Biological image analysis methods for assessing and detecting color anomalies through color distribution.

Abstract

A biomedical image analysis method for distinguishing melanoma from benign skin lesions in digital images of skin tissue. The method is based on color anomaly detection and provides a suitable metric for quantitatively assessing the probability that a skin lesion is melanoma. This method is not AI-based (AI faces challenges such as insufficient training data, poor interpretability, and high computational requirements), and it is independent of camera sensor specifications and lighting conditions during image capture. The method can also be integrated into a smartphone application for use by non-professional individuals, requiring minimal computational power.

Description

Technical Field

[0001] This invention relates to a biomedical image analysis method for detecting lesions in digital images of tissues, specifically for detecting melanoma. The invention also relates to apparatus suitable for performing the method and a computer program for the method. Background Technology

[0002] Melanoma is an aggressive skin cancer with a rapidly rising incidence. In its early stages, melanoma can be successfully treated with surgery alone, resulting in high survival rates; however, once metastasis occurs, survival rates drop significantly. Therefore, early and accurate diagnosis is crucial to ensuring patients receive the best possible outcome. Medical malpractice claims related to pathology and dermatology due to melanoma misdiagnosis exceed those for any other cancer except breast cancer, because malignant melanoma can clinically masquerade as benign lesions (false negatives), while benign pigmented lesions can clinically mimic malignant melanoma (false positives). Therefore, a tool to assist clinicians and improve workflow is clearly needed.

[0003] Over the past few decades, digital microscopy has become an indispensable tool in pathological research. The generation of digital images of tissue sections has enabled the creation and application of image analysis algorithms. Image analysis can be defined as a combination of computational techniques used to automatically process carefully captured images to generate measurement datasets and extract useful information. Images themselves can come from a variety of sources, from digital cameras or smartphones to more advanced sensors integrated into sophisticated instruments. Image analysis has been proposed for extracting biologically relevant information in applications such as the detection of plant diseases, the determination of food quality and safety, and the diagnosis of skin cancer. In all such cases, the purpose of analyzing images (bioimage analysis) is to identify abnormalities attributable to biological phenomena compared to known (normal) conditions.

[0004] Bio-image analysis can achieve excellent results by employing specialized instruments with specific camera sensor characteristics. However, in today's era of smartphone diagnostics, building specialized instruments is considered outdated and fails to meet the demand for rapid, convenient, and low-cost "on-demand diagnostics" or self-testing. Smartphone applications performing bio-image analysis for diagnostic purposes lack accuracy due to a) variations in smartphone camera specifications, b) changing lighting conditions during image capture, and c) photos taken by non-professionals. As an alternative, deep learning techniques, known as AI methods, have become a new trend due to their immunity to these issues. However, AI-based bio-image analysis requires high-quality photos that have undergone manual preprocessing, which can be time-consuming and vary depending on individual skill levels. Furthermore, existing bio-image analysis techniques face challenges such as insufficient training data, poor interpretability, and high computational demands. Most importantly, existing diagnostic bio-image analysis methods based on color anomaly detection do not provide a suitable metric for quantitatively assessing the likelihood of skin lesions being melanoma.

[0005] Therefore, there is still a need for image analysis methods that are easy to implement and highly accurate to differentiate between malignant melanoma and benign skin lesions. Summary of the Invention

[0006] The present invention described herein provides a method as described in claim 1, which solves the problems of the prior art. Specifically, the present invention provides a computer-implemented method for detecting melanoma in digital skin images, the method comprising the following steps: a) Determine the region of interest (AOI) in the digital image using an image segmentation algorithm; b) Calculate the mean and standard deviation of the AOI and the region outside the AOI to extract the risk threshold; c) Calculate the geometric center of the AOI, and thereby define the initial centroid; d) Divide the AOI by color to obtain its outline; e) If multiple contours are identified in step (d), calculate the Euclidean distance of the centroid of each identified contour. f) Calculate the sum of the average of all measured distances and the standard deviation, and then divide by the ratio of each profile distance; g) The ratio is compared with the risk threshold extracted in step (b), thereby classifying the contour as having a normal or abnormal color distribution; and h) Repeat steps (d) to (g) by applying multiple thresholds to the image until the size of the contour covers the AOI; The presence of multiple contours with abnormal color distribution and whose number increases during the process indicates melanoma.

[0007] The method described in this paper helps detect color anomalies during biological image analysis to characterize specific skin conditions as "normal" or "abnormal" compared to known physiological states, and specifically to differentiate between malignant melanomas and benign nevi. This method can also be applied to a range of medical, biological, or other applications, such as detecting plant diseases or food quality deterioration. The method is not AI-based and does not depend on camera sensor specifications or lighting conditions during image capture. It can also be integrated into smartphone applications for use by non-professional individuals with low computational requirements.

[0008] This method evaluates the color distribution in a specific region of interest in an image by employing a series of mathematical calculations that must be performed in a specific order. These calculations result in identifying color contours within the regions of interest and the relative distances between them. Compared to prior art image analysis methods, the disclosed method has the advantage of providing a quantitative method for assessing the likelihood that a particular skin lesion is melanoma or a benign nevus. As will be shown in the accompanying embodiments, the method of the present invention distinguishes malignant melanomas with an accuracy of 93% or higher, which is higher than the accuracy typically achieved by the naked eye of a trained medical expert, which rarely exceeds 70%.

[0009] The present invention also provides an apparatus for detecting melanoma in digital images of skin tissue, the apparatus comprising a camera adapted to acquire the image, a processor adapted to process the image data according to the methods disclosed herein, and an output device adapted to indicate the likelihood of the presence or absence of melanoma.

[0010] A computer program including instructions is also provided, which, when executed by a computer, cause the computer to perform the methods disclosed herein.

[0011] Other aspects and benefits of the invention will become apparent from the detailed description below. Attached Figure Description

[0012] The invention will now be described with reference to certain embodiments illustrated in the accompanying drawings. It should be noted that the drawings illustrate preferred embodiments of the invention and should not be considered as limiting the scope of the invention.

[0013] Figure 1 A schematic flowchart of a method for detecting abnormal morphological features in digital images of skin tissue according to the present invention is shown.

[0014] Figure 2 The workflow of a method for detecting abnormal morphological features in digital images of skin tissue according to a preferred embodiment of the present invention is shown.

[0015] Figure 3 An exemplary output graph according to the present invention is shown. As depicted by the top curve (A), an increase in the number of contours classified as having an abnormal color distribution (“risk contours”) during the process (repetition) indicates the presence of melanoma. A stable or decreasing number of contours with an abnormal color distribution (B) corresponds to benign skin conditions.

[0016] Figure 4 This is illustrated through a series of images showing the contour processing applied to melanocytic nevi (A) and melanoma (B). In the nevi, the color spreads uniformly, while in the melanoma, the color spreads randomly.

[0017] Figure 5 This is a graphical representation of the process for calculating the number of peaks around the boundary of an AOI. (A) is a digital image corresponding to melanoma, while (B) is a digital image corresponding to a mole.

[0018] Figure 6 It is to calculate the optimal R c A diagram illustrating the process. (A), (C), (E), and (G) represent results from the training set; (B), (D), (F), and (H) represent results from the test set. The tested R... c The values ​​are 0.25 (A, B), 0.24 (C, D), 0.22 (E, F), and 0.22 (G, H). The result obtained at the end of the process is R. c =0.232 (I).

[0019] Figure 7 Dermoscopic images of moles (A) and melanomas (C) are shown, along with corresponding outputs for moles (B) and melanomas (D). Detailed Implementation

[0020] This disclosure provides a computer-implemented image analysis method for diagnosing melanoma in digital skin images. The invention also provides an apparatus for detecting melanoma in digital images of tissue and a computer program for implementing the method.

[0021] The term “centroid” or “center” as used throughout the specification and claims refers to the geometric center of the profile.

[0022] The term “sub-profile” as used throughout the specification and claims refers to any profile detected within a profile.

[0023] This method involves evaluating the color distribution in an image, which can be used to characterize a specific skin condition as "normal" or "abnormal" when compared to a known physiological state. The images can be captured by a digital camera, a smartphone camera, or any other type of sensing element integrated into the instrument. Abnormal or non-physiological conditions are not naturally occurring but can be attributed to biological phenomena such as the growth of cancer cells.

[0024] This invention discloses an iterative process for quantitatively analyzing digital images of melanotic and non-melanotic skin lesions to detect abnormal color distribution. Specifically, the inventors have discovered that an increase in the number of contours classified as having abnormal color distribution during the process indicates the presence of melanoma. A stable or decreasing number of contours classified as having abnormal color distribution corresponds to benign skin conditions, such as moles.

[0025] Color anomalies in a specific image are detected by following a series of different mathematical functions. Figure 1 In step 101, an image segmentation algorithm is first used to define the region of interest (AOI) corresponding to the portion of the image where color anomalies are being checked. Thus, the AOI is separated from the outer regions (OA). This step is crucial for color normalization. In a preferred embodiment, the AOI and OA are converted to grayscale.

[0026] In step 102, the mean and standard deviation of color intensity for the AOI and the region outside the AOI are calculated. This step calculates a risk threshold based on the mean and standard deviation. In step 107, the risk threshold is used to determine whether each of the identified contours has a normal or abnormal color distribution.

[0027] In step 103, the method then calculates the geometric center of the AOI. Since there is only one shape of interest, the centroid and centroid are the same. In a preferred embodiment, the method also calculates the symmetry of the AOI. If the AOI is asymmetrical, it indicates a malignant lesion. In another preferred embodiment, noise at the boundary of the AOI is measured, specifically the peak value around that boundary. This is achieved by first measuring the distance of each pixel at the boundary of the region of interest and then calculating the point where the mean and that distance intersect. An anomalous boundary of the AOI indicates melanoma. In a more preferred embodiment, the method calculates both the symmetry of the AOI and the noise at the boundary of the AOI.

[0028] In step 104, the colors of the AOI are segmented to obtain contours. During this step, a threshold is applied to the image, and the intensity threshold creates one or more contours. In a preferred embodiment, the threshold starts from a first intensity level detected by a global threshold.

[0029] In step 105, if more than one profile was identified in the previous steps, the Euclidean distance of the centroid of each identified profile is calculated, and in step 106, for each identified profile, the sum of the average of all measured distances plus the standard deviation is calculated and then divided by the ratio of the distances of each profile.

[0030] In step 107, the obtained ratio is compared with the risk threshold calculated in step 102, thereby classifying each identified contour as having a normal or abnormal color distribution.

[0031] In a preferred embodiment, the method additionally includes a subsequent step whereby, if a sub-contour is detected within a contour, the average color intensity of the region surrounding the center of the sub-contour and the average color intensity of the region surrounding the center of the corresponding main contour are measured to determine whether the intensity of the sub-contour is significantly different from the intensity of the corresponding main contour. Preferably, the region around the center of the sub-contour and the corresponding main contour where their average color intensity is measured should extend 5 pixels or more in all directions. More preferably, the region around the center of the sub-contour and the corresponding main contour where their average color intensity is measured should extend 5 to 15 pixels in all directions. Sub-contours whose intensity is significantly different from the intensity of the main contour are classified as having an anomalous color distribution.

[0032] Repeat steps 104, 105, 106, and 107 until the size of the contour covers the AOI. During each repetition of step 104, apply different threshold values ​​by increasing them so that the contour covers a larger area each time.

[0033] In step 108, the number of contours (and optional sub-contours) with abnormal color distribution identified during each repetition is calculated. As will be shown in the embodiments, the presence of multiple contours with abnormal color distribution and whose number increases during the process is highly correlated with the presence of melanoma.

[0034] The methods disclosed herein are used to determine the likelihood that a skin lesion is melanoma. Given the serious nature of melanoma and the extreme importance of early detection of the disease, there is a need for a method that allows individuals who may be concerned about having melanoma to obtain rapid initial screening using computer-based image analysis and pattern recognition tools that are easily accessible, for example, via the Internet and can utilize readily available imaging technologies such as smartphone cameras or conventional digital cameras. This invention extends the methods disclosed in the prior art to provide this capability.

[0035] Figure 2This is a flowchart illustrating a preferred embodiment of a computer-implemented method 200 for detecting melanoma in a digital image of tissue. Method 200 is exemplary. In use, method 200 may include additional steps, omit some steps, and / or perform certain steps in a different order. Method 200 is shown as an overview. Numerous details of performing the various steps of method 200 are described in detail herein.

[0036] In step 201, the skin image (such as a dermoscopy image or an image taken by a smartphone camera) used to examine for the presence of lesions is divided into two parts: the area of ​​interest (AOI) and the outer region (OA). The region of interest is unified into a single unit, and the outer part is masked (masked image). Subsequently, both regions are converted to grayscale.

[0037] In step 202, two vectors are generated, one corresponding to the OA and the other to the AOI. The mean and standard deviation of the color intensity for both regions are then calculated, preferably by randomly selecting 60% to 70% of the corresponding vectors. A risk threshold is then calculated based on the mean and standard deviation calculated during this step. A high standard deviation indicates a high probability of unwanted noise, such as blurred images, the presence of hair, freckles, etc. Subsequently, the geometric center or centroid of the AOI and OA is calculated based on pixel intensity.

[0038] In step 203, the symmetry of the AOI is calculated. To define the symmetry of the lesion, an ellipse is fitted over the masked region. If the region is symmetrical, the ellipse and the masked region should converge by approximately 90% or more, preferably 90% ± 2% or more. If the AOI is asymmetrical, it is assigned a risk value indicating an increased probability of malignancy.

[0039] In step 204, noise at the boundary of the AOI is measured, specifically the peak value around that boundary. This is achieved by first measuring the distance of each pixel at the boundary of the region of interest and then calculating the points where the mean and that distance intersect. If more than 5 intersection points are calculated, the AOI has an anomalous boundary, and it is assigned a risk value indicating an increased likelihood of lesions.

[0040] In step 205, a bilateral filter is applied to smooth noise or unwanted details. The applied filter (which is a function of the standard deviation of the AOI) depends on the resolution of the image being analyzed. For example, a higher recommended value is often used when analyzing dermoscopic images, while a lower recommended value is often used when analyzing mobile phone images taken with a high-resolution camera.

[0041] In step 206, the AOI's color is preferably segmented to obtain contours using binary (global) thresholding. During this step, a threshold is applied to the grayscale image, and this grayscale intensity threshold creates one or more contours. The threshold starts from the first intensity level detected by the global threshold.

[0042] In step 207, if more than one contour was identified in the previous steps, the Euclidean distance of the centroid of each identified contour is calculated. In step 208, for each identified contour, the sum of the average and standard deviation of all measured distances is calculated and divided by the ratio of the distances of each contour. In step 209, the resulting ratio is compared with the risk threshold calculated in step 202, thereby classifying each identified contour as having a normal or abnormal color distribution.

[0043] In step 210, if a sub-contour is detected within the main contour, the average color intensity of the region around the center of the sub-contour and the average color intensity of the region around the center of the corresponding main contour are measured to determine whether the intensity of the sub-contour is significantly different from the intensity of the corresponding main contour. Preferably, the region around the center of the sub-contour and the corresponding main contour where their average color intensity is measured should extend 5 pixels or more in all directions. More preferably, the region around the center of the sub-contour and the corresponding main contour where their average color intensity is measured should extend 5 to 15 pixels in all directions. Sub-contours whose intensity is significantly different from the intensity of the main contour are classified as having an anomalous color distribution. As an example, if a sub-contour corresponds to a base color different from the base color of the corresponding main contour, the average color intensity of the sub-contour is considered significantly different from the average color intensity of the corresponding main contour. If the image is grayscale, the average color intensity of the sub-contour is considered significantly different from the average color intensity of the corresponding main contour if the sub-contour corresponds to a different base color within the grayscale range. For example, if the difference between the means of the two regions (sub-contour and main contour) is greater than 30, the average color intensity of the sub-contour is considered significantly different from the average color intensity of the corresponding main contour. Given that there are 8 basic colors (white, black, and the six basic colors derived from the three primary colors), the value 30 appears. Grayscale values ​​range from 0 to 255, so 255 divided by 8 equals 31.2, making white shading correspond to values ​​from 0 to 30, yellow shading to values ​​from 30 to 60, and so on. If the average color intensities of the sub-outlines and the outlines are, for example, 150 and 100 respectively, then they belong to different hue "groups".

[0044] Subsequently, steps 206, 207, 208, 209, and 210 are repeated until the size of the contour covers the AOI. Preferably, steps 206, 207, 208, 209, and 210 are repeated until the size of the contour reaches 75% to 80% of the lesion area (AOI). The value of 75% to 80% corresponds to the optimal percentage because it covers the largest AOI and does not include unnecessary information (such as shadows and dark spots) at the boundaries, which may be caused by imperfect segmentation. Since each skin lesion has different distributions of different colored shadows, the contour convergence is also different. Therefore, the range of the threshold level depends on the object depicted in the image, and for this reason, the range of the threshold level is not always the same. During each repetition of step 206, different thresholds are applied by increasing the value, so that the contour gradually covers a larger area. Preferably, the intensity threshold applied during each repetition of step 206 is increased by a value of 1.

[0045] The evaluation results of the above steps are used to determine whether the color distribution indicates the presence of melanoma. More specifically, in Figure 3 The diagram illustrates the differences between the distributions of contours with anomalous color distributions detected in melanoma and non-melanoma lesions during the iterations of the described process. As the algorithm proceeds from one repetition to another and detects contours, contours classified as having anomalous color distributions (“risk contours”) are plotted relative to the number of repetitions occurring each time throughout the process. In the case of melanoma, an ascending curve is generated as the number of repetitions increases, indicating the observation of an anomalous color distribution in this type of lesion. Figure 3 A). When the lesion is a non-melanotic lesion, the curve shown is stable or declining ( Figure 3 (B) depicts a relatively uniform color distribution in non-melanotic lesions compared to melanotic lesions. Similarly, Figure 4 A series of images illustrates the effect of contour processing when applied to melanocytic nevi ( Figure 4 A), the color spreads evenly, but when applied to melanoma ( Figure 4 B), the colors expand in a random manner.

[0046] During each repetition of steps 206 to 210, a risk value can be calculated to provide a quantitative assessment of the rate of increase in the number of contours with anomalous color distributions during the iteration process. The risk value increases whenever a contour or sub-contour is classified as having an anomalous color distribution. For example, one unit is added to the risk value each time a contour or sub-contour is classified as having an anomalous color distribution. A cumulative risk value is then calculated at the end of the iteration process (step 211), which provides a quantitative assessment of the probability of the presence of melanotic lesions. In another preferred embodiment, the symmetry of the AOI and / or noise at the boundaries of the AOI are also considered when calculating the cumulative risk value. For example, if the AOI is symmetrical and has smooth boundaries, two units can be extracted from the cumulative risk value; if the AOI is asymmetrical and has anomalous boundaries, two units can be added to the cumulative risk value; if the AOI is asymmetrical and has smooth boundaries or symmetrical and has anomalous boundaries, one unit can be extracted from the cumulative risk value.

[0047] If the cumulative risk value is calculated in step 211, then in step 212, the ratio of the cumulative risk value to the total number of repetitions is compared with the classification threshold. The inventors have analyzed a large number of medical images from melanoma and non-melanoma cases (Example 1) and have derived the classification threshold R. c The range is 0.18 to 0.25, (0.18, 0.25). The exact value depends on the image, i.e., whether the image is a dermoscopy image or a smartphone image. The goal is to determine the classification threshold R for the specific image used. c Those skilled in the art will use several known samples (melanotic and non-melanotic) to calibrate the algorithm for the exact value, as explained in Example 1.

[0048] If the ratio of the cumulative risk value to the total number of repetitions is less than the classification threshold R... c If the ratio is higher than R, then the lesion is non-melanotic, while if the ratio is higher than R... c If so, the lesion is melanin-related.

[0049] Example

[0050] Example 1

[0051] Method Calibration-R c Value Calculation

[0052] The methods disclosed in this paper were used to analyze 44 medical images containing 15 known melanoma cases and 29 non-melanoma (nevus) cases. Figure 4 Images are randomly selected to avoid bias in the results. The methodological steps are incorporated into a computer algorithm that analyzes each image independently.

[0053] [1] By applying the KNN algorithm, each image in the analysis is divided into two parts: the region of interest and the outer part (surrounding area). The region of interest is then unified into a whole, and the outer part is masked (masked image).

[0054] [2] Convert both regions to grayscale. Therefore, the pixels of each region are read as a one-dimensional vector. For the conversion, a luminance method is used, and the vector... 图像 =0.299·Red + 0.587·Green + 0.114·Blue. Additionally, the masked image is converted to grayscale using the same method.

[0055] [3] From step [2], two vectors are generated, one corresponding to the external region and the other to the region of interest (OA, AOI). By randomly selecting 60% to 70% of each vector, the mean and standard deviation of each vector are calculated, i.e., (E OA ,σ OA ), (E) AOI , σ AOI These values ​​help determine the condition and intensity texture of the skin (OA) and AOI. Increased σ values ​​correspond to more “noisy” texture. More specifically, “noisy” texture is defined by an uneven color distribution, meaning the skin has many small points of differentiation (e.g., freckles), and these skin areas have a higher risk of developing skin cancer originating from moles. Additionally, E... AOI The lower the value, the darker the color.

[0056] [4] The centroid is defined by using the “image moment” equation on the masked portion, which is the center of the shape / surface of the region of interest (initial centroid).

[0057]

[0058] Where I(x, y) is the pixel intensity corresponding to the (x, y) coordinate.

[0059] [5] Moles or benign lesions tend to be more rounded in shape and are classified as symmetrical. This can also be seen in... Figure 4 Among them, benign skin lesions ( Figure 4 A) and melanoma ( Figure 4 B) Compared to having a more rounded shape. To define the symmetry of the lesion, an ellipse is fitted over the masking region. If the region is symmetrical, the ellipse and the masking region should converge ≥90%. (By El) 面积The area of ​​an ellipse is calculated using the formula π·A·B, where A is the length of the semi-major axis and B is the length of the semi-minor axis. The convergence is calculated as follows:

[0060] [6] Noise is measured at the boundary of the region of interest, specifically the peak value around that boundary. Each pixel on the boundary of the lesion can be considered as a (x, y) 2D spatial vector. Each (x, y) pixel is calculated using the following formula. i ,y i ) point and its lesion center (c x , c y Euclidean distance of )

[0061] Noise can be defined as frequent and dense fluctuations in the distance value distribution, indicating that lesions with five or more peaks have an irregular or complex shape. In contrast, a consistent, smooth lesion will have fewer such fluctuations (four peaks or less).

[0062] Next, the calculated distances are normalized by subtracting the average of all calculated distances. Let... For each normalized result. If the lesion is of perfect shape, we expect The constant is zero (the average of the constant minus the constant equals zero). If the lesion is elliptical, it will look like a triangular diagram with four peaks (two extremes and two minimums). If the lesion's geometry is imperfect, the number of peaks will be greater than five. In the latter case, the boundary is considered anomalous ("noisy"). If This indicates the presence of peaks. The number of peaks is calculated using the following formula:

[0063] Figure 5 It is a graphical representation of the process of calculating the number of peaks around the boundaries of the AOI. Figure 5 A corresponds to the digital image of melanoma, while Figure 5 B corresponds to the digital image of a mole.

[0064] [7] Apply a Gaussian filter to smooth noise or unwanted details. For σ... x ∈(5, 11) and σ y The parameter set ∈ (5,11) depends on the image being analyzed. For example, when analyzing dermoscopy images, σ x and σ y The recommended value is: σ x =11, σy =11. When analyzing images taken with a high-resolution camera from a mobile phone, σ x and σ y The recommended value is low, σ x =7, σ y =7.

[0065] [8] Considering the result of step [3], define the risk threshold R. th .

[0066]

[0067] In the above equation, "O" refers to "external" and "I" refers to "internal".

[0068] [9] The region of interest is segmented by color to obtain a contour, starting with a darker contour and progressing to a lighter one. To achieve this, binary (global) thresholding is used to perform multiple conversions from the smoothed masked image to a binary image (black and white). In this process, multiple thresholds are applied to the grayscale image. The threshold level starts from the first intensity level detected by the global threshold and continues until it reaches 75% to 80% of the lesion area. The range of the threshold level depends on the objects depicted in the image, and for this reason, the range of the threshold level is not always the same.

[0069]

[10] One or more contours are created for each grayscale intensity threshold. In each case where there is more than one contour, the centroid of all contours is measured using the following formula:

[0070] A i It is the area of ​​each contour, and if The formula can be written as:

[0071] The second step of the process involves calculating the Euclidean distance between each contour center and the new center C′.

[0072]

[0073] This creates a one-dimensional array, D=(D′0, D′1, D′2, …, D′). n ), where n is the number of contour centers. Finally, each contour is classified as having a normal or abnormal color distribution:

[0074] And the process returns:

[0075]

[11] For each sub-contour detected within the contour during step

[10] , the region around its center is measured to determine whether it represents a significant change within the main contour. The measured region extends 5 to 15 pixels in all directions from both the contour and its sub-contour region, but if the sub-contour is located at the center of the main contour, another region is randomly selected. Any significant change at this step means that the intensities of the two regions are completely different. If the difference between the means of the two regions (sub-contour and contour) is greater than 30, the average color intensity of the sub-contours is considered to be significantly different. If a sub-contour with a significant change has been detected, a risk value is added to the result of step

[10] , meaning that from that point onwards, the value r is defined as r i+1 =r i +1.

[0076]

[12] Repeat steps

[10] and

[11] until step [9] reaches the desired coverage area. The method returns a risk value r for each iteration. i This means adding a one-dimensional (1D) array. .

[0077]

[13] Finally, the symmetry and boundary from steps [5] and [6] respectively are added as additional features r′ to the resulting 1D array from step

[12] . middle.

[0078]

[0079]

[14] The classification is defined by the following formula:

[0080] k is the number of stimulations required to terminate the method. It is an array The last value of R, and R c It is the threshold for defining the classification.

[0081] To calculate R c The data was divided into a training set (80%) and a test set (20%).

[0082] Starting with Rc values ​​from 0.18 to 0.25, test all values ​​until the lowest Rc value is calculated that classifies the most well-known melanoma cases as melanoma and the percentage of nevus lesions classified as melanoma is found. c value. Figure 6 This is a diagrammatic representation of the process. Figure 6 A, Figure 6 C Figure 6 E and Figure 6 G represents the result from the training set; Figure 6 B Figure 6 D、 Figure 6 F and Figure 6 H represents the result from the test set. The tested R... c The value is 0.25 ( Figure 6 A, Figure 6 B), 0.24 Figure 6 C Figure 6 D), 0.22 Figure 6 E, Figure 6 F) and 0.22 ( Figure 6 G, Figure 6 H).

[0083] Evaluate the Rc value using test data and record the score.

[0084] Repeat the same process multiple times by shuffling the data.

[0085] At the end of the process, R c It is the average score from all trials.

[0086] The result obtained at the end of this process is R c =0.232, ( Figure 6 I) The scoring accuracy was 88%, sensitivity was 86%, and specificity was 89.6%.

[0087] like Figure 7 As shown, the number of contours with abnormal color distributions (“risk contours”) that are repeated each time is plotted relative to the number of repetitions. The following measures are extracted from the output plot: For moles ( Figure 7 A, Figure 7 B) ,and For melanoma ( Figure 7 C Figure 7 D), .

[0088] Example 2

[0089] Digital image analysis for detecting malignant or benign skin lesions .

[0090] Using the image analysis method disclosed herein, the inventors analyzed 600 digital images of skin lesions to determine whether the lesions were melanomas or moles. The method steps were incorporated into a computer algorithm that independently analyzed each image. After calibration in Example 1, the classification threshold R was... c Set to 0.232.

[0091] The results showed that the detection of color abnormalities was correlated with melanoma status. Color abnormalities were detected in 96% of melanoma images, compared to only 10% of non-melanoma images. This corresponds to an accuracy of 93% for the method, which is higher than the accuracy typically achieved by trained medical experts with the naked eye, which rarely exceeds 70%, primarily due to its low level of specificity without a preliminary diagnosis.

Claims

1. A computer-implemented method for detecting melanoma in digital skin images, the method comprising the following steps: a) Determine the region of interest (AOI) in the digital image using an image segmentation algorithm, mask the AOI from the outer region, and convert the masked AOI and the region outside the AOI to grayscale; b) Calculate the mean and standard deviation of the grayscale intensity of the AOI and the region outside the AOI, and extract a risk threshold based on the mean and standard deviation; c) Calculate the geometric center of the AOI, thereby defining the initial centroid; d) Obtain the outline by dividing the color of the AOI by applying a grayscale intensity threshold to the image; e) If multiple contours are identified in step (d), the centroids of all identified contours are measured, and the Euclidean distance from the center of each contour to the centroids of all contours is calculated. f) For each contour identified in step (d), calculate the sum of the average of all measured distances plus the standard deviation, and then divide by the ratio of the distance of each contour to the centroid of all contours identified in step (e). g) The ratio is compared with the risk threshold extracted in step (b), and if the ratio is less than the risk threshold extracted in step (b), the contour is classified as having an abnormal color distribution; as well as h) Repeat steps (d) to (g) by applying multiple thresholds to the image until the size of the contour covers the AOI; The presence of multiple contours with abnormal color distribution and increasing in number during the process indicates melanoma.

2. The method according to claim 1, wherein, If a sub-contour, i.e. a contour within a contour, is detected, step (d) further includes: measuring the average color intensity of a region extending 5 to 15 pixels in all directions around the centroid of the contour, and measuring the average color intensity of a region extending 5 to 15 pixels in all directions around the centroid of the sub-contour; wherein, if the difference between the average values ​​of the two regions is greater than 30, the sub-contour is classified as having an anomalous color distribution.

3. The method according to claim 1 or 2, further comprising: A risk value is calculated during each repetition, wherein the risk value increases whenever the contour and / or sub-contour is classified as having an abnormal color distribution; the cumulative risk value obtained at the end of the process is used to classify the AOI as melanoma or a non-melanoma.

4. The method according to claim 3, wherein, The ratio of the cumulative risk value obtained at the end of the process to the total number of repetitions is compared with a classification threshold ranging from 0.18 to 0.

25. If the ratio is greater than the classification threshold, the ratio indicates melanoma.

5. The method according to any one of the preceding claims, wherein, Step (b) further includes: calculating the symmetry of the AOI, wherein an asymmetric AOI indicates melanoma.

6. The method according to any one of the preceding claims, wherein, Step (b) further includes: measuring the noise at the boundary of the AOI by calculating the Euclidean distance of each pixel at the boundary of the AOI from the initial centroid defined in step (c), i.e. the presence of fluctuations in the distribution of Euclidean distance values ​​at the boundary of the AOI; normalizing the calculated distance by subtracting the average of the calculated distances; and identifying the presence of peaks whenever the normalized distance changes sign between two consecutive points. Wherein, when the number of peaks is greater than five, the boundary of the AOI is abnormal; Furthermore, the abnormal boundaries of the AOI indicate melanoma.

7. An apparatus for detecting melanoma in a digital image of tissue, the apparatus comprising: A camera, the camera being adapted to acquire the image; A processor adapted to process image data according to the method according to any one of the preceding claims; and an output device, the output device being adapted to indicate the likelihood of the presence or absence of melanoma.

8. The apparatus according to claim 7, wherein, The device is a smartphone or tablet.

9. A computer program comprising instructions that, when executed by a computer, cause the computer to perform the method according to claims 1 to 6.

Images