 17 can transmit its signal to a prompt sound device connected to the motherboard, such as a speaker. If a polyp is detected, confirm whether the prompt tone switch is on or not (step S48), and when the prompt tone switch is on, confirm the volume of the three-tone scale prompt tone (step S49), and the prompt tone device sends the setting according to the prompt volume state. Prompt sound (step S50). When the display module marks the location block, a prompting sound device is triggered to emit a prompting sound according to the volume of the prompting sound.

As shown in Figure 8, during a colonoscopy examination, the colorectal polyp image detection method, device and system of the embodiment of the present invention can assist the physician to detect the presence of polyp in real time, and mark the polyp in the colonoscopy image The location of the meat and a beep sound to inform the doctor. When the doctor scans the colonoscope with both hands, when the doctor wants to confirm the colonoscope image or temporarily observe the original image of the colonoscope, he can use the foot pedal to turn off the automatic detection of large bowel polyps, and then turn on the detection when needed , you can also control the switch and volume of the prompt sound.

As shown in Figure 9, when the inference switch is turned on, the doctor can see the colonoscope image A on the display device 30, and the large intestine polyp region B detected by artificial intelligence; when the inference switch is turned off, the doctor can only see the colonoscope In image A, you need to judge the position of the large intestine polyp by yourself. In addition, the inference switch state C1 , the prompt sound switch state C2 or the prompt volume state C3 can be known at any time through the graphical user interface.

The present invention has been disclosed through the above-mentioned embodiments and various examples. The above are only some preferred embodiments of the invention, but they are not used to limit the present invention. The aforementioned technical features and embodiments, and the equivalent changes or modifications made within the spirit and scope of the present invention, still belong to the scope covered by the present invention, and the scope of patent protection of the present invention depends on the requirements attached to this specification. The one defined in the item shall prevail.

1: Colorectal polyps image detection system 10: Image detection device 11: Initial module 12: Input detection module 13: Execute the module 14: Image reading module 15: Deduction module 16: Display module 17:Notification module 20:Control Pedal Set 21: Inference switch pedal 22: Prompt sound switch pedal 23: Cue Volume Pedal 30: Display device 40: Endoscope device A: Colonoscopy image B: Large intestine polyp area C1: Deduce switch state C2: Prompt sound switch status C3: Prompt volume status

FIG. 1 is a block diagram of a colorectal polyp image detection system according to an embodiment of the present invention.

FIG. 2 is a schematic diagram of an image detection device according to an embodiment of the present invention.

FIG. 3 is a flow chart of the initial module of the embodiment of the present invention.

FIG. 4 is a flowchart of an image reading module according to an embodiment of the present invention.

FIG. 5 is a flowchart of the inference module of the embodiment of the present invention.

6A, 6B, and 6C respectively show the colonoscope image, the colonoscope image detection heat map, and the colonoscope scan detection result map according to the embodiment of the present invention.

FIG. 7 is a flowchart of a display module and a notification module according to an embodiment of the present invention.

Fig. 8 is a schematic diagram of the detection situation of large intestine polyp according to the embodiment of the present invention.

Fig. 9 is a schematic diagram showing the detection of large intestine polyp according to the embodiment of the present invention.

1: Colorectal polyps image detection system

10: Image detection device

11: Initial module

12: Input detection module

13: Execute the module

14: Image reading module

15: Deduction module

16: Display module

17:Notification module

20:Control Pedal Set

21: Inference switch pedal

22: Prompt sound switch pedal

23: Cue Volume Pedal

30: Display device

40: Endoscope device

Claims (10)

An image detection system for large intestinal polyps, comprising: an image detection device, used with an endoscope device and a display device, which includes: an image reading module, which is used to read the An image of the viewing mirror device; an inference module, which reduces the resolution of the image to 672 x 352 when an inference switch is turned on, and then inputs the image into an inference model to obtain a plurality of location block information and their respective relative Corresponding to the probability information, the inference model adopts a supervised convolutional deep learning model, which is trained by using a Resnet50 classification model with polyp images marked by the method of class activation mapping, and the inference model is executed in a graphics processor. The inference model adopts a Yolov4 architecture, and operates on a TensorRT framework, and the TensorRT framework adopts 16 floating-point precision; and a display module, the location block corresponding to the probability information greater than a filter threshold The information is subjected to a repeated detection screening, and a location block is marked on the image according to the filtered location block information; and a deduction switch pedal is connected with the image detection device, which sends a first control signal to switch The inference switch. The detection system as described in claim 1, further comprising: a beep sound switch pedal connected to the image detection device, which sends a second control signal to switch a beep sound switch; and a beep volume pedal, is connected with the image detection device, which sends out a third control signal to switch the volume of a prompt sound; Wherein the image detection device further includes a notification module, when the notification sound switch is on, when the display module marks the location block, a notification sound device is triggered to emit a notification sound according to the volume of the notification sound. The detection system as described in claim 1, wherein the filter threshold is 0.15, and the repeated detection filter is to sort the probability information greater than the filter threshold in order of size, and calculate every two of the probability information successively according to the sorting Corresponding to one of the overlapping information between each two location block information, when the overlapping information exceeds a certain value, the location block information corresponding to the smaller probability information among the two location block information delete. The detection system as described in claim 1, wherein the image reading module reads the image received from the endoscope device by a cache space capture card. The detection system as described in claim 2, wherein the image detection device further includes an input detection module for continuously confirming the first control signal, the second control signal, or the third control signal. The detection system as described in claim 2, wherein the image detection device further includes an initial module for setting the prompt sound device, setting a pedal message channel and determining the existence of a program file. A large intestine polyp image detection device is used with an endoscope device and a display device, which includes: an image reading module, which is used to read an image received from the endoscope device; The inference module, when an inference switch is turned on, reduces the resolution of the image to 672 x 352 and then inputs the image into an inference model to obtain a plurality of location block information and their corresponding probability information, the inference model A supervised convolutional deep learning model was used for training on polyp images labeled with a Resnet50 classification model coupled with class activation mapping. The inference model was implemented on In a graphics processor, the inference model adopts a Yolov4 architecture and operates on a TensorRT framework, and the TensorRT framework adopts 16 floating-point precision; and a display module, the probability information is greater than a filter threshold A repetition detection filter is performed on the corresponding location block information, and a location block is marked on the image according to the filtered location block information. The detection device as described in claim 7, further comprising: a notification module, when a prompt sound switch is on, when the display module marks the location block, trigger a prompt sound device according to a prompt sound volume Beep. An image detection method for large intestinal polyps, comprising: reading an image received from an endoscope device; when an inference switch is turned on, reducing the resolution of the image to 672 x 352 and then inputting the image into an inference model In order to obtain a plurality of location block information and their corresponding probability information, the inference model uses a supervised convolutional deep learning model, which uses a Resnet50 classification model and a class activation mapping method to mark polyp images for training. The inference model is executed in a graphics processor, the inference model adopts a Yolov4 architecture, and operates on a TensorRT framework, and the TensorRT framework adopts 16 floating-point precision; and the probability information is greater than a filtering threshold. A repetition detection filter is performed on the corresponding location block information, and a location block is marked on the image according to the filtered location block information. The detection method as described in claim 9, further comprising: when a prompt sound switch is on, when the location block is marked, triggering a prompt sound device to emit a prompt sound according to a volume of a prompt sound.

Images