Endoscopic image recognition method

Abstract

An endoscopic image recognition method includes converting an original image into a first color space, extracting an image of the first color channel of the original image in the first color space; converting the original image into a second color space, extracting an image of the second color channel of the original image in the second color space. The image of the first color channel and the image of the second color channel are weighted and normalized to generate a normalized probability map of the enhanced target area. Based on the normalized probability map, the probability distribution similarity of adjacent regions of the image is obtained; based on the probability distribution similarity of the adjacent regions, a target area extraction threshold is obtained; and the target area is identified based on the target area extraction threshold. The method of the present invention has high image processing efficiency and can effectively detect bleeding symptoms in smaller areas. The method of the present invention can be applied to the extraction of specific areas of images where the area to be detected is significantly different from the background area. It has a small amount of calculation and steps and is suitable for application in different environments.

Description

Technical Field

[0001] The present invention relates to an endoscope image processing method, and in particular to a capsule endoscope image target area recognition method. Background Art

[0002] With the advancement of computer image processing technology, image processing and analysis are no longer limited to human observation. Computer-assisted screening has become crucial in the medical field. For example, capsule endoscopes are often used to detect bleeding areas in the intestine during intestinal examinations. The capsule endoscope is swallowed by the patient and passes through the esophagus and stomach to the intestines, where it captures and provides images of the intestines. A single intestinal examination using a capsule endoscope can last eight hours. Based on a sampling rate of two images per second, approximately 57,600 images are collected during a single examination. Relying solely on visual screening and manual inspection by a doctor, the image review and screening of all images for a single intestinal examination would be laborious, inefficient, and result in a high rate of diagnostic error.

[0003] Faced with the massive demand for image diagnosis, threshold segmentation methods based on color feature space are often used to detect specific target areas within an image, such as bleeding areas. In practice, the bleeding areas presented in the image to be assessed vary widely. Depending on the patient's specific situation, the patient's intestines may contain large or very small areas of bleeding. Traditional methods for determining small areas of bleeding are often affected by background interference, resulting in inaccurate judgments.

[0004] The present invention aims to provide an improved endoscopic image recognition method to more accurately and effectively detect intestinal bleeding areas through endoscopic images, thereby assisting medical staff in making diagnoses. Summary of the Invention

[0005] The present invention provides an endoscopic image recognition method, comprising: converting an original image into a first color space and extracting an image of a first color channel of the original image in the first color space; converting the original image into a second color space and extracting an image of a second color channel of the original image in the second color space; weighting and normalizing the image of the first color channel and the image of the second color channel to generate a normalized probability map for enhancing a target region; obtaining a probability distribution similarity of adjacent regions of the image based on the normalized probability map; obtaining a target region extraction threshold based on the probability distribution similarity of the adjacent regions; and identifying a target region based on the target region extraction threshold.

[0006] Compared with the existing technology, the method of the present invention has higher image processing efficiency and can effectively detect bleeding symptoms in smaller areas. The method of the present invention is applicable to extracting specific areas of images with large differences between the detection area and the background area. It has low computational complexity and steps, and is suitable for application in different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a flowchart of an endoscopic image recognition method according to an embodiment of the present invention. Detailed description

[0008] It will be appreciated that, in addition to the described example embodiments, the components of the embodiments as generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Accordingly, the following more detailed description of the example embodiments as represented in conjunction with the figures is not intended to limit the scope of the claimed embodiments, but is merely representative of example embodiments.

[0009] Reference throughout this specification to "one embodiment," "another embodiment," or "an embodiment" (or similar descriptions) means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Thus, the appearances of the phrases "in one embodiment" or "in an embodiment," etc., in various places throughout this specification are not necessarily all referring to the same embodiment.

[0010] In addition, the described features, structures or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to provide a thorough understanding of the embodiments. Those skilled in the relevant art will recognize that the various embodiments can be practiced without one or more specific details, or in the case of other methods, components, materials, etc. That is, in other cases, some or all known structures, materials or operations may not be shown or described in detail to avoid confusion.

[0011] The present invention will be described below with reference to the accompanying drawings. Figure 1 FIG. 1 is a schematic diagram of an endoscopic image recognition method according to an embodiment of the present invention, for example, a method for detecting intestinal bleeding areas based on endoscopic images. Figure 1 As shown, the endoscope image recognition method 100 according to this embodiment includes the following steps:

[0012] Step 110: Capture the original image. Step 110 includes using an image acquisition device to capture an image of the subject to be inspected. For example, in the medical field for intestinal examination, step 110 includes using an image acquisition device such as a capsule endoscope to capture an image of the subject's intestines.

[0013] Step 122: Convert the collected original image into a first color space, such as a LAB color space. Specifically, step 122 includes inputting the original image into an image processing system, and outputting an image in the LAB color space after image processing.

[0014] Step 124: Extracting an image of a first color channel of the LAB color space, for example, an A channel image of the LAB color space. Specifically, step 124 includes inputting the image of the LAB color space into the image processing system, and after image processing, outputting the A channel image of the LAB color space from the image processing system.

[0015] Step 132: Convert the captured original image to a second color space, such as a CMYK color space. Specifically, step 132 includes inputting the original image into the image processing system, and outputting an image in the CMYK color space from the image processing system after image processing.

[0016] Step 134: Extracting a second color channel image of the CMYK color space, such as an image of the M channel. Specifically, step 134 includes inputting an image of the CMYK color space into the image processing system, and outputting an image of the M channel of the CMYK color space from the image processing system after image processing.

[0017] Step 140: Weighted normalization is performed on the image of the A channel of the LAB color space and the image of the M channel of the CMYK color space to generate a normalized probability map. For example, assuming that A(x) is the value corresponding to pixel x in the A channel of the LAB color space (LAB-A), and M(x) is the value corresponding to pixel x in the M channel of the CMYK color space (CMYK-M), the two images are first normalized separately and then weighted summed according to the following formula:

[0018] F(x) = wA(x) + (1-w)M(x) (1)

[0019] Where F(x) is the value corresponding to pixel x in the normalized probability map. w is the preset weight, 0 <w<1。

[0020] In this embodiment, the image of the A channel of the LAB color space and the image of the M channel of the CMYK color space are weighted and normalized according to a weight ratio of 0.5 to 0.5 to generate a normalized probability map. It will be appreciated by those skilled in the art that other weights may also be used for weighted normalization. For example, the image of the A channel of the LAB color space and the image of the M channel of the CMYK color space are weighted and normalized according to a weight ratio of 0.4 to 0.6. Alternatively, the image of the A channel of the LAB color space and the image of the M channel of the CMYK color space are weighted and normalized according to a weight ratio of 0.3 to 0.7, and so on.

[0021] Step 150: Obtaining the probability distribution similarity of adjacent regions of the image based on the normalized probability map.

[0022] Step 160: Based on the similarity of probability distribution of adjacent regions in the image, obtain the optimal extraction threshold of the target region, for example, the optimal extraction threshold of the bleeding region displayed in the image.

[0023] Step 170: Determine the target region based on the optimal extraction threshold of the target region, for example, determine the bleeding region based on the optimal extraction threshold of the bleeding region. Specifically, step 170 includes: determining the region in the normalized probability map that is greater than the optimal extraction threshold as the bleeding region.

[0024] According to a preferred embodiment, step 150 obtains the probability distribution similarity of adjacent regions of the image based on the normalized probability map, including performing region growing threshold segmentation on the normalized probability map using multiple selected threshold parameters to obtain the probability distribution similarity of adjacent regions of the image.

[0025] According to a preferred embodiment, step 150 obtains the probability distribution similarity of adjacent regions of the image based on the normalized probability map, performs region growing threshold segmentation on the normalized probability map using multiple selected threshold parameters, and further includes obtaining the probability distribution similarity of adjacent regions according to the following formula:

[0026]

[0027] where p n 、p n-1 They represent the nth and n-1th thresholds respectively, and the threshold set is P = {p1, p2, ..., p n-1 ,p n ,…,p N}, each threshold is in the range of 0 to 1, and the threshold values decrease in sequence, that is, p1>p2,>…,>p n-1 >p n >…,>p N . H(p n ,i) indicates that p n The value of the i-th channel in the probability histogram corresponding to the image area obtained for the segmentation threshold, Where |·| represents the total number of elements in the calculation set, Rn is the number of elements in p n is the pixel set obtained by the segmentation threshold, R n ={x|Pr(x)>p n}, where x is a pixel in the image, and Pr(x) represents the value of the pixel x in the normalized probability image obtained in step 150. n,i ={x|x∈R n,i·s>Pr(x)>(i-1)s}, where s is the width of each histogram channel set when calculating the probability histogram, and s=1 / K.

[0028] In this preferred embodiment, N, p1 and p N are set to 5, Prm-0.1 and Prm-0.3 respectively, where Prm represents the maximum value of all pixel positions in the current normalized probability map, p1 and p N The threshold values between them are equally spaced; in this preferred solution, K is set to 20.

[0029] It will be understood by those skilled in the art that the various parameter values set in this preferred solution (such as N, p1 and p N The settings of K and K are only exemplary. Those skilled in the art can appropriately adjust the parameter values according to actual needs.

[0030] According to a preferred solution, step 160 obtains the optimal extraction threshold of the target region based on the similarity of probability distribution of adjacent regions of the image, including the following steps:

[0031] If the probability distribution similarity S2 between adjacent regions is less than the threshold t0, then p2 is selected as the optimal extraction threshold;

[0032] If the probability distribution similarity S2 between adjacent regions is greater than or equal to the preset threshold t0, then continue to check the probability distribution similarity of subsequent adjacent regions. If Sk>Sk-1 or the absolute value of the difference between the two is greater than the first threshold t1, then p k was selected as the optimal extraction threshold;

[0033] If none of the above conditions are met, the threshold p N In this embodiment, t0 and t1 are set to 0.4 and 0.07 respectively.

[0034] Those skilled in the art will appreciate that the various parameter values set in this preferred solution (such as the setting of t0 and t1) are merely exemplary, and those skilled in the art may appropriately adjust the parameter values according to actual needs.

[0035] As quoted herein, unless expressly stated otherwise, the singular "a" and "an" may be construed to include the plural "one or more".

[0036] The present invention has been presented for purposes of illustration and description, but is not intended to be exhaustive or limiting. Many modifications and variations will be apparent to those skilled in the art. The exemplary embodiments have been chosen and described in order to explain the principles and practical applications, and to enable those skilled in the art to understand that the various embodiments of the present disclosure may be used with various modifications as are suitable for the particular application contemplated.

[0037] Therefore, although illustrative example embodiments have been described herein with reference to the accompanying drawings, it should be understood that the description is not restrictive and that, therefore, a person skilled in the art may implement various changes, modifications, substitutions or reorganizations without departing from the scope of the present disclosure and the technical solutions provided.

Claims

1. An endoscopic image recognition method, characterized in that: The method comprises: Converting the original image to a first color space, wherein the first color space is a LAB color space; Extracting an image of the original image in a first color channel of the first color space, where the first color channel is an A channel; Converting the original image to a second color space, wherein the second color space is a CMYK color space; Extracting an image of a second color channel of the original image in the second color space, where the second color channel is the M channel; Weighting and normalizing the image of the first color channel and the image of the second color channel to generate a normalized probability map of the enhanced target area; Obtaining the probability distribution similarity of adjacent regions of the image based on the normalized probability image includes: The normalized probability map is segmented by region growing threshold using multiple selected threshold parameters to obtain the similarity of probability distribution of adjacent regions of the image; and the similarity of probability distribution of adjacent regions is obtained based on the following formula: in, S n is the probability distribution similarity between adjacent regions, p n 、p n-1 They represent the nth and n-1th thresholds respectively. Assume that the threshold set is P = {p1, p2, ..., p n-1 ,p n ,…,p N }, each threshold is in the range of 0 to 1, and the threshold values decrease in sequence, that is, p1>p2,>…,>p n-1 >p n >…,>p N , H(p n ,i) indicates that p n The value of the i-th channel in the probability histogram corresponding to the image area obtained for the segmentation threshold: Among them, |·| represents the total number of elements in the calculation set, R n For p n The set of pixels obtained for the segmentation threshold: R n ={x|Pr(x)>p n , Where x is a pixel in the image, Pr(x) represents the value of the pixel x in the normalized probability map obtained in step (2), and the expression H(p n ,i) in X n,i Limited to: X n,i ={x|x∈R n ,i s>Pr(x)>(i-1)s} Where s is the width of each channel of the histogram set when calculating the probability histogram, s = 1 / K; The optimal extraction threshold of the target region is obtained based on the probability distribution similarity of the adjacent regions, including: if S2 is less than the threshold t0, then p2 is selected as the optimal extraction threshold; if S2 is greater than or equal to the threshold t0, then the subsequent probability distribution similarity of the adjacent regions is continuously checked, and if S k >S k-1 Or the absolute value of the difference between the two is greater than t1, then select p k is the optimal extraction threshold; if none of the above conditions are met, then p is selected N is the optimal extraction threshold; and A target region is identified based on the target region extraction threshold.

Images