A multi-dimensional detection method based on micro-lesions in CT images of children's kidneys

Abstract

The present application relates to the technical field of microlesion detection, and discloses a multi-dimensional detection method for microlesions in a child kidney CT image, which is based on a constructed child kidney anatomical template database, and uses the multi-dimensional parameters of the microlesions in the kidney CT image obtained by calculation to construct a standard comprehensive quality evaluation model and a dynamic age correction model and obtain a standard score threshold, which is used as an evaluation standard; then, the uncertainty of the standard comprehensive quality evaluation model is quantitatively analyzed to obtain a quality evaluation matrix and a real-time comprehensive quality evaluation model, and the real-time comprehensive quality evaluation model is used to obtain a real-time comprehensive quality score, i.e. a single detection result; finally, the real-time comprehensive quality score is compared with the standard score threshold to realize accurate identification of the microlesions in the child kidney CT image, and solve the technical problems of age difference and lack of consensus mechanism in the identification process.

Description

Technical Field

[0001] This invention relates to the field of minimal lesion detection technology, and in particular to a multi-dimensional detection method for minimal lesions in pediatric renal CT images. Background Technology

[0002] Dynamic contrast-enhanced renal imaging is an important technique for identifying hidden small lesions in the kidneys. Renal CT imaging generates a large amount of image data, and manual image screening is labor-intensive and inefficient. Due to the diversity and complexity of kidney lesions, the quantification of many indicators and the assessment of pathological grading rely on manual counting and semi-quantitative evaluation. The entire evaluation process is time-consuming, labor-intensive, and subject to subjective differences. Furthermore, manual visual evaluation can overlook some minute lesions, leading to missed diagnoses.

[0003] To address the aforementioned technical problems, a method for detecting minute lesions in pediatric renal CT images has been disclosed by those skilled in the art (publication number: CN116167966A). This method involves: labeling and constructing a database of minute lesions in pediatric renal CT images; extracting features of minute lesions in pediatric renal CT images based on a dynamic weighted skip network; locating minute lesions in pediatric renal CT images based on a non-maximum suppression strategy; and outputting lesion results through the dynamic weighted skip network. Thus, the designed dynamic weighted skip network enables automatic extraction of lesion features, and the designed non-maximum suppression strategy enables accurate identification of minute lesions in pediatric renal CT images.

[0004] However, the above-mentioned patents still have the following problems in actual implementation:

[0005] 1. Neglecting age differences: The 300% developmental difference in kidney volume among children aged 1-14 years was not taken into account, resulting in inaccurate assessment standards and failing to meet the precise testing needs of children of different ages;

[0006] 2. Insufficient consensus mechanism: Due to the large differences in the subjective identification of children's kidney CT images, the identification of a single doctor has certain limitations. The method lacks mathematical modeling for the discrepancies in annotations by multiple experts, making it difficult to identify abnormal annotations.

[0007] 3. Lack of assessment and verification capabilities: Because the identification results of small lesions in children's kidney CT images are easily affected by a variety of subjective and objective factors, the real-time detection results have great uncertainty and cannot stably and reliably reflect the lesion condition. There is a lack of verification and assessment capabilities for the detection results.

[0008] Therefore, those skilled in the art need a detection method that can perform multi-dimensional analysis of small lesions in pediatric renal CT images, thereby improving the detection rate of small renal lesions and reducing the rate of missed diagnosis of small renal lesions. Summary of the Invention

[0009] The purpose of this invention is to address the problems of poor reliability and significant discrepancies in existing methods for detecting minute lesions in pediatric renal CT images, and to design a multi-dimensional detection method for minute lesions in pediatric renal CT images.

[0010] The technical solution of the present invention to achieve the above objectives is a multi-dimensional detection method for minute lesions in pediatric renal CT images, the method comprising the following steps:

[0011] Step 1: Construct a database of pediatric kidney anatomy templates and group the pediatric kidney anatomy templates in the database according to the children's age;

[0012] Step 2: Based on the pediatric renal anatomy template database, calculate the spatial anatomical fit, morphological rationality index, expert consensus, and clinical consistency factor of small lesions in renal CT images, and construct a standard comprehensive quality assessment model, which will serve as the basis for lesion detection and outcome evaluation.

[0013] Step 3: Introduce an age correction function to perform nonlinear calibration on the standard comprehensive quality assessment model and obtain a dynamic age correction model. Use the dynamic age correction model to output a standard scoring threshold. The main function of this precise scoring threshold is to serve as a standard for judging the reliability of the test results.

[0014] Step 4: Quantitatively analyze the uncertainty of the standard comprehensive quality assessment model and obtain the quality assessment matrix. Then, use the quality assessment matrix to construct a real-time comprehensive quality assessment model and obtain a real-time comprehensive quality score. The real-time comprehensive quality score will be used as the single detection result of small lesions in children's kidney CT images.

[0015] Step 5: Compare the real-time comprehensive quality score with the standard scoring threshold to assess the reliability of the real-time comprehensive quality score. If the real-time comprehensive quality score is reliable, it will be output as the result of the detection of small lesions in pediatric renal CT images. If the real-time comprehensive quality score is unreliable, it will be canceled and the operation in Step 4 will be repeated until a reliable real-time comprehensive quality score is obtained. Through the above comparison operation, the accurate identification of small lesions in pediatric renal CT images can be achieved.

[0016] The pediatric kidney anatomy template database in step one includes, but is not limited to, annotation masks and anatomical templates; the child's age is no more than fourteen years old.

[0017] In step two, the spatial anatomical fit S C The calculation formula is:

[0018]

[0019] In the formula, M is the annotation mask, and T is the age-matched anatomical template. C M C T To label the coordinates and template centroid coordinates, D T The maximum diameter of the template kidney is given by ω1 = 0.6 and ω2 = 0.4.

[0020] Morphological rationality index M P The calculation formula is:

[0021]

[0022] In the formula, A is the area of ​​the labeled region, P is the perimeter of the labeled region, and σ r / μ r S is the ratio of the standard deviation in the radial direction to the mean. k The sigmoid function is a mathematical transformation function used to measure the similarity to the Hausdorff distance of a standard ellipsoid.

[0023] Expert consensus level E C The calculation formula is:

[0024]

[0025] In the formula, p i Let be the annotation probability graph for the i-th expert. D KL The value is the Kullback-Leibler divergence, where N is the number of pediatric imaging diagnostic experts involved in the annotation and N > 3;

[0026] Clinical consistency factor C f The calculation formula is:

[0027]

[0028] In the formula, TP is the number of real microlesions correctly identified in the annotation results; FP is the number of normal tissue areas incorrectly marked as lesions in the annotation results; FN is the number of real microlesions that were not annotated; and λ1 is the penalty coefficient for missed diagnosis.

[0029] The standard comprehensive quality assessment model in step two is as follows:

[0030]

[0031] In the formula, Q is the standard comprehensive quality score, and α, β and γ are all weight coefficients, which can be dynamically calibrated through a Bayesian optimization algorithm.

[0032] The dynamic calibration model for the weighting coefficients based on the Bayesian optimization algorithm is as follows:

[0033]

[0034] In the formula, θ is the parameter to be optimized, θ=(α,β,γ); θ * To optimize the parameters, Q gold Q is the gold standard score recognized by the expert panel, λ is the standard comprehensive quality score, and λ is the adjustment coefficient.

[0035] The dynamic age correction model in step three is as follows:

[0036]

[0037] In the formula, Q final The standard scoring threshold is 'age', where 'age' is the age of the child being tested.

[0038] The formula for quantifying uncertainty in step four is as follows:

[0039]

[0040] In the formula, U Q Here, ∈ represents the quantitative analysis value of uncertainty, ∈ represents the factor to be eliminated, ∈ = 0.01μ represents the historical mean, and μ s μ is the historical mean of spatial anatomical fit. M μ is the historical average of the morphological rationality index. E This represents the historical average of expert consensus.

[0041] The quality assessment matrix QM in step four is:

[0042]

[0043] The real-time comprehensive quality assessment model in step four is as follows:

[0044] Q = ||QM|| F ·σ max (QM) (10)

[0045] In the formula, ||·|| F For Frobenius norm, σ max It is the largest singular value.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This method constructs a standard model for evaluation and a real-time model for detection based on multi-dimensional influencing factors. At the same time, it fully considers the differences in kidney development among children of different ages, introduces the age mechanism of children into the detection and evaluation process, effectively controls the inaccuracy rate of detection results, and thus better meets the accurate detection needs of children of different ages.

[0048] 2. This method addresses the limitations of individual experts in identifying minute lesions, while also taking into full account the subjective differences among different experts in recognizing minute lesions in pediatric renal CT images. It constructs a mathematical model that satisfies multi-expert annotation and recognition, thereby better meeting the needs for accurate detection of minute lesions.

[0049] 3. This method effectively controls the influence of subjective and objective factors on the test results through comparative analysis, thereby providing a stable and reliable reflection of the lesion condition. Attached Figure Description

[0050] Figure 1 This is a flowchart of a multi-dimensional detection method for minute lesions in pediatric renal CT images, as described in this invention.

[0051] Figure 2 This is a system framework diagram of a multi-dimensional detection method for minute lesions in pediatric renal CT images as described in this invention;

[0052] Figure 3 This is a comparative analysis table of the detection effects of the multi-dimensional detection method for minute lesions in pediatric renal CT images described in this invention and the detection effects of existing minute lesion detection methods;

[0053] Figure 4 This is the age-corrected data table for children's quality score (1-14 years old) described in this invention;

[0054] Figure 5 This is the comparative table for evaluating the quality of annotation of pediatric kidney CT images as described in this invention. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1-5 As shown;

[0056] A multidimensional detection method for minute lesions in pediatric renal CT images, comprising the following steps:

[0057] Step 1: Construct a pediatric kidney anatomy template database and group the pediatric kidney anatomy templates according to the children's age. The process is as follows: Collect CT images of 500 children's kidneys, divide them into 4 groups according to age to construct a pediatric kidney anatomy template database, and the data in the database includes the independent annotation results of 3 senior physicians.

[0058] Step 2: Based on the pediatric renal anatomy template database, calculate the spatial anatomical fit, morphological rationality index, expert consensus, and clinical consistency factor of small lesions in renal CT images. Then, construct a standard comprehensive quality assessment model for small lesions in renal CT images based on the spatial anatomical fit, morphological rationality index, expert consensus, and clinical consistency factor.

[0059] Step 3: Introduce an age correction function to perform nonlinear calibration on the standard comprehensive quality assessment model and obtain a dynamic age correction model. Use the dynamic age correction model to output the standard scoring threshold for small lesions in renal CT images.

[0060] Step 4: Quantify the uncertainty of the standard comprehensive quality assessment model, construct a quality assessment matrix based on the results of the quantitative analysis of the uncertainty of the standard comprehensive quality assessment model, and then use the quality assessment matrix to construct a real-time comprehensive quality assessment model and obtain a real-time comprehensive quality score.

[0061] Step 5: Compare the real-time comprehensive quality score with the standard scoring threshold. When the real-time comprehensive quality score is greater than or equal to the standard scoring threshold, the identification result of small lesions in children's kidney CT images is reliable; when the real-time comprehensive quality score is less than the standard scoring threshold, the identification result of small lesions in children's kidney CT images is unreliable, thus achieving accurate identification of small lesions in children's kidney CT images.

[0062] The innovation of this application lies in constructing a standardized comprehensive quality assessment model by integrating multiple dimensions of factors, including spatial anatomical fit, morphological rationality index, expert consensus, and clinical consistency factors, using a pediatric renal anatomical template. This model serves as the foundation for lesion detection and assessment. By fully considering the age-specific characteristics of children, the standardized comprehensive quality assessment model is optimized to obtain a standard scoring threshold for assessment. This threshold serves as the standard for evaluating detection results and analyzing the reliability of each test result. Uncertainty analysis is then performed on the standardized comprehensive quality assessment model to obtain a real-time comprehensive quality score. This real-time comprehensive quality score serves as the result of detecting minimal lesions in pediatric renal CT images. This result represents only a single detection of minimal lesions in pediatric renal CT images. Finally, the reliability of the single detection result is evaluated through comparative analysis, thereby ensuring the accuracy of detecting minimal lesions in pediatric renal CT images. The comparative analysis of the detection effect of this invention with the detection effect of existing minimal lesion detection methods is as follows: Figure 3 As shown.

[0063] In the technical solution of this application, the pediatric kidney anatomy template database includes, but is not limited to, annotation masks and anatomical templates; and the children are no more than fourteen years old.

[0064] It should be noted that the calculation of spatial anatomical fit S C The formula is:

[0065]

[0066] In the formula, M is the annotation mask, and T is the age-matched anatomical template. C M C T To label the coordinates and template centroid coordinates, D T The maximum diameter of the template kidney is given by ω1 = 0.6 and ω2 = 0.4.

[0067] Calculate the morphological rationality index M P The formula is:

[0068]

[0069] In the formula, A is the area of ​​the labeled region, P is the perimeter of the labeled region, and σ r / μ r S is the ratio of the standard deviation in the radial direction to the mean. k The sigmoid function is a mathematical transformation function used to measure the similarity to the Hausdorff distance of a standard ellipsoid.

[0070] Calculate the degree of expert consensus E C The formula is:

[0071]

[0072] In the formula, p i Let be the annotation probability graph for the i-th expert. D KL The value is the Kullback-Leibler divergence, where N is the number of pediatric imaging diagnostic experts involved in the annotation and N > 3;

[0073] Calculate the clinical consistency factor C f The formula is:

[0074]

[0075] In the formula, TP (True Positives) represents the number of correctly identified true microlesions in the annotation results. It is calculated by comparing the annotated area with the gold standard (pathological biopsy or expert confirmation), focusing on the portion where the spatial overlap between the annotated area and the true lesion (e.g., DICE coefficient ≥ 0.7) and the pathological characteristics are consistent. It primarily reflects the algorithm's ability to capture true lesions. FPFP (False Positives) represents the number of normal tissue areas incorrectly labeled as lesions in the annotation results. It is calculated by the portion where the annotated area does not overlap with the gold standard lesion (DICE coefficient < 0.1) and the pathological verification is negative. Its main function is to measure the risk of overdiagnosis by the algorithm. FNFN (False Positives) represents the number of normal tissue areas incorrectly labeled as lesions in the annotation results. FPFP represents the number of normal tissue areas incorrectly labeled as lesions in the annotation results. It is calculated by the portion where the annotated area does not overlap with the gold standard lesion (DICE coefficient < 0.1) and the pathological verification is negative. FPFP primarily measures the risk of overdiagnosis by the algorithm. Negatives (false negatives) are the number of real but unlabeled minute lesions. They are calculated as the portion of lesions confirmed by the gold standard that are not covered in the labeling results (DICE coefficient < 0.1). Their main function is to reflect the possibility of missed diagnoses by the algorithm. λ1 is the missed diagnosis penalty coefficient, used to adjust the weighting of the impact of false negatives on the score. λ1 = 0.3 indicates that missed diagnoses are highly harmful in pediatric kidney disease (e.g., delayed treatment), requiring a penalty but not excessively reducing the score sensitivity. This is determined through ROC curve analysis. When λ = 0.3, C f The AUC value reached 0.91 (the optimal balance point); its main function is to weaken the contribution of missed diagnoses to the denominator when λ<1, avoid the score being overly sensitive to the number of missed diagnoses, and ensure that the score achieves a balance between the detection rate and the false alarm rate.

[0076] Furthermore, the standard comprehensive quality assessment model constructed is as follows:

[0077]

[0078] In the formula, Q is the standard comprehensive quality score, and α, β and γ are all weight coefficients. The weight coefficients can be dynamically calibrated by the Bayesian optimization algorithm, and the values ​​of the weight coefficients are: α = 0.45, β = 0.3, γ = 0.25.

[0079] The dynamic calibration model for the weighting coefficients, based on the Bayesian optimization algorithm, is as follows:

[0080]

[0081] In the formula, θ is the parameter to be optimized, θ=(α,β,γ); θ * To optimize the parameters, Q gold Q is the gold standard score recognized by the expert panel, λ is the standard comprehensive quality score, and λ is the adjustment coefficient.

[0082] The constructed dynamic age correction model is as follows:

[0083]

[0084] In the formula, Q final The standard scoring threshold is 'age', where 'age' is the age of the child being tested.

[0085] In the technical solution of this application, the formula for quantifying uncertainty is:

[0086]

[0087] In the formula, U Q Here, ∈ represents the quantitative analysis value of uncertainty, ∈ = 0.01, μ is the historical mean, and μ s μ is the historical mean of spatial anatomical fit. M μ is the historical average of the morphological rationality index. E This represents the historical average of expert consensus.

[0088] Taking into account the quantitative analysis results of uncertainty, as well as the calculation results of spatial anatomical fit, morphological rationality index, expert consensus, and clinical consistency factor, the quality assessment matrix QM can be obtained as follows:

[0089]

[0090] Then, based on the quality assessment matrix QM, the real-time comprehensive quality assessment model can be obtained as follows:

[0091] Q = ||QM|| F ·σ max (QM) (10)

[0092] In the formula, ||·|| F For Frobenius norm, σ max It is the largest singular value.

[0093] Example 1:

[0094] The system framework for deploying this method based on a hospital's PACS system includes: a data input layer, a multi-dimensional analysis engine, a dynamic correction module, and a visualization output layer, such as... Figure 2 As shown;

[0095] in,

[0096] The data input layer can access the hospital's PACS system and obtain the right to collect data and communicate. Its main function is to process the collected data into multi-dimensional data (such as spatial anatomical fit, morphological rationality index, expert consensus and clinical consistency factor, etc.) for output.

[0097] The role of the multi-dimensional analysis engine is to receive multi-dimensional data output from the data input layer, analyze the received multi-dimensional data, and then output a real-time comprehensive quality score.

[0098] The function of the dynamic correction module is to receive the real-time comprehensive quality score output by the multi-dimensional analysis engine and the multi-dimensional data output by the data input layer. The dynamic correction module can generate a standard scoring threshold based on the multi-dimensional data, compare the real-time comprehensive quality score with the standard scoring threshold, and output a reliable real-time comprehensive quality score as the detection result based on the comparison structure.

[0099] The visualization output layer can receive the detection results output by the dynamic correction module. Its main function is to visualize the received detection results. The visualization output can be displayed on computers, smartphones, and tablets, among other things.

[0100] It is important to note that Figure 4 The baseline score threshold for the age-corrected children's quality score data table is Q = 0.82 (determined by ROC curve). The peak correction coefficient reaches its maximum value of 1.05 at age 7 (corresponding to the phase angle π / 2 of the sine function), while the nonlinear mapping reflects the growth pattern of children's kidney volume (the number of glomeruli increases by 240% from 1 to 7 years old).

[0101] also, Figure 5 The criteria for quality assessment of CT annotated images of kidneys in children included: a sample size of 200 cases (including 100 cases of lesions <5mm); the equipment used was GE Revolution CT vs Siemens Somatom Force; and the gold standard was ultrasound-guided puncture biopsy.

[0102] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A multi-dimensional detection method for minute lesions in pediatric renal CT images, characterized in that, The method includes the following steps: Step 1: Construct a database of pediatric kidney anatomy templates and group the pediatric kidney anatomy templates in the database according to the children's age; Step 2: Based on the pediatric renal anatomy template database, calculate the spatial anatomical fit, morphological rationality index, expert consensus, and clinical consistency factor of small lesions in renal CT images, and construct a standard comprehensive quality assessment model; Step 3: Introduce an age correction function to perform nonlinear calibration on the standard comprehensive quality assessment model and obtain a dynamic age correction model. Use the dynamic age correction model to output the standard scoring threshold. Step 4: Quantitatively analyze the uncertainty of the standard comprehensive quality assessment model and obtain the quality assessment matrix. Then, use the quality assessment matrix to construct a real-time comprehensive quality assessment model and obtain a real-time comprehensive quality score. Step 5: Compare the real-time comprehensive quality score with the standard scoring threshold to achieve accurate identification of minute lesions in pediatric kidney CT images.

2. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 1, characterized in that, The pediatric kidney anatomy template database in step one includes, but is not limited to, annotation masks and pediatric kidney anatomy templates.

3. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 1, characterized in that, In step two, the spatial anatomical fit S C The calculation formula is: （1） In the formula, M is the annotation mask, and T is the age-matched anatomical template. C M C T To label the coordinates and the centroid coordinates of the template, D T The maximum diameter of the template kidney is given by ω1=0.6, ω2=0.4; Morphological rationality index M P The calculation formula is: （2） In the formula, A The area of ​​the marked region. P The perimeter of the labeled area. σ r / μ r It is the ratio of the standard deviation in the radial direction to the mean. S k The sigmoid function is a mathematical transformation function used to measure the similarity to the Hausdorff distance of a standard ellipsoid. Expert consensus level E C The calculation formula is: （3） In the formula, p i Let be the annotation probability graph for the i-th expert. D KL For Kullback-Leibler divergence, N The number of pediatric imaging diagnostic experts who participated in the annotation and N >3; Clinical consistency factor C f The calculation formula is: （4） In the formula, TP is the number of real microlesions correctly identified in the annotation results; FP is the number of normal tissue areas incorrectly marked as lesions in the annotation results; FN is the number of real microlesions that were not annotated; and λ1 is the penalty coefficient for missed diagnosis.

4. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 3, characterized in that, The standard comprehensive quality assessment model in step two is as follows: （5） In the formula, Q For the standard comprehensive quality score, α, β, and γ are all weighting coefficients, which can be dynamically calibrated using a Bayesian optimization algorithm.

5. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 4, characterized in that, The dynamic calibration model for the weighting coefficients is as follows: （6） In the formula, θ For parameters to be optimized, θ =( α , β , γ ); For the optimized parameters, Q gold This is the gold standard score recognized by the expert panel. Q For the standard comprehensive quality score, λ This is for adjusting the coefficient.

6. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 1, characterized in that, The dynamic age correction model in step three is as follows: （7） In the formula, Q final The standard scoring threshold is 'age', where 'age' is the age of the child being tested. Q This is the standard comprehensive quality score.

7. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 1, characterized in that, The formula for quantifying uncertainty in step four is as follows: （8） In the formula, U Q This is a quantitative analysis value for uncertainty. To prevent the elimination of zero factors, =0.01, μ s This represents the historical average of spatial anatomical fit. μ M The historical average of the morphological rationality index. μ E S represents the historical average of expert consensus. C For spatial anatomical adaptation, M P E is the morphological rationality index. C To achieve expert consensus, C f It is a clinical consistency factor.

8. The multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 7, characterized in that, The quality assessment matrix QM in step four is: （9）。 9. A multi-dimensional detection method for minute lesions in pediatric renal CT images according to claim 8, characterized in that, The real-time comprehensive quality assessment model in step four is as follows: （10） In the formula, F It is the Frobenius norm. σ max It is the largest singular value.

Images