Image recognition-based radiotherapy body surface cross mark detection method

Abstract

This invention discloses a method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition. Belonging to the field of image recognition, it constructs a detection model comprising a pre-localization module for the region of interest (ROI), a cross-projection localization module, and a cross-center structure correction module. First, it outputs an ROI response sub-map using multi-scale initial response features. Second, it constructs a cross-direction balanced response value based on the ROI response sub-map to output a preliminary cross-center localization result. Finally, it constructs a dual-inversion coupled response based on the preliminary localization result to output the final cross-center localization result, thereby achieving automatic detection of cross-shaped marks on the surface of a radiotherapy body.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a method for detecting cross-shaped marks on the surface of radiotherapy bodies based on image recognition. Background Technology

[0002] In modern radiotherapy, surface cross markers are important markers for patient positioning, treatment area localization, and center alignment during radiotherapy. Their accuracy directly affects the accuracy of target localization and dose delivery consistency. Current methods for identifying surface cross markers in radiotherapy typically rely on manual observation, manual recording, and semi-automatic image processing based on fixed thresholds. Manual methods suffer from low efficiency, high subjectivity, and poor repeatability. Fixed thresholds or simple template matching methods are easily affected by changes in lighting, skin texture, surface reflection, differences in marker thickness, and changes in shooting angle, leading to missed detections, false detections, or center positioning deviations of the cross markers.

[0003] Several existing patents exist for the detection of radiotherapy surface cross marks based on image recognition. Patent publication number CN120747230A discloses an absolute linear position detection system and method based on image recognition. This system enhances image quality through multi-scale filtering, extracts positional encoding information features from the enhanced image, obtains absolute positional encoding values ​​through encoding-decoding relationship mapping, and finally outputs high-precision linear positional information through relationship calibration. Another patent publication number CN113393458A discloses a hand wound detection method based on a wound weighted saliency algorithm. This method uses GrabCut foreground extraction, calculates region distances in the Lab color space, and introduces wound color feature weighting and a moving visual focus strategy to effectively highlight and detect the location of hand wounds.

[0004] Existing detection methods typically only output the target position in the image plane without further establishing the correspondence between the target pixel coordinates and the physical coordinates of the body surface. This makes it difficult to directly use the detection results for radiotherapy positioning correction, isocenter alignment, and body surface geometric error compensation. Therefore, there is an urgent need for a radiotherapy body surface cross mark detection method that can stably locate cross marks in complex body surface images and map the detection results to body surface spatial coordinates. Summary of the Invention

[0005] This invention provides a radiotherapy surface cross-marker detection method based on image recognition. Addressing the problems of uneven grayscale distribution, complex backgrounds, and insufficient center localization accuracy in existing radiotherapy surface cross-marker detection methods, this invention proposes a radiotherapy surface cross-marker detection model consisting of a surface ROI pre-localization module, a cross-projection localization module, and a cross-center structure correction module. This model first uses a standardized surface cross-marker image to quickly locate candidate cross-marker regions by constructing multi-scale initial response features, outputting an ROI response sub-map. Second, it constructs a cross-direction balanced response value based on the ROI response sub-map and outputs a preliminary cross-center localization result. Finally, it constructs a dual-inversion coupled response based on the preliminary cross-center localization result and performs structural suppression correction to obtain the final cross-center localization result.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps: S1. Collect images of cross-shaped markings on the body surface in the radiotherapy room and construct a standardized dataset of cross-shaped marking images on the body surface; S2. Based on standardized body surface cross-marker images, fast localization of candidate regions for body surface cross-markers is achieved by constructing multi-scale initial response features, and ROI response sub-maps are output. S3. Construct balanced response values ​​for the cross directions based on the ROI response subgraph and output preliminary cross center location results; S4. Based on the preliminary cross center positioning results, construct the dual inversion coupled response and perform structural suppression correction to obtain the final cross center positioning results; S5, together with steps S2, S3 and S4, jointly constructs a cross-shaped marker detection model for radiotherapy body surface; S6. Based on the radiotherapy surface cross mark detection model, process the standardized surface cross mark image and output the final cross center localization result.

[0007] Preferably, in step S1, images of cross-shaped markings on the body surface are acquired in the radiotherapy room. These images include cross-shaped markings on the patient's body surface. The acquired cross-shaped marking images are then subjected to size standardization processing to construct a standardized cross-shaped marking image dataset. This provides a unified data foundation for subsequent ROI pre-location, preliminary detection of the cross center, and correction of the cross center structure.

[0008] Preferably, in the scenario of detecting cross markers on the body surface during radiotherapy, the patient's body surface image is easily affected by skin texture, ambient lighting, and shooting angle. If detection is performed directly based on the gray level of a single pixel, false detections are likely to occur. Therefore, by aggregating gray levels in the local neighborhood, the subsequent gray level potential analysis focuses more on the local deviation features of the cross markers relative to the surrounding body surface area. Secondly, in the detection of cross markers on the body surface during radiotherapy, the cross markers usually have more obvious linear gray level changes relative to the skin background. Multi-scale initial response features can highlight such local differences and reduce the influence of different patient skin colors, different shooting brightness, and local surface reflections on the ROI pre-location results.

[0009] Preferably, in step S2, based on the standardized body surface cross-marker image, the candidate region of the body surface cross-marker is quickly located by constructing multi-scale initial response features, and the ROI response sub-map is output. Based on the standardized cross-shaped image of the body surface, the gray values ​​in the local neighborhood set around the current pixel are aggregated and calculated through local neighborhood gray-level aggregation to generate a local neighborhood average result, so that the subsequent gray-level potential difference analysis pays more attention to the local deviation characteristics of the cross mark relative to the surrounding body surface area.

[0010] Preferably, a gray-level offset relationship is constructed based on the gray-level potential difference between the standardized cross-marked image on the body surface and the local neighborhood average result, and a gray-level potential difference value is constructed by combining it with a stability constant for constraint processing.

[0011] Preferably, a local deviation relationship between the gray potential difference value and its neighborhood average result at different scales is constructed, and the local deviation relationship is enhanced by combining the deviation enhancement index. The enhancement result is combined with the scale weight coefficient to jointly construct multi-scale initial response features.

[0012] Preferably, based on multi-scale initial response features and combined with the ROI response determination threshold, low-response background areas are suppressed, while high-response areas where the cross marks on the body surface are retained, thus generating ROI response sub-maps.

[0013] Furthermore, by constructing a local grayscale reference benchmark around the current pixel, the impact of isolated noise points, local reflective points, and single-pixel grayscale abrupt changes on subsequent judgments can be weakened. By calculating the relative deviation of the current pixel from the average result of its local neighborhood, absolute grayscale changes can be converted into local relative difference expressions. Through grayscale potential difference deviation enhancement under multi-scale neighborhood, the response expressions of fine edges of cross markers, main lines, and slightly blurred areas can be taken into account simultaneously, avoiding insufficient adaptability of a single scale to a certain type of marker morphology and improving the robustness of pre-localization of surface ROIs. The generation of ROI response sub-images using intensity response method enhances the stability and anti-interference ability of ROI localization. In summary, step S2 constructs a pre-localization module for surface ROIs, providing a stable candidate region input and background interference suppression basis for the radiotherapy surface cross marker detection model.

[0014] Preferably, after obtaining the ROI response sub-map through the ROI pre-positioning module on the body surface, although the interference of the background on the body surface has been weakened, the ROI response sub-map still mainly reflects the intensity of local grayscale deviation and cannot directly determine the center of the cross intersection. Therefore, it is necessary to construct a cross projection positioning module to extract the projection response from the horizontal and vertical directions and determine the preliminary cross center positioning result through the balance of the responses in the two directions.

[0015] Preferably, in step S3, a cross-direction balanced response value is constructed based on the ROI response subgraph, and a preliminary cross center positioning result is output; Based on the ROI response subgraph, the horizontal and vertical projection responses are extracted respectively, and the horizontal and vertical projection response values ​​are generated.

[0016] Preferably, based on the relative balance of the projected response values ​​in the horizontal and vertical directions, a cross-equalization enhancement index is introduced for nonlinear modulation, and then synergistically fused with the average intensity of the projected response values ​​in the two directions to generate a cross-equalization response value.

[0017] Preferably, based on the balanced response value of the cross direction, the candidate cross center point set is determined by adjusting the candidate threshold coefficient and combining it with the local maximum constraint, and the preliminary cross center positioning result is further determined by combining the candidate cross center point set.

[0018] Furthermore, by performing horizontal and vertical projections within the ROI response sub-map, the structural features of the cross markers' intersection can be utilized to reduce redundancy in the full-image search and interference from non-target textures. By combining the horizontal and vertical responses and their balance, the true cross intersection area can be highlighted, suppressing single linear interference. By adjusting the candidate threshold coefficient and constraining local maxima, false detections and localization offsets can be reduced. In summary, step S3 constructs a cross-projection localization module, improving the accuracy and stability of the initial cross center localization.

[0019] Preferably, after obtaining the preliminary cross center positioning result through the cross projection positioning module, although the candidate intersection position has been determined based on the horizontal and vertical responses, the preliminary positioning result may still have center offset or adjacent pseudo-center response due to the influence of surface undulations, uneven thickness of marker lines, local shadows and residual linear interference. Therefore, it is necessary to further construct a cross center structure correction module to suppress pseudo-response by using the double inversion coupling response relationship around the preliminary positioning result, and to perform structural correction on the center position to obtain the final cross center positioning result.

[0020] Preferably, in step S4, a dual inversion coupled response is constructed based on the preliminary cross center localization result and structural suppression correction is performed to obtain the final cross center localization result; The dual inversion coupled response is constructed based on the preliminary cross-center positioning results.

[0021] Preferably, based on the dual inversion coupling response, the preliminary cross center positioning result is subjected to structural suppression correction to output the final cross center positioning result.

[0022] Furthermore, by introducing a dual inversion coupling mechanism, the positioning error caused by structural deviation and inconsistency in strength can be characterized simultaneously, improving the ability to identify surface texture, patch edges, and shadow interference. Based on the dual inversion coupling response, structural suppression correction is applied to the preliminary cross center positioning result, which can adaptively weaken the abnormal response area, making the final cross center positioning result closer to the real cross intersection position. In summary, step S4 constructs a cross center structural correction module, which improves the accuracy and anti-interference ability of the final positioning result.

[0023] Preferably, in step S5, steps S2, S3 and S4 are combined to construct a cross-shaped marker detection model for radiotherapy body surface. The standardized body surface cross-marked image is used to construct the body surface ROI pre-location module through step S2, and outputs the ROI response sub-map.

[0024] Preferably, the ROI response subgraph outputs preliminary cross center positioning results through the cross projection positioning module constructed in step S3.

[0025] Preferably, the preliminary cross center positioning result is used to output the final cross center positioning result through the cross center structure correction module constructed in step S4.

[0026] Furthermore, the surface ROI pre-location module suppresses low-response background, the cross-projection positioning module enhances the salience of the cross center, and the cross center structure correction module eliminates structural interference by using dual inversion coupling response. The radiotherapy surface cross mark detection model constructed by the three modules in sequence has both accuracy and efficiency, and effectively realizes the automated detection and positioning of radiotherapy surface cross marks. Attached Figure Description

[0027] Figure 1 This is a step-by-step diagram of a radiotherapy surface cross mark detection method based on image recognition.

[0028] Figure 2 This is a structural diagram of a cross-shaped marker detection model for radiotherapy.

[0029] Figure 3 A structural diagram of the pre-positioning module for the ROI on the body surface.

[0030] Figure 4 This is a structural diagram of the cross-projection positioning module.

[0031] Figure 5 This is a structural diagram of the cross-shaped center structure correction module.

[0032] Figure 6 Input the standardized cross-mark image of the body surface into the detection result image of the radiotherapy body surface cross-mark detection model. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Other implementation methods obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.

[0034] Please see the appendix Figure 1 To be continued Figure 6 This invention provides a method for detecting radiotherapy surface cross-shaped markers based on image recognition. The method constructs a radiotherapy surface cross-shaped marker detection model, which consists of a surface ROI pre-location module, a cross-projection localization module, and a cross-center structure correction module. The surface ROI pre-location module inputs a standardized surface cross-shaped marker image, constructs multi-scale initial response features to quickly locate candidate regions of the surface cross-shaped markers, and outputs an ROI response sub-map. The cross-projection localization module inputs the ROI response sub-map, constructs a cross-direction balanced response value, and outputs a preliminary cross-center localization result. The cross-center structure correction module inputs the preliminary cross-center localization result, constructs a dual-inversion coupling response, and performs structural suppression correction to obtain the final cross-center localization result. The specific process is as follows: Figure 1 As shown.

[0035] Construct a cross-shaped marker detection model for radiotherapy body surface, the model structure is as follows: Figure 2 As shown.

[0036] S1. Collect images of cross-shaped markings on the body surface in the radiotherapy room and construct a standardized dataset of cross-shaped marking images on the body surface.

[0037] Furthermore, a total of 1,000 images of cross-shaped markings on the patient's body surface were collected in the radiotherapy room. These images included the patient's body surface area and the cross-shaped markings located on the body surface area. The collected images of the cross-shaped markings were standardized to a size of 224×224, thus constructing a standardized dataset of cross-shaped marking images. This provides a unified data foundation for subsequent ROI pre-location, cross center detection, and body surface coordinate mapping.

[0038] S2. Based on the standardized body surface cross-marked image, the candidate region of the body surface cross-marked is quickly located by constructing multi-scale initial response features, and the ROI response sub-map is output.

[0039] Furthermore, the structure diagram of the surface ROI pre-location module is as follows: Figure 3 As shown, based on a standardized cross-shaped image of the body surface, the gray values ​​within the local neighborhood set surrounding the current pixel are aggregated and calculated through local neighborhood gray-level aggregation to generate a local neighborhood average result. The mathematical model is as follows: ; in, Indicates pixel position The local neighborhood average result at that location, This represents the row index of the current pixel in the normalized crosshair-marked image. This represents the column index of the current pixel in the normalized cross-shaped image. Represents the row index of a pixel in the local neighborhood set. This represents the column index of a pixel in the local neighborhood set. Indicates the current pixel A local neighborhood set centered on; This represents the number of pixels in the local neighborhood set. This indicates the position of a normalized surface crosshair image in its neighborhood pixels. grayscale value at that location This indicates local neighborhood grayscale aggregation.

[0040] Furthermore, based on the gray-level difference between the standardized cross-marked image and the local neighborhood average, a gray-level shift relationship is constructed, and constrained by a stability constant to construct the gray-level potential difference value. The mathematical model is as follows: ; in, Represents a grayscale potential difference map. This indicates the grayscale potential difference map at the pixel location. The gray potential difference value at that location, This indicates the current pixel position in the standardized crosshair image. The grayscale value at that location; This represents the stability constant, which takes the value of [value missing] in this embodiment. .

[0041] Furthermore, local deviation relationships between gray-level potential difference values ​​and their neighborhood averages at different scales are constructed, and these local deviation relationships are enhanced using a deviation enhancement index. The enhanced results, combined with scale weighting coefficients, are used to jointly construct multi-scale initial response features. The mathematical model is as follows: ; in, Indicates the first Multi-scale initial response feature maps at various scales Indicates the first Multi-scale initial response feature maps at various scales at pixel locations Multi-scale initial response characteristics at the location; This represents the scale number, which in this embodiment is taken as a value of , , ; Indicates the first Pixel position at each scale The gray potential difference value at the location; Indicates the first The neighborhood average result at each scale; Indicates the first The local neighborhood set corresponding to each scale; This represents the deviation enhancement index, which in this embodiment takes the value of ; Indicates the first The scale weight coefficients corresponding to each scale; In this embodiment, , , Corresponding to , , The neighborhood scale, The neighborhood is used to capture fine-grained grayscale changes at the edges of the crosshair marker. The neighborhood is used to accommodate the main line width range of the crosshairs in conventional radiotherapy surface images. The neighborhood is used to enhance the perception of slightly thicker marking lines, slightly blurred markings, and locally diffused edges. The combination of the three scales can cover three common situations: thin lines, normal line width, and slightly diffused line width, which are suitable for the needs of cross marking detection on the body surface under different patients, different marking thicknesses, and different shooting distances. In this embodiment, , , The values ​​are respectively , and Among these considerations, the crosshairs on the body surface used in conventional radiotherapy typically have a certain line width at standard acquisition distances. The scale corresponds to a slightly higher weight, making it easier to form a continuous and stable subject response; The scale is used to preserve edge details and is set to [scale value]. ; The scale is used to improve tolerance in cases of blurriness and thick lines, but setting it too high will introduce more background texture. ; In this embodiment, Pick ,when Greater than At the same time, it can appropriately amplify the local deviation differences between the cross-marked area and the surrounding body surface area; simultaneously, The exponential value is not too large, which avoids excessive amplification of noise in non-cross-shaped areas of skin texture and shadow edges, thus balancing cross-marker response enhancement and false detection suppression.

[0042] Furthermore, based on multi-scale initial response features and combined with the ROI response determination threshold, low-response background regions are suppressed, while high-response regions containing the cross marks on the body surface are preserved, generating ROI response sub-maps. The mathematical model is as follows: ; in, This represents the ROI response subplot. This indicates the pixel location of the ROI response submap. The response value at that location, This represents a response gating function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise. This represents the threshold for determining the ROI response, which is adaptively determined by the overall response distribution of the ROI response subplot. The mathematical model for the composite initial response value obtained by fusing the initial response features from three scales according to scale weight coefficients is as follows: ; in, This indicates that the response intensity of the same pixel location at three scales is accumulated to comprehensively reflect the response at that location at fine, regular, and large scales.

[0043] S3. Construct balanced response values ​​for the cross directions based on the ROI response subgraph and output the preliminary cross center location results.

[0044] Furthermore, the structural diagram of the cross-projection positioning module is as follows: Figure 4 As shown, the projected responses in the horizontal and vertical directions are extracted based on the ROI response subgraph, generating horizontal and vertical projected response values. The mathematical model is as follows: ; ; in, Indicates the local coordinate position of the ROI The horizontal projection response value at the location; Indicates the local coordinate position of the ROI The vertical projection response value at the location; This represents the horizontal projection response diagram; This represents the vertical projection response diagram; In the ROI response subplot, the first... line, number The response value at the column position; In the ROI response subplot, the first... line, number The response value at the column position; Indicates the row index within the vertical projection window; Indicates the column index within the horizontal projection window; This represents the half-width of the projection window, which in this embodiment is taken as a value. When the projection window extends beyond the ROI boundary, summation is performed only within the valid ROI area. In this embodiment, This represents the half-width of the projection window, which in this embodiment is taken as a value. The corresponding length is The directional projection window of one pixel can cover the main line width range of the radiotherapy surface crosshair at the conventional acquisition distance, while retaining a certain edge tolerance.

[0045] Furthermore, based on the relative balance of the projected response values ​​in the horizontal and vertical directions, a cross-equalization enhancement index is introduced for nonlinear modulation, and then synergistically fused with the average intensity of the projected response values ​​in both directions to generate a cross-directional equalization response value. The mathematical model is as follows: ; in, This represents the cross-directional equilibrium response diagram. Indicates the local coordinate position of the ROI Equalized response value at the cross direction; This indicates the operation of finding the minimum value; This indicates the operation of retrieving the maximum value; This represents the cross-response stability constant, which in this embodiment takes the value of ; This represents the cross-balance enhancement index, which in this embodiment takes the value of When both the horizontal and vertical responses are strong and relatively balanced, It can further enhance the response at the intersection center position; when there are only unidirectional lines, the ratio of the minimum response to the maximum response is small, and the response is suppressed after exponential enhancement, thereby reducing the possibility of false detection caused by non-cross-shaped structures.

[0046] Furthermore, based on the balanced response values ​​in the cross directions, the set of candidate cross center points is determined by adjusting the candidate threshold coefficients and combining them with local maximum constraints. The mathematical model is as follows: ; in, Represents the set of candidate cross center points; This represents the mean of all response values ​​in the cross-directional equilibrium response diagram. Represents the cross-directional equilibrium response diagram The standard deviation of all response values; This represents the candidate threshold adjustment coefficient, which takes the value of [value] in this embodiment. ; In this embodiment, Pick After ROI constraint and cross-balancing enhancement, the true cross center usually has a higher-than-average response within the ROI, using mean addition. Using a standard deviation multiple of 1 / 2 can effectively exclude weak response regions and most non-central regions, while avoiding the loss of low-contrast crosshair center points due to excessively high thresholds; combined with the candidate crosshair center point set, the preliminary crosshair center localization result is determined, and the mathematical model is as follows: ; in, This indicates the preliminary results of the cross center location; This means selecting the position with the largest cross-direction equilibrium response value from the candidate point set as the initial cross center.

[0047] S4. Based on the preliminary cross center positioning results, construct the dual inversion coupled response and perform structural suppression correction to obtain the final cross center positioning results.

[0048] Furthermore, the structural diagram of the cross-shaped center structure correction module is as follows: Figure 5 As shown, a dual inversion coupled response is constructed based on the preliminary cross-center positioning results. The mathematical model is as follows: ; in, This represents the dual inversion coupled response. Represents a non-linear activation function. Indicates the kernel size as Convolution operation, Indicates the kernel size as The convolution operation.

[0049] Furthermore, based on the aforementioned dual inversion coupling response, structural suppression correction is applied to the preliminary cross center localization result to output the final cross center localization result. The mathematical model is as follows: ; in, This indicates the final crosshair center positioning result. This represents the Sigmoid mapping function. This indicates element-wise multiplication.

[0050] Steps S5, S2, S3, and S4 are combined to construct a cross-shaped marker detection model for radiotherapy body surface.

[0051] Furthermore, the structural diagram of the cross-marker detection model on the body surface is as follows: Figure 2 As shown, the standardized body surface cross-marked image, through the body surface ROI pre-location module constructed in step S2, outputs an ROI response submap. The mathematical model is as follows: ; in, Represented as a standardized cross-shaped image of the body surface. This indicates the pre-location module for the ROI on the body surface. This represents the ROI response subplot.

[0052] Furthermore, the ROI response submap, through the cross-projection localization module constructed in step S3, outputs preliminary cross-center localization results. The mathematical model is as follows: ; in, This indicates the cross-projection positioning module. This indicates the preliminary results of the cross center positioning.

[0053] Furthermore, the preliminary crosshair center localization result is used by the crosshair center structure correction module constructed in step S4 to output the final crosshair center localization result. The mathematical model is as follows: ; in, This indicates the cross-shaped center structure correction module. This indicates the final crosshair center positioning result.

[0054] S6. Based on the radiotherapy surface cross mark detection model, process the standardized surface cross mark image and output the final cross center localization result.

[0055] Furthermore, as shown in the appendix Figure 1 The radiotherapy surface cross-mark detection model described in S6 requires an input image of the surface cross-mark, as shown in the attached image. Figure 6 As shown in (a), the output shows the location of the cross center on the body surface during radiotherapy, as attached. Figure 6 As shown in (b).

[0056] Furthermore, the operating system platform used to train and load the radiotherapy surface cross mark detection model was CentOS, the language was Python 3.9.2, the processor used was Jetson Xavier, the image processing library used was OpenCV, the server physical memory was 64G, Adam was selected as the optimizer during training, the initial learning rate was set to 0.001, and the batch size was 4.

Claims

1. A method for detecting cross-shaped marks on the surface of a radiotherapy patient based on image recognition, characterized in that, Includes the following steps: S1. Collect images of cross-shaped markings on the body surface in the radiotherapy room and construct a standardized dataset of cross-shaped marking images on the body surface; S2. Based on standardized body surface cross-marker images, fast localization of candidate regions for body surface cross-markers is achieved by constructing multi-scale initial response features, and ROI response sub-maps are output. S3. Construct balanced response values ​​for the cross directions based on the ROI response subgraph and output preliminary cross center location results; S4. Based on the preliminary cross center positioning results, construct the dual inversion coupled response and perform structural suppression correction to obtain the final cross center positioning results; S5, together with steps S2, S3 and S4, jointly constructs a cross-shaped marker detection model for radiotherapy body surface; S6. Based on the radiotherapy surface cross mark detection model, process the standardized surface cross mark image and output the final cross center localization result.

2. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 1, characterized in that, Based on the standardized cross-shaped image, the gray values ​​of the local neighborhood set around the current pixel are aggregated and calculated through local neighborhood gray-level aggregation to generate the local neighborhood average result.

3. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 2, characterized in that, Based on the gray-level potential difference between the standardized cross-marked image and the local neighborhood average, a gray-level offset relationship is constructed, and a stability constant is used for constraint processing to construct the gray-level potential difference value.

4. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 3, characterized in that, We construct the local deviation relationship between the gray potential difference value and its neighborhood average result at different scales, and enhance the local deviation relationship by combining the deviation enhancement index. The enhancement result is combined with the scale weight coefficient to jointly construct the multi-scale initial response feature.

5. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 4, characterized in that, Based on multi-scale initial response features and combined with ROI response determination thresholds, low-response background regions are suppressed, while high-response regions where the cross marks on the body surface are preserved, generating ROI response sub-maps.

6. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 5, characterized in that, Based on the ROI response subgraph, the horizontal and vertical projection responses are extracted respectively, and the horizontal and vertical projection response values ​​are generated.

7. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 6, characterized in that, Based on the relative balance of the horizontal and vertical projection response values, a cross-equalization enhancement index is introduced for nonlinear modulation, and then synergistically fused with the average intensity of the two-direction projection response values ​​to generate a cross-direction equalization response value.

8. The method for detecting radiotherapy surface cross marks based on image recognition according to claim 7, characterized in that, Based on the cross-direction balanced response value, the candidate cross center point set is determined by adjusting the candidate threshold coefficient and combining it with the local maximum constraint. The preliminary cross center location result is then determined by combining the candidate cross center point set.

9. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 8, characterized in that, Based on the preliminary cross center positioning results, a dual inversion coupled response is constructed. Based on the dual inversion coupled response, structural suppression correction is performed on the preliminary cross center positioning results to output the final cross center positioning results.

10. The method for detecting cross-shaped marks on the surface of a radiotherapy body based on image recognition according to claim 9, characterized in that, The standardized body surface cross-marked image is used to construct the body surface ROI pre-location module in step S2, which outputs the ROI response sub-map; the ROI response sub-map is used to construct the cross projection localization module in step S3, which outputs the preliminary cross center localization result; the preliminary cross center localization result is used to construct the cross center structure correction module in step S4, which outputs the final cross center localization result.

Images