Machine learning-based intelligent assessment system and method for nursing wound images

Abstract

The present application relates to the technical field of image recognition, and particularly relates to a nursing wound image intelligent evaluation system and method based on machine learning. The method comprises the following steps: obtaining a nursing wound image through an image acquisition device and transmitting the image to a processing unit. First, a local window containing a wound area is selected, and a radial pixel-by-pixel scanning is performed to record a gray scale change sequence. Next, the gray scale sequence is input into a lightweight machine learning model for trend discrimination to filter out effective wound boundary feature sequences. Then, first-order differentiation is performed on the effective sequences to extract extreme points to generate a candidate boundary point set. Finally, based on the angle distribution of the scanning path, the candidate point set is ring-recombined to form a closed boundary curve, and a wound state evaluation result is output through gray scale difference, thereby completing the intelligent evaluation of the wound image. The present application improves the accuracy and stability of image processing by accurately identifying wound boundaries, optimizing gray scale change analysis and automatically correcting boundary discontinuous areas.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a machine learning-based intelligent assessment system and method for nursing wound images. Background Technology

[0002] Currently available wound image assessment technologies typically rely on traditional boundary detection algorithms or simple image processing methods. These methods often exhibit low accuracy and stability when dealing with complex wound morphologies. Specifically, traditional methods cannot effectively identify complex wound boundaries, especially when the edges are blurred or irregular, leading to significant errors in the assessment results.

[0003] Furthermore, many existing technologies cannot adaptively handle local discontinuities or breaks in wound images, often requiring manual intervention or overly cumbersome post-processing steps, thus reducing the system's automation level and processing efficiency. Summary of the Invention

[0004] Therefore, it is necessary to provide a machine learning-based intelligent assessment system and method for nursing wound images to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a machine learning-based intelligent assessment method for nursing wound images is proposed, the method comprising the following steps:

[0006] Step S1: Acquire images of the wound through an image acquisition device and transmit them to the processing unit;

[0007] Step S2: Select a local window containing the wound area, and scan pixel by pixel along the radial direction with the center pixel of the local window as the reference, and record the grayscale change sequence on each scanning path;

[0008] Step S3: Input the grayscale change sequence into the pre-trained lightweight machine learning model for trend discrimination and filter out the effective grayscale sequences corresponding to the wound boundary change features;

[0009] Step S4: Perform first-order difference on the effective gray-level sequence to obtain the gray-level gradient change set, and extract the extreme points to generate the wound boundary candidate point set;

[0010] Step S5: Based on the angular distribution of each scanning path, the candidate point set of the wound boundary is reorganized circumferentially to form a closed boundary curve, and the wound status assessment result is output according to the gray-scale difference inside and outside the closed boundary curve.

[0011] This invention also provides a machine learning-based intelligent assessment system for nursing wound images, used to perform the machine learning-based intelligent assessment method for nursing wound images as described above. The machine learning-based intelligent assessment system for nursing wound images includes:

[0012] The image acquisition module is used to acquire images of the wound through the image acquisition device and transmit them to the processing unit;

[0013] The radial scanning module is used to select a local window containing the wound area, and scans pixel by pixel along the radial direction with the center pixel of the local window as the reference, recording the grayscale change sequence on each scanning path.

[0014] The trend discrimination module is used to input grayscale change sequences into a pre-trained lightweight machine learning model for trend discrimination and to filter out effective grayscale sequences corresponding to wound boundary change features.

[0015] The gradient extraction module is used to perform first-order difference on the effective gray-level sequence to obtain a set of gray-level gradient changes, and extract extreme points to generate a set of candidate points for the wound boundary.

[0016] The boundary reconstruction module is used to perform circumferential reconstruction of the candidate point set of the wound boundary based on the angular distribution of each scanning path, forming a closed boundary curve, and outputting the wound status assessment result based on the gray-scale difference inside and outside the closed boundary curve.

[0017] The beneficial effects of this invention lie in acquiring wound images through an image acquisition device and scanning and recording grayscale change sequences along the radial direction. A pre-trained lightweight machine learning model is used for trend discrimination to filter out effective grayscale sequences of the wound boundary. Next, a candidate point set for the wound boundary is formed through first-order difference and extreme point extraction. These candidate points are then circumferentially reorganized according to the angular distribution of the scanning path to construct a closed wound boundary curve. Furthermore, the method analyzes the grayscale difference distribution between the inner and outer sides of the boundary through annular grayscale sampling and constructs a wound boundary stability criterion, thereby providing a wound status assessment result. For potential boundary discontinuities or omissions, the method also implements automatic interpolation and boundary correction mechanisms to ensure complete boundary closure and consistency with the actual wound morphology. Through intelligent correction of local boundary protrusions, depressions, or gaps, the method further optimizes the accuracy of the wound boundary. In addition, during the annular laying process, the annular position is adjusted according to the grayscale trend to effectively compensate for local discontinuous areas, ensuring complete coverage of the wound area. Overall, this method has significant advantages in accurately identifying and correcting wound boundaries, improving assessment stability and applicability, and can provide precise and real-time wound assessment technical support for clinical nursing, demonstrating high practicality and innovation. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of a machine learning-based intelligent assessment method for wound images in nursing care.

[0019] Figure 2 for Figure 1A detailed flowchart illustrating the implementation steps of step S5.

[0020] Figure 3 This is a schematic diagram of wound care based on a machine learning-based intelligent assessment method for wound images, as described in this application.

[0021] Figure 4 This is a schematic diagram of the annular grayscale sampling process of the intelligent assessment method for nursing wound images based on machine learning proposed in this application;

[0022] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0025] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] To achieve the above objectives, please refer to Figures 1 to 4 A machine learning-based intelligent assessment method for nursing wound images, the method comprising the following steps:

[0027] Step S1: Acquire images of the wound through an image acquisition device and transmit them to the processing unit;

[0028] Step S2: Select a local window containing the wound area, and scan pixel by pixel along the radial direction with the center pixel of the local window as the reference, and record the grayscale change sequence on each scanning path;

[0029] Step S3: Input the grayscale change sequence into the pre-trained lightweight machine learning model for trend discrimination and filter out the effective grayscale sequences corresponding to the wound boundary change features;

[0030] Step S4: Perform first-order difference on the effective gray-level sequence to obtain the gray-level gradient change set, and extract the extreme points to generate the wound boundary candidate point set;

[0031] Step S5: Based on the angular distribution of each scanning path, the candidate point set of the wound boundary is reorganized circumferentially to form a closed boundary curve, and the wound status assessment result is output according to the gray-scale difference inside and outside the closed boundary curve.

[0032] In one embodiment, using a hospitalized patient's lower limb pressure ulcer wound as the detection object, nursing staff use a handheld image acquisition device (such as a high-definition camera terminal with supplemental lighting) to acquire wound images at a fixed distance (e.g., 20cm), and transmit them wirelessly to a processing unit (such as an embedded edge computing device or a nursing terminal tablet). To reduce ambient light interference, a ring light is activated during the acquisition process to ensure uniform brightness in the wound area.

[0033] The processing unit first preprocesses the image (e.g., grayscale conversion, noise reduction) and determines the local window containing the wound (e.g., a 128×128 pixel region) through simple thresholding or manual selection. Starting from the center pixel of this window, it scans outward pixel by pixel according to a preset angular resolution (e.g., one path per 10°, a total of 36 radial paths). On each scanning path, it records the sequence of grayscale values ​​changing with distance; for example, on a certain path, the grayscale gradually changes from "dark red (low grayscale) → pink → normal skin (higher grayscale)," thus forming a set of grayscale change sequence data.

[0034] The processing unit first preprocesses the image (e.g., grayscale conversion, noise reduction) and determines the local window containing the wound (e.g., a 128×128 pixel region) through simple thresholding or manual selection. Starting from the center pixel of this window, it scans outward pixel by pixel according to a preset angular resolution (e.g., one path per 10°, a total of 36 radial paths). On each scanning path, it records the sequence of grayscale values ​​changing with distance; for example, on a certain path, the grayscale gradually changes from "dark red (low grayscale) → pink → normal skin (higher grayscale)," thus forming a set of grayscale change sequence data.

[0035] The selected valid grayscale sequences undergo first-order differencing, which calculates the grayscale difference (ΔGray) between adjacent pixels. In the valid sequences, the wound boundary typically corresponds to the location of the most drastic grayscale change, thus appearing as a maximum or minimum value in the differencing result. These extreme points are extracted and mapped back to the original image coordinates, forming a set of discrete wound boundary candidate points. For example, approximately 30 valid candidate points might be obtained from 36 paths.

[0036] Based on the angle information of each scanning path (e.g., 0°, 10°, ..., 350°), candidate points are circumferentially reconstructed according to angle order, and adjacent points are connected using interpolation or curve fitting methods (e.g., spline curves) to form a closed wound boundary curve. Subsequently, the average gray value difference between the region inside this closed curve and the adjacent external region is calculated. For example, if the gray value of the internal region is significantly lower than that of the external region (indicating a darker color, possible necrosis or exudation), it is assessed as "severe wound condition"; if the gray value difference is small and the boundary is clear, it is assessed as "good recovery".

[0037] In another embodiment, circumferential recombination may refer to Figure 3 The images still use pressure ulcers on the lower extremities of hospitalized patients as the primary images of the wounds.

[0038] As an example of the present invention, reference is made to Figure 2 As shown, step S5 in this example includes:

[0039] Step S51: Sort the candidate point set of the wound boundary according to the polar angle parameters of each scanning path, group the candidate points in the same angle interval into a group, and record the radial distance distribution of the candidate points in each group.

[0040] Step S52: Within each angular interval, select effective boundary points based on the density of radial distance distribution, and connect the effective boundary points of adjacent angular intervals end to end to construct an initial circumferential boundary chain;

[0041] Step S53: Perform ring-type gray-scale sampling on the initial circumferential boundary chain to obtain the inner gray-scale sequence and the outer gray-scale sequence, and calculate the segmented gray-scale difference distribution between the two.

[0042] Step S54: Construct a wound boundary stability criterion based on the fluctuation range and consistency of the segmented gray-level difference distribution, and use the wound boundary stability criterion to compare the gray-level contrast of the wound area to obtain the wound status assessment result.

[0043] In one embodiment, reference Figure 4Candidate points are sorted according to the polar angle parameters corresponding to each scanning path (e.g., divided into 10° intervals within the range of 0° to 360°), and candidate points falling into the same angle interval (e.g., each interval is 10°) are grouped together. For example, there may be multiple candidate points within a 20° interval, with radial distances of 30 pixels, 32 pixels, and 35 pixels, respectively. The radial distances of all candidate points in this group are statistically analyzed to form the radial distance distribution data for that angle interval, providing a basis for subsequent screening.

[0044] Within each angular interval, effective boundary points are selected based on the density of radial distances. Specifically, clustering or density judgment methods (such as selecting the region with the most concentrated distance distribution) can be used to determine representative points. For example, in the aforementioned 20° interval, if pixels 30 and 32 are relatively close, their average position (approximately 31 pixels) is selected as the effective boundary point of this interval, while pixels 35, which deviate significantly, are discarded. After selecting effective points for each angular interval, the effective boundary points of adjacent intervals are connected sequentially according to the angle size pattern (e.g., 20° connects to 30°, 30° connects to 40°), and closed connections are made at the beginning and end (e.g., 0° and 360°), thereby constructing an initial cyclic boundary chain.

[0045] A ring-shaped grayscale sampling process is performed on the initial circumferential boundary chain. Specifically, an inner sampling band is formed by offsetting inwards by a certain distance (e.g., 3 pixels) along the boundary chain, and an outer sampling band is formed by offsetting outwards by the same distance. Grayscale value sequences are extracted from both the inner and outer sampling bands and segmented according to angular order (e.g., every 30° segment). For example, if the average grayscale value in a segment is 80 on the inner side and 140 on the outer side, then the grayscale difference for that segment is 60. By calculating over the entire boundary, a set of segmented grayscale difference distributions can be obtained.

[0046] A wound boundary stability criterion is constructed based on the fluctuation range and consistency of the grayscale difference distribution. For example, when the grayscale difference in each segment is relatively large and the changes are stable (e.g., most intervals are between 50 and 70 with small fluctuations), it indicates that the wound boundary is clear and the tissue demarcation is obvious; if the grayscale difference is small or the fluctuation is drastic (e.g., some intervals are only 20, while others reach 80), it indicates that the boundary is blurred or that inflammation has spread. This stability criterion is combined with the overall grayscale contrast for analysis. For example: high contrast and stability → assessed as "good healing stage"; low contrast and instability → assessed as "high risk of deterioration or infection".

[0047] Preferably, before performing circumferential grayscale sampling on the initial circumferential boundary chain, the method further includes:

[0048] The initial circumferential boundary chain is continuously detected, and the segments where the angular interval between adjacent boundary points exceeds a preset threshold are marked as discontinuous regions.

[0049] For discontinuous regions, the grayscale change value of the angle intervals on both sides of the missing angle is taken;

[0050] Within the missing angle segment, boundary points are inserted sequentially along the radial direction according to the proportion of adjacent gray level changes;

[0051] Insert the boundary points into the initial circumferential boundary chain to form a continuous closed boundary curve.

[0052] In one embodiment, continuity detection is performed on the initial circumferential boundary chain. Specifically, the angular interval between adjacent points is calculated sequentially according to the polar angle order of the boundary points (e.g., 0°→360°). For example, normally the interval between adjacent points is 10°. If an interval of 30° or greater is detected between two points (exceeding a preset threshold, such as 15°), then a discontinuity is determined in that segment, and that angular interval (e.g., 40°~70°) is marked as a missing region.

[0053] Subsequently, grayscale change information of adjacent angle intervals is extracted for the discontinuous region. For example, for the 40° point on the left and the 70° point on the right of the missing interval, the corresponding grayscale change features (such as boundary gradient value or inner and outer grayscale difference value) are read and recorded as G1 and G2 respectively, for subsequent interpolation reference.

[0054] Next, boundary point interpolation is performed within the missing angle segment. In this embodiment, the 40°–70° interval is divided into two interpolation points, 50° and 60°, according to the original angular resolution (e.g., every 10°). Along the radial direction, linear interpolation is performed based on the proportional relationship between the grayscale changes on adjacent sides. For example, the grayscale change value at the 50° position is the midpoint between G1 and G2, and the corresponding radial boundary position is inferred from this (e.g., estimated through the extreme value position of the grayscale gradient). This generates new boundary point coordinates (e.g., polar coordinates r, θ).

[0055] Then, the newly generated boundary points are added to the original initial circumferential boundary chain and reordered according to their angles. For example, the newly generated 50° and 60° points are inserted between 40° and 70°, thereby restoring the segment to a continuous structure.

[0056] Finally, after the above processing, the original missing boundary chain was repaired into a complete and continuous closed curve. At this point, the boundary points are uniformly distributed within the range of 0° to 360°, and there are no longer any large-angle discontinuities.

[0057] Preferably, step S53 includes:

[0058] Based on the continuous closed boundary curve, several concentric rings are generated on the inner and outer sides along the normal direction of the curve, with each ring having the same width and covering the tissue area near the boundary.

[0059] Pixel grayscale values ​​are collected sequentially along the radial direction of each ring, forming an inner grayscale sequence and an outer grayscale sequence respectively;

[0060] The inner grayscale sequence and the outer grayscale sequence are divided into several segments of equal length, and the average grayscale difference of each segment is calculated to form the grayscale difference distribution of the segments.

[0061] In one embodiment, based on a continuously closed boundary curve, several concentric rings are generated on both the inner and outer sides along the curve's normal direction. In a specific implementation, the normal direction of each boundary point on the boundary curve is calculated (this can be obtained by fitting the tangent to adjacent boundary points and then calculating the normal), and then offset inwards (towards the wound area) and outwards (towards the normal skin area) by fixed distances. For example, setting the ring width to 2 pixels, three inner rings (2px, 4px, and 6px from the boundary) and three outer rings (2px, 4px, and 6px from the boundary) are generated, thus forming a multi-layered sampling region covering the tissue near the boundary.

[0062] Subsequently, pixel grayscale values ​​are sequentially acquired along the radial direction of each ring. In specific implementation, for each boundary point, the pixel grayscale values ​​corresponding to the inner and outer ring positions are extracted sequentially along its normal direction and arranged according to the angular order of the boundary points (e.g., 0° to 360°). For example, at a certain angular position, the grayscale values ​​of the inner three layers may be [75, 80, 85], and the outer three layers may be [130, 135, 140]. The grayscale values ​​of multiple layers on the same side can be averaged or weighted to form the inner and outer grayscale values ​​corresponding to that angle, thereby constructing a complete inner and outer grayscale sequence.

[0063] Next, the inner and outer grayscale sequences are divided into several segments of equal length. In this embodiment, the segments can be divided according to the angle range, for example, 30° per segment, dividing 360° into 12 segments. For multiple sampling points within each segment, the average inner grayscale value and the average outer grayscale value are calculated respectively.

[0064] Finally, the average grayscale difference for each segment is calculated, forming a segmented grayscale difference distribution. For example, in a certain segment, the average grayscale value on the inner side is 82, and the average grayscale value on the outer side is 138, then the grayscale difference is 56. After repeating this calculation process for all segments, a set of segmented difference data reflecting the grayscale contrast of the entire boundary area can be obtained (such as [52, 55, 60, 48,…]).

[0065] Preferably, based on a continuously closed boundary curve, several concentric rings are generated on the inner and outer sides along the curve's normal direction, including:

[0066] Based on the continuous closed boundary curve, extend several equidistant points inward and outward along the normal direction of each sampling point on the boundary curve;

[0067] Connect the inner and outer extension points sequentially along the boundary curve to form several complete concentric rings, with each ring having a consistent width;

[0068] The boundary tissue area of ​​the wound area is covered by concentric rings;

[0069] Repeat the above operation for all concentric rings.

[0070] In one embodiment, a multi-layered concentric ring structure for grayscale sampling is constructed based on the obtained continuous closed wound boundary curve to cover the tissue area near the wound boundary.

[0071] First, using a continuously closed boundary curve as a reference, the normal direction of each sampling point on the boundary curve is calculated. In the specific implementation, the tangent direction can be fitted between this point and its adjacent points before and after it, and the perpendicular direction is taken as the normal direction. Then, the normal direction is extended inward (pointing towards the inside of the wound) and outward (pointing towards the normal skin area) with a fixed step size. For example, if the step size is set to 2 pixels, then an inner point (2px, 4px, 6px) and an outer point (2px, 4px, 6px) are generated at each boundary point, forming several equally spaced sets of extension points.

[0072] Next, the extension points of each inner and outer layer are connected in circumferential order according to the boundary curve. In practice, for the same layer (e.g., all inner points 2px from the boundary), they are connected sequentially in angular order to form a closed curve; similarly, this process is repeated for layers of 4px, 6px, etc. In this way, multiple closed curves similar in shape to the original boundary curve can be formed on both the inner and outer sides of the boundary curve, and these closed curves constitute several concentric rings.

[0073] Subsequently, the width of the loop is defined by the region between two adjacent closed curves. For example, the region between 2px and 4px from the boundary constitutes the first loop, and the region between 4px and 6px constitutes the second loop. Since the extension step is fixed, the width of each loop remains consistent (e.g., all are 2 pixels), thus ensuring the uniformity and comparability of the sampling area.

[0074] Furthermore, these concentric rings cover a certain range of tissue areas both inside and outside the wound boundary. For example, the inner ring covers the inner edge tissue of the wound, while the outer ring covers the transition area of ​​normal skin, allowing subsequent grayscale sampling to simultaneously reflect the differences between diseased and healthy tissue.

[0075] Finally, the above generation process is repeated for all concentric rings, that is, the operation of "normal extension - circumferential connection - ring construction" is performed for each boundary point and each layer offset distance until a complete multi-layer ring structure system is formed.

[0076] Preferably, the boundary tissue area covered by the concentric rings of the wound area specifically includes:

[0077] Determine the continuously closed boundary curves and the resulting set of concentric rings;

[0078] Each ring is laid sequentially along the circumference of the boundary curve, ensuring that the width of each ring is consistent with that of the adjacent rings, forming a continuous covering ring layer sequence;

[0079] Based on the annular layer sequence, each annular band is marked with a gray-scale gradient reference in the inward and outward directions, so that the annular band coverage area can automatically distinguish between the internal and external tissues of the wound.

[0080] During the installation of the ring bands, for areas with uneven or gapped boundaries, the position of the ring bands is adjusted according to the grayscale trend of adjacent ring bands to achieve local continuous compensation until the concentric ring bands cover the boundary tissue area of ​​the wound area.

[0081] In one embodiment, to ensure that the generated concentric rings can completely and continuously cover the tissue area near the wound boundary, the process of laying and adaptively adjusting the rings is refined.

[0082] First, a continuously closed boundary curve and a set of concentric rings generated based on this boundary curve are determined. In the specific implementation, the boundary curve (an ordered set of polar coordinate points) and the corresponding multiple layers of inner and outer rings (e.g., 3 inner layers and 3 outer layers) have been obtained through the aforementioned steps. These rings expand outward and inward from the boundary with a fixed step size, forming an initial set of ring structure.

[0083] Subsequently, the rings are laid out sequentially along the circumference of the boundary curve. During implementation, the points of each ring layer are connected point by point according to the angular order (e.g., 0° to 360°) to ensure that each ring layer is a closed curve. At the same time, the distance between adjacent rings is kept consistent (e.g., 2 pixels each), thus forming a regular "ring layer sequence" (e.g., inner layer 1, 2, 3; outer layer 1, 2, 3), achieving continuous coverage of the boundary area.

[0084] Next, grayscale gradient reference labels are applied to each annulus based on the annular layer sequence. In practice, the positional attribute (inner or outer) of each annulus relative to the boundary is recorded and labeled in conjunction with the grayscale sampling results. For example, the inner annulus typically corresponds to lower grayscale (wound area), while the outer annulus corresponds to higher grayscale (normal skin). Based on this, "inner / outer attribute labels" or grayscale gradient direction identifiers are established for each annulus, enabling subsequent processing to automatically distinguish between internal and external wound tissues.

[0085] During the installation of the ring-shaped paving, local compensation is also required for complex boundary morphologies. For example, when there are obvious protrusions (sharp corners) or depressions (gap) at the boundary, simple normal extension may lead to overlap or gaps between the rings. Therefore, dynamic adjustments are made based on the grayscale variation trends of adjacent rings.

[0086] If the spacing between the rings in a certain area is too large (gaps appear), then compensation points are inserted in that section based on the grayscale continuity of the adjacent areas, so that the rings extend into the gap area;

[0087] If the rings in a certain area overlap or are too dense, the local ring points are smoothed or extrapolated to restore a uniform spacing.

[0088] For example, in a concave region, if the grayscale change of the inner ring still maintains a continuous increasing trend, the outer ring can be appropriately compensated inward to keep the overall grayscale gradient consistent.

[0089] Ultimately, through the aforementioned "circumferential laying-gradient marking-local compensation" process, all concentric rings form a continuous, uniform, and adaptive coverage structure in space, completely enveloping the tissue area near the wound boundary.

[0090] Preferably, during the laying of the ring belts, for local boundary unevenness or gap areas, adjusting the position of the ring belts according to the grayscale trend of adjacent ring belts to achieve local continuous compensation includes:

[0091] Continuous detection is performed on each ring along the boundary curve in the circumferential direction to identify abnormal sections with protrusions, depressions or covering gaps, and the corresponding angle range is recorded.

[0092] Within the abnormal section, extract the gray values ​​of the current ring and its adjacent rings at corresponding positions, calculate the difference in gray value changes at each position, and select the point with the maximum gray value change difference as the priority adjustment point.

[0093] Using the priority adjustment point as a reference, the current ring is shifted and corrected along the normal direction corresponding to that point, so that the grayscale change at that point is consistent with that of the adjacent ring.

[0094] Centered on the priority adjustment point, the adjustment is expanded point by point in the circumference to both sides. The displacement of adjacent positions is corrected in turn according to the preset step size, and the displacement of each position is set in a decreasing manner.

[0095] The corrected annular bands were then reviewed.

[0096] In one embodiment, to address the issue of discontinuous rings caused by complex wound boundary morphology (such as local protrusions, depressions, or uneven ring coverage), a local continuity compensation mechanism based on grayscale trends is introduced during the ring laying process.

[0097] First, the continuity of each concentric ring is detected circumferentially along the boundary curve. In the specific implementation, the points of each ring are traversed in angular order (e.g., sampling every 10°), and the spatial spacing and grayscale change consistency between adjacent points are calculated. When an abnormal increase in the ring spacing (e.g., exceeding a set threshold), local overlap, or abrupt grayscale change is detected in a certain angular segment, the segment is determined to be an abnormal segment, and its corresponding angular range (e.g., 120° to 150°) is recorded.

[0098] Subsequently, the grayscale values ​​of the current annular band and its adjacent annular bands (such as inner or outer adjacent layers) at corresponding positions are extracted within the abnormal section. For example, at a certain angle, the grayscale value of the current annular band is 90, the adjacent inner annular band is 70, and the adjacent outer annular band is 130. The grayscale variation difference is calculated (e.g., |90-70|=20, |130-90|=40), and the distribution of the difference at each point is compared. Within this section, the point with the largest grayscale variation difference (e.g., the point with a difference of 40) is selected as the priority adjustment point, serving as the core reference for local correction.

[0099] Next, using the priority adjustment point as a reference, the current ring position is shifted and corrected along the normal direction of that point. For example, if the current ring is biased outward at that point, causing an abnormal increase in grayscale difference, it is finely adjusted inward along the normal direction (such as moving 1 to 2 pixels) to make its grayscale value closer to the change trend of the adjacent ring, thereby restoring the continuity of the grayscale gradient.

[0100] After correcting the priority adjustment point, the correction is performed continuously, expanding outwards from that point in a circumferential direction. Specifically, adjacent points are adjusted sequentially at preset step sizes (e.g., every 5° or per sampling point), with decreasing displacement amounts. For example: the priority point displacement is 2 pixels; the first adjacent point displacement is 1.5 pixels; the second point displacement is 1 pixel; and subsequent displacements gradually decrease to 0.

[0101] This adjustment strategy of "maximum at the center and decreasing on both sides" allows for a smooth transition in the local ring morphology, avoiding new abrupt changes.

[0102] Finally, the corrected ring bands are reviewed. During the review, the ring band spacing and grayscale variation difference of the section are recalculated to determine whether the continuity requirements are met (e.g., the grayscale variation tends to be stable and the ring band spacing returns to uniformity). If abnormalities still exist, the above correction process can be repeated until the preset continuity standard is met.

[0103] Preferably, within the missing angle segment, inserting boundary points radially in proportion to the change in adjacent gray levels includes:

[0104] The missing angle segment is divided into several interpolation units and numbered in circumferential order;

[0105] In the radial direction corresponding to each interpolation unit, obtain the gray value and radial position of the existing boundary points on both sides of the unit;

[0106] Based on the grayscale difference between the two boundary points, determine the proportion of grayscale change along the radial direction and determine the candidate insertion position;

[0107] Multiple candidate insertion points are set in each interpolation unit. The gray-scale changes corresponding to each candidate insertion point are compared, and the point with the most continuous gray-scale transition on both sides is selected as the actual insertion point.

[0108] Repeat the above operation for all interpolation units to complete the boundary point insertion in the circumferential direction.

[0109] In one embodiment, the missing angle segment is divided into several interpolation units along the circumferential direction, for example, by each sampling point interval or every 5° angle. Each interpolation unit is numbered sequentially in the circumferential direction for easy sequential processing and interpolation management.

[0110] For each interpolation unit, identify the existing boundary points on both sides of the unit (i.e., the continuous boundary points at both ends of the missing segment) and obtain their grayscale values. and and the corresponding radial position and .

[0111] Based on the grayscale difference between the two boundary points and radial spacing Calculate the grayscale change ratio along the radial direction: Based on this, the radial position of the candidate insertion point is determined. Corresponding target grayscale value: .

[0112] Within each interpolation unit, several candidate insertion points (e.g., 3-5) are evenly arranged radially, and the actual grayscale value of each candidate point is obtained. Compare the grayscale values ​​of the candidate points with the target grayscale values. Compare and calculate the grayscale difference: The candidate point with the smallest gray-scale difference is selected as the actual insertion point, so that the gray-scale transition is the smoothest and most continuous after the boundary point is inserted.

[0113] Repeat the above operation for all interpolation units in the missing segment, and add the actual insertion point of each unit to the boundary point set in turn, and arrange them in circumferential order to form a continuous closed boundary curve.

[0114] Preferably, based on the grayscale difference between the two boundary points, the radial grayscale variation ratio is determined, and the candidate insertion position is determined as follows:

[0115] Obtain the grayscale values ​​and radial distances of the two boundary points on both sides of the missing angle segment in the radial direction, and determine the grayscale difference and distance interval;

[0116] Based on the grayscale difference and distance interval, calculate the grayscale change per unit distance and divide the grayscale transition intervals.

[0117] Within the grayscale transition range, multiple candidate location points are generated radially according to a preset step size;

[0118] For each candidate location point, read its corresponding pixel grayscale value and match and compare it with the expected grayscale transition range;

[0119] Candidate locations that meet the condition of continuous grayscale variation are selected as candidate insertion locations.

[0120] In one embodiment, two boundary points are determined radially for the missing angle segment: one side is the current ring or inner boundary point, and the other side is the adjacent ring or outer boundary point. The grayscale values ​​of the two boundary points are recorded. and and its radial distance and Calculate the grayscale difference and radial distance interval .

[0121] Calculate the grayscale change per unit pixel along the radial direction based on the grayscale difference and distance interval: The radial segment is then divided into several transition intervals with equal grayscale changes, forming a grayscale gradient reference sequence.

[0122] Within the grayscale transition range, multiple candidate location points are generated radially according to a preset step size (e.g., every 1 to 2 pixels), and each candidate point corresponds to a predicted grayscale value: For each candidate location point, read the actual image pixel grayscale value. and the expected grayscale value Perform a matching comparison and calculate the grayscale difference: ,like If the point is within the preset continuity threshold range, it is considered to meet the grayscale continuity requirement.

[0123] The set of candidate points that meet the grayscale continuity condition is recorded as the candidate set of insertion positions in the radial direction. Subsequently, when inserting boundary points radially in the missing angle segment, the optimal point is selected from this candidate set for insertion, thereby forming a continuous closed boundary curve.

[0124] Of particular importance, step S2 includes:

[0125] Based on the boundary location of the wound area, determine the local window that includes the wound area, ensuring that the window's range can completely cover the wound boundary and surrounding tissue area;

[0126] Using the center pixel of the local window as a reference, the radial scanning direction is determined, and this direction is divided into several equally spaced pixel scanning paths;

[0127] Along each scanning path, scan pixel by pixel from the center of the window outwards, record the gray value of each pixel position, and calculate the gray value difference between adjacent pixels;

[0128] For each scanning path, calculate the average value and standard deviation of its grayscale changes, and determine whether the grayscale changes of the path meet the preset continuity standard.

[0129] When the grayscale changes do not meet the preset standards, the scanning path range or scanning step size is readjusted, and the grayscale changes are recorded again to ensure that the grayscale sequence with high continuity is effectively acquired.

[0130] Integrate the grayscale change sequences of all scan paths.

[0131] In one embodiment, the radial scanning and grayscale sequence acquisition process in step S2 is specifically implemented using an image of a pressure ulcer on a patient's back as the processing object, in order to improve the stability and accuracy of subsequent boundary recognition.

[0132] First, a local window is determined based on the boundary position of the wound area. In a specific implementation, the approximate outline of the wound can be obtained through preliminary segmentation (such as threshold segmentation or simple edge detection), and a rectangular or circular local window is generated by expanding this outline by a certain number of pixels (such as 10 to 20 pixels). For example, if the wound area is approximately 80×80 pixels, a 128×128 pixel window can be selected to ensure complete coverage of the wound boundary and the surrounding normal tissue area.

[0133] Subsequently, the radial scanning direction is determined based on the center pixel of the local window. In this embodiment, the 360° range is evenly divided into several scanning directions (e.g., one every 10°, for a total of 36 paths), and each direction is defined as a pixel scanning path extending outward from the center. The paths are evenly spaced angularly, thereby achieving omnidirectional coverage of the wound area.

[0134] Next, a pixel-by-pixel grayscale scan is performed along each scan path. Specifically, starting from the center of the window, the system moves outward pixel by pixel in a specified direction, recording the grayscale value (Gray(i)) of each pixel, and calculating the grayscale difference between adjacent pixels ΔGray(i) = Gray(i+1) - Gray(i). For example, on a certain path, a grayscale sequence might be obtained: [60, 62, 65, 90, 130…], with a corresponding difference sequence of: [2, 3, 25, 40…], where larger differences usually correspond to the wound boundary position.

[0135] Then, for each scanning path, its grayscale change characteristic parameters are calculated, including the average and standard deviation of the grayscale difference. For example, if the average grayscale difference of a path is large and the standard deviation is moderate, it indicates that the path has a clear and stable grayscale transition; conversely, if the average is small or the standard deviation is too large (violent fluctuations), it may be affected by noise or invalid regions. The above statistical results are compared with preset continuity standards (such as the average difference threshold and the upper limit of the standard deviation) to determine whether the path is a valid path.

[0136] When the grayscale variation of a certain scanning path does not meet the preset standard, it is adaptively adjusted. In specific implementation, one or a combination of the following methods can be adopted:

[0137] Reduce or expand the scan path range (e.g., limit the maximum scan radius);

[0138] Adjust the scan step size (e.g., change from pixel-by-pixel sampling to interval sampling or vice versa);

[0139] Fine-tune the scanning angle (e.g., offset by ±2° to avoid noisy areas).

[0140] After adjustment, grayscale acquisition and difference calculation are re-executed until the path meets the continuity requirement, thereby ensuring that a high-quality grayscale change sequence is acquired.

[0141] Finally, the grayscale change sequences of all scanning paths are integrated. The paths are arranged in angular order to form a structured multi-directional grayscale change data set (such as a two-dimensional array: path number × grayscale sequence), which provides basic data support for grayscale trend discrimination, boundary extraction and circumferential reconstruction in subsequent steps.

[0142] Of particular importance is that, based on the center pixel of the local window, determining the radial scan direction includes:

[0143] A polar coordinate reference system is established with the center pixel of the local window as the origin, and the pixel position in the image plane is converted into a representation with radial and angular parameters.

[0144] Within a preset angle range, the radial direction is discretized according to equal angular intervals to generate several scanning paths extending outward from the center pixel;

[0145] Boundary constraints are applied to each scan path to ensure that its termination position is within the local window boundary, thus guaranteeing that all scan paths cover the entire area without exceeding the boundary.

[0146] The pixels on each scanning path are sorted according to their radial distance to form an ordered pixel sequence from the inside out.

[0147] In one embodiment, a polar coordinate reference system is established with the center pixel of the local window as the origin. In a specific implementation, the two-dimensional Cartesian coordinates (x, y) in the image are converted to polar coordinates (r, y). ), where r represents the radial distance from the pixel to the center. This represents the polar angle of the point relative to the horizontal reference direction. This transformation allows the pixel distribution in the image to be uniformly described in the form of "distance + direction," providing a coordinate basis for subsequent radial scanning.

[0148] Subsequently, the radial direction is discretized within a preset angle range. Typically, the angle range is set to 0° to 360°, and divided at equal angular intervals (such as 5°, 10°, or 15°). For example, when the interval is 10°, 36 scanning paths can be generated. Each path starts from the center pixel and extends outward along the corresponding angular direction, forming a radial structure covering the entire local window.

[0149] Next, boundary constraints are applied to each scanning path. During implementation, the maximum allowable radial distance for each path is calculated based on the size of the local window (e.g., 128×128 pixels), and the scanning path terminates when it reaches the window boundary. For example, sampling stops when the path extends to the image edge or reaches the window border, thus ensuring that all scanning paths remain within the local window area and avoiding out-of-bounds access or invalid data acquisition.

[0150] Then, the pixels along each scan path are sorted according to their radial distance. Specifically, starting from the center pixel along the path, the pixels are recorded sequentially. Increasing pixels, forming an ordered sequence (such as...) For example, obtaining a pixel sequence in a certain direction. ,in Closest to the center Located near the window boundary. This sorting ensures that grayscale variation analysis can be performed strictly in a "from inside to outside" spatial order.

[0151] This invention also provides a machine learning-based intelligent assessment system for nursing wound images, used to perform the machine learning-based intelligent assessment method for nursing wound images as described above. The machine learning-based intelligent assessment system for nursing wound images includes:

[0152] The image acquisition module is used to acquire images of the wound through the image acquisition device and transmit them to the processing unit;

[0153] The radial scanning module is used to select a local window containing the wound area, and scans pixel by pixel along the radial direction with the center pixel of the local window as the reference, recording the grayscale change sequence on each scanning path.

[0154] The trend discrimination module is used to input grayscale change sequences into a pre-trained lightweight machine learning model for trend discrimination and to filter out effective grayscale sequences corresponding to wound boundary change features.

[0155] The gradient extraction module is used to perform first-order difference on the effective gray-level sequence to obtain a set of gray-level gradient changes, and extract extreme points to generate a set of candidate points for the wound boundary.

[0156] The boundary reconstruction module is used to perform circumferential reconstruction of the candidate point set of the wound boundary based on the angular distribution of each scanning path, forming a closed boundary curve, and outputting the wound status assessment result based on the gray-scale difference inside and outside the closed boundary curve.

[0157] The beneficial effects of this invention lie in acquiring wound images through an image acquisition device and scanning and recording grayscale change sequences along the radial direction. A pre-trained lightweight machine learning model is used for trend discrimination to filter out effective grayscale sequences of the wound boundary. Next, a candidate point set for the wound boundary is formed through first-order difference and extreme point extraction. These candidate points are then circumferentially reorganized according to the angular distribution of the scanning path to construct a closed wound boundary curve. Furthermore, the method analyzes the grayscale difference distribution between the inner and outer sides of the boundary through annular grayscale sampling and constructs a wound boundary stability criterion, thereby providing a wound status assessment result. For potential boundary discontinuities or omissions, the method also implements automatic interpolation and boundary correction mechanisms to ensure complete boundary closure and consistency with the actual wound morphology. Through intelligent correction of local boundary protrusions, depressions, or gaps, the method further optimizes the accuracy of the wound boundary. In addition, during the annular laying process, the annular position is adjusted according to the grayscale trend to effectively compensate for local discontinuous areas, ensuring complete coverage of the wound area. Overall, this method has significant advantages in accurately identifying and correcting wound boundaries, improving assessment stability and applicability, and can provide precise and real-time wound assessment technical support for clinical nursing, demonstrating high practicality and innovation.

[0158] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0159] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A machine learning-based intelligent assessment method for nursing wound images, characterized in that, Includes the following steps: Step S1: Acquire images of the wound through an image acquisition device and transmit them to the processing unit; Step S2: Select a local window containing the wound area, and scan pixel by pixel along the radial direction with the center pixel of the local window as the reference, and record the grayscale change sequence on each scanning path; Step S3: Input the grayscale change sequence into the pre-trained lightweight machine learning model for trend discrimination and filter out the effective grayscale sequences corresponding to the wound boundary change features; Step S4: Perform first-order difference on the effective gray-level sequence to obtain the gray-level gradient change set, and extract the extreme points to generate the wound boundary candidate point set; Step S5: Based on the angular distribution of each scanning path, the candidate point set of the wound boundary is reorganized circumferentially to form a closed boundary curve, and the wound status assessment result is output according to the gray-scale difference inside and outside the closed boundary curve.

2. The intelligent assessment method for nursing wound images based on machine learning according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Sort the candidate point set of the wound boundary according to the polar angle parameters of each scanning path, group the candidate points in the same angle interval into a group, and record the radial distance distribution of the candidate points in each group. Step S52: Within each angular interval, select effective boundary points based on the density of radial distance distribution, and connect the effective boundary points of adjacent angular intervals end to end to construct an initial circumferential boundary chain; Step S53: Perform ring-type gray-scale sampling on the initial circumferential boundary chain to obtain the inner gray-scale sequence and the outer gray-scale sequence, and calculate the segmented gray-scale difference distribution between the two. Step S54: Construct a wound boundary stability criterion based on the fluctuation range and consistency of the segmented gray-level difference distribution, and use the wound boundary stability criterion to compare the gray-level contrast of the wound area to obtain the wound status assessment result.

3. The intelligent assessment method for nursing wound images based on machine learning according to claim 2, characterized in that, Before performing circumferential grayscale sampling on the initial circumferential boundary chain, the following steps are also included: The initial circumferential boundary chain is continuously detected, and the segments where the angular interval between adjacent boundary points exceeds a preset threshold are marked as discontinuous regions. For discontinuous regions, the grayscale change value of the angle intervals on both sides of the missing angle is taken; Within the missing angle segment, boundary points are inserted sequentially along the radial direction according to the proportion of adjacent gray level changes; Insert the boundary points into the initial circumferential boundary chain to form a continuous closed boundary curve.

4. The intelligent assessment method for nursing wound images based on machine learning according to claim 2, characterized in that, Step S53 includes: Based on the continuous closed boundary curve, several concentric rings are generated on the inner and outer sides along the normal direction of the curve, with each ring having the same width and covering the tissue area near the boundary. Pixel grayscale values ​​are collected sequentially along the radial direction of each ring, forming an inner grayscale sequence and an outer grayscale sequence respectively; The inner grayscale sequence and the outer grayscale sequence are divided into several segments of equal length, and the average grayscale difference of each segment is calculated to form the grayscale difference distribution of the segments.

5. The intelligent assessment method for nursing wound images based on machine learning according to claim 4, characterized in that, Based on a continuously closed boundary curve, several concentric rings are generated on the inner and outer sides along the curve's normal direction, including: Based on the continuous closed boundary curve, extend several equidistant points inward and outward along the normal direction of each sampling point on the boundary curve; Connect the inner and outer extension points sequentially along the boundary curve to form several complete concentric rings, with each ring having a consistent width; The boundary tissue area of ​​the wound area is covered by concentric rings; Repeat the above operation for all concentric rings.

6. The intelligent assessment method for nursing wound images based on machine learning according to claim 5, characterized in that, The boundary tissue area covered by the concentric rings of the wound area is specifically as follows: Determine the continuously closed boundary curves and the resulting set of concentric rings; Each ring is laid sequentially along the circumference of the boundary curve, ensuring that the width of each ring is consistent with that of the adjacent rings, forming a continuous covering ring layer sequence; Based on the annular layer sequence, each annular band is marked with a gray-scale gradient reference in the inward and outward directions, so that the annular band coverage area can automatically distinguish between the internal and external tissues of the wound. During the installation of the ring bands, for areas with uneven or gapped boundaries, the position of the ring bands is adjusted according to the grayscale trend of adjacent ring bands to achieve local continuous compensation until the concentric ring bands cover the boundary tissue area of ​​the wound area.

7. The intelligent assessment method for nursing wound images based on machine learning according to claim 6, characterized in that, During the installation of the ring belts, for areas with uneven or gapped boundaries, the position of the ring belts is adjusted based on the grayscale trend of adjacent ring belts to achieve local continuous compensation, including: Continuous detection is performed on each ring along the boundary curve in the circumferential direction to identify abnormal sections with protrusions, depressions or covering gaps, and the corresponding angle range is recorded. Within the abnormal section, extract the gray values ​​of the current ring and its adjacent rings at corresponding positions, calculate the difference in gray value changes at each position, and select the point with the maximum gray value change difference as the priority adjustment point. Using the priority adjustment point as a reference, the current ring is shifted and corrected along the normal direction corresponding to that point, so that the grayscale change at that point is consistent with that of the adjacent ring. Centered on the priority adjustment point, the adjustment is expanded point by point in the circumference to both sides. The displacement of adjacent positions is corrected in turn according to the preset step size, and the displacement of each position is set in a decreasing manner. The corrected annular bands were then reviewed.

8. The intelligent assessment method for nursing wound images based on machine learning according to claim 3, characterized in that, Within the missing angle segment, boundary points are inserted radially in sequence according to the proportion of adjacent gray-level changes, including: The missing angle segment is divided into several interpolation units and numbered in circumferential order; In the radial direction corresponding to each interpolation unit, obtain the gray value and radial position of the existing boundary points on both sides of the unit; Based on the grayscale difference between the two boundary points, determine the proportion of grayscale change along the radial direction and determine the candidate insertion position; Multiple candidate insertion points are set in each interpolation unit. The gray-scale changes corresponding to each candidate insertion point are compared, and the point with the most continuous gray-scale transition on both sides is selected as the actual insertion point. Repeat the above operation for all interpolation units to complete the boundary point insertion in the circumferential direction.

9. The intelligent assessment method for nursing wound images based on machine learning according to claim 8, characterized in that, Based on the grayscale difference between the two boundary points, the radial grayscale variation ratio is determined, and the specific candidate insertion positions are determined as follows: Obtain the grayscale values ​​and radial distances of the two boundary points on both sides of the missing angle segment in the radial direction, and determine the grayscale difference and distance interval; Based on the grayscale difference and distance interval, calculate the grayscale change per unit distance and divide the grayscale transition intervals. Within the grayscale transition range, multiple candidate location points are generated radially according to a preset step size; For each candidate location point, read its corresponding pixel grayscale value and match and compare it with the expected grayscale transition range; Candidate locations that meet the condition of continuous grayscale variation are selected as candidate insertion locations.

10. A machine learning-based intelligent assessment system for nursing wound images, characterized in that, For performing the machine learning-based intelligent assessment method for nursing wound images as described in claim 1, the machine learning-based intelligent assessment system for nursing wound images comprises: The image acquisition module is used to acquire images of the wound through the image acquisition device and transmit them to the processing unit; The radial scanning module is used to select a local window containing the wound area, and scans pixel by pixel along the radial direction with the center pixel of the local window as the reference, recording the grayscale change sequence on each scanning path. The trend discrimination module is used to input grayscale change sequences into a pre-trained lightweight machine learning model for trend discrimination and to filter out effective grayscale sequences corresponding to wound boundary change features. The gradient extraction module is used to perform first-order difference on the effective gray-level sequence to obtain a set of gray-level gradient changes, and extract extreme points to generate a set of candidate points for the wound boundary. The boundary reconstruction module is used to perform circumferential reconstruction of the candidate point set of the wound boundary based on the angular distribution of each scanning path, forming a closed boundary curve, and outputting the wound status assessment result based on the gray-scale difference inside and outside the closed boundary curve.

Images