Automatic detection and diagnosis method of lung lesions based on CT image

Abstract

The application provides a lung lesion automatic detection and diagnosis method based on CT images, and relates to the field of medical image assisted diagnosis, and comprises the following steps: S1, multi-modal feature fusion and segmentation are performed on the input CT image lung region; S2, lung nodule detection is realized based on dynamic self-adaptive dilated convolution; S3, the size of the lung nodule is quantified in combination with multi-modal features and an improved Snakes model; S4, dynamic tracking is performed by using a space-time dynamic tracking module; S5, a multi-dimensional feature of the nodule is extracted by a respective feature extraction module, and input is provided for diagnosis decision; and S6, a clinical diagnosis model is used to output a benign / malignant diagnosis result. The method can accurately identify the position of the lung nodule and judge the size of the nodule, thereby completing lung nodule assisted diagnosis, and avoiding problems such as low efficiency, large human subjective factors and low accuracy caused by excessive dependence on artificial.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image assisted diagnosis, and particularly relates to a lung lesion automatic detection and diagnosis method based on CT images. BACKGROUND

[0002] Pulmonary diseases are one of the main diseases threatening human health, among which, pulmonary nodules are a common manifestation of pulmonary diseases; pulmonary nodules, i.e. focal circular or quasi-circular density increase shadows with a diameter of not more than 3cm, can be single or multiple, which are found through pulmonary imaging examination (such as CT images); most pulmonary nodules have no obvious symptoms and are often found accidentally during physical examination or other disease screening. Pulmonary nodules include benign and malignant, among which, benign nodules are usually caused by inflammation (such as pneumonia, tuberculosis) sequelae, intrapulmonary lymph nodes, benign tumors (such as hamartoma), pulmonary alveolar hemorrhage, fibrosis, etc., which are usually smooth in edge, uniform in density, and have no obvious size change in a short period of time, and some may be accompanied by calcification; malignant nodules are usually caused by early lung cancer (such as adenocarcinoma in situ, microinvasive adenocarcinoma), lung metastases (other cancer metastases to the lung), etc., which are usually irregular in edge (spicules, lobulation), uneven in density (containing air bubbles or pleural traction sign), and may increase in size in a short period of time.

[0003] Early screening and treatment of pulmonary nodules can greatly slow down nodule lesions, thereby improving the survival rate of patients; at present, traditional pulmonary nodule detection, recognition and diagnosis mainly rely on observation and analysis of medical images by doctors, which not only has strong subjective factors and high human error rate, is prone to misdiagnosis and missed diagnosis, and has problems such as large workload, low efficiency, high dependence on professional experience of doctors, etc., which prolongs the entire diagnosis time and further causes delayed diagnosis of malignant pulmonary nodules. At the same time, the prior art also has a method of pulmonary nodule auxiliary detection and diagnosis through machine learning algorithm and deep learning algorithm, for example: using a 2D convolutional neural network to extract two-dimensional features of CT slices, and then generating candidate nodules through a region proposal network (RPN), which easily ignores the three-dimensional spatial information of the nodules, is not accurate in positioning the cross-layer nodules (such as nodules with a diameter of more than 10mm), and small nodules (such as nodules with a diameter of less than 3mm) are easily submerged in background noise and lack sensitivity; a 3D convolutional neural network and U-Net segmentation are used to directly process three-dimensional CT data, so as to realize extraction of multi-scale features, but this method has large calculation resource consumption, long training time, and low segmentation accuracy of nodules in anatomically complex regions such as lung apex and paratracheal region. SUMMARY

[0004] In view of the problems in the prior art, the purpose of the present application is to provide a lung lesion automatic detection and diagnosis method based on CT images, which can not only accurately identify the position of lung nodules and accurately judge the size of nodules under the premise of network lightweight, but also effectively complete the auxiliary diagnosis of benign and malignant lung nodules, and avoid the problems of low efficiency, large subjective factors and low accuracy caused by excessive reliance on artificial detection and diagnosis process.

[0005] The purpose of the present application is achieved by the following technical solutions:

[0006] A lung lesion automatic detection and diagnosis method based on CT images, comprising:

[0007] Step S1, multi-modal feature fusion and segmentation are performed on the input CT image lung region;

[0008] Step S2, lung nodule detection is realized based on dynamic adaptive dilated convolution;

[0009] Step S3, multi-modal features and improved Snakes model are combined to realize 3D contouring of lung nodules, so as to quantify the size of lung nodules;

[0010] Step S4, a space-time dynamic tracking module is used to realize dynamic tracking of lung nodules across time frames;

[0011] Step S5, according to the specific differences of benign and malignant nodules, the morphological, density and dynamic features of lung nodules are extracted through the morphological feature module, the density feature module and the dynamic feature module respectively, to provide input for diagnosis decision;

[0012] Step S6, a clinical diagnosis model is used to output the diagnosis results of benign / malignant.

[0013] Based on the further optimization of the above scheme, the multi-modal feature fusion is aimed at lung window (clear edge) and mediastinal window (density sensitive) features, i.e. and mediastinal window feature map (wherein, C represents the number of channels, H , W , D respectively represent the height, width and depth of the feature map), through a cross-window attention mechanism, the feature weights of different window widths are dynamically allocated:

[0014] ;

[0015] In the formula: w lu , w me respectively represent the feature weights of the lung window and the mediastinal window. Convdenotes 1x1x1 convolution dimension reduction; denotes Sigmoid function;

[0016] The lung window feature map and the mediastinal window feature map are fused to obtain a fused feature map:

[0017] .

[0018] Based on further optimization of the above scheme, the encoder of the segmentation network structure in step S1 takes MobileNetV3 as a feature extraction backbone, and the internal structure thereof includes an initial feature layer, a group of intermediate bottleneck modules, and a feature compression layer. The initial feature layer performs preliminary feature extraction on the input lung image through Conv3x3 convolution. The group of intermediate bottleneck modules (bneck) includes a plurality of 3x3 and 5x5 separated convolution blocks (the number of separated convolution blocks is determined according to actual conditions), which are matched with batch normalization (BN) and H-Switch activation function, etc., and are used to extract high-level semantic features (such as the outline and regional pattern of lung tissue). The feature compression layer uses Conv1x1 convolution to replace the fully connected layer, and combines a pooling operation to compress the extracted features into a form suitable for the decoder. The decoder of the segmentation network structure adopts the architecture of “down-sampling-up-sampling-skip connection” of U-Net, and the core thereof is cross-level feature fusion. The encoder generates multi-scale feature maps through “down-sampling and bneck modules”. The decoder restores the feature map size through up-sampling (such as bilinear interpolation and transposed convolution), and fuses the “feature map of the corresponding level of the encoder” and the “feature map after up-sampling of the decoder” through skip connection.

[0019] The loss function of the segmentation network structure adopts a hybrid loss of weighted Dice and cross-entropy:

[0020] ;

[0021] ;

[0022] In the formula: w CE 、 w Dice denote the weight coefficients of cross-entropy and Dice , respectively; p denotes the prediction output of the model (i.e., the prediction probability of the output lung parenchyma category), y denotes the real label.

[0023] Based on further optimization of the above scheme, step S2 is specifically:

[0024] The dynamic adaptive dilated convolution (DADC) module adjusts the dilated rate in real time according to the local feature complexity, including: obtaining the lung area feature map segmented in step S1 The local entropy and gradient norm of each voxel ( x, y, z ):

[0025] ;

[0026] In the formula: K , represents the number of gray scale levels; p i , represents the probability of the gray scale in the 3*3*3 neighborhood; i

[0027] The dynamic dilated rate is obtained by the local entropy and gradient norm r(x, y, z) , avoiding feature rupture caused by excessive dilatation:

[0028] ;

[0029] In the formula: , respectively represent the entropy weight coefficient and the gradient weight coefficient; r 0 represents the basic dilated rate; clip represents the truncation function (1 and 5 are the upper and lower limits of the truncation function, i.e. the dynamic dilated rate is constrained between 1 and 5);

[0030] The detection network module is improved based on RetinaNet , integrating lightweight, adaptive feature extraction and complex sample optimization, including 3D MobileNetV3 as the backbone network (i.e. Backbone) for feature extraction; using the feature pyramid network (FPN) and the dynamic adaptive dilated convolution (DADC) module as the neck structure (i.e. Neck) of the network, for multi-scale fusion and adaptive complex morphology capture; using the double-branch head structure of RetinaNet as the detection head (i.e. Heads) of the network, including the classification head and the regression head, wherein the classification head uses the improved Focal Loss to predict the probability of each anchor box outputting lung nodules or background:

[0031] ;

[0032] In the formula: p t , represents the prediction probability of the model that the sample belongs to lung nodules; , represents the class weight; , represents the focusing parameter.

[0033] Based on the further optimization of the above scheme, the step S3 is specifically:

[0034] ​The energy function of the improved Snakes model includes multi-modal external energy (fusing lung window edge and mediastinum window density), i.e. including internal energy E int , external energy E ext and constraint energy E con

[0035]

[0036]

[0037] wherein: a(s) , b(s) denote smoothing coefficients respectively used to control bending smoothness and stretching smoothness of the contour; U denote contour vertices, i.e. points on the 3D active contour, used to describe geometric position of the active contour; k 1, k 2 respectively denote corresponding weight coefficients; denote average density of the nodule; C z denote centroid of the candidate nodule; R yue denote constraint radius, avoiding the contour overflowing the candidate nodule region;

[0038] After convergence of the improved Snakes model, a continuous 3D contour is obtained, which is converted into voxel representation of the 3D image, and the number of voxels inside the Snakes contour is obtained by traversing voxels of the 3D image N ts ; and the volume of the nodule is calculated by the number of voxels inside the Snakes contour V F and the equivalent diameter d F

[0039]

[0040] wherein: d x , d y , d z denote voxel ratio of CT respectively.

[0041] Based on further optimization of the above scheme, the step S4 is specifically:

[0042] First, a feature vector of each nodule is constructed F n

[0043] ​​​​​​ ;

[0044] wherein, f DADC represents the output feature of a dynamic adaptive dilated convolution (DADC) module; cx 、 cy 、 cz represents the centroid coordinate;

[0045] After that, Kalman filtering is used for state prediction and observation update:

[0046] State prediction:

[0047] ;

[0048] wherein: represents the prior state estimation of the current t frame; A represents a state transition matrix (describing the transition rule of the system state from t -1 frame to t frame); C t-1 represents the posterior state estimation of the t -1 frame; B represents a control input matrix; represents a control input vector; represents the prior estimation covariance matrix of the current t frame; P t-1 represents the posterior estimation covariance matrix of the t -1 frame; Q represents a process noise covariance matrix:

[0049] ;

[0050] wherein: I represents an identity matrix;

[0051] Observation update:

[0052] ;

[0053] wherein: K t represents a Kalman gain; O represents an observation matrix ( O = I ); Z r Observation noise covariance matrix:

[0054] ;

[0055] Then, the fusion feature and position similarity, cross time frame node feature matching is carried out:

[0056] ;

[0057] In the formula: S(n t-1,i ,n t,j ) The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; t The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; i The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; t The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; j The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; t The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; i The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; t The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; j The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; i The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; C t,j The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; t The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; j The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity; The first node of the first (t-1) frame and the first node of the first (t) frame are fused similarity;

[0058] Through feature matching, if a node is matched, the change trend (such as volume growth rate) of the node is output; if a node is not matched, a new node is marked, and the node in the previous frame is marked as disappeared.

[0059] Based on the further optimization of the above scheme, the morphological feature module includes edge irregularity calculation and pleural traction sign detection:

[0060] Edge irregularity calculation:

[0061] The node contour obtained by improving the Snake model U ={ u 1, u 2,…, u N}, u i The contour fluctuation under multiple scales (i.e. 2mm, 4mm, 6mm) is obtained:

[0062] ;

[0063] In the formula: L i+m The contour fluctuation under multiple scales (i.e. 2mm, 4mm, 6mm) is obtained: u i The contour fluctuation under multiple scales (i.e. 2mm, 4mm, 6mm) is obtained: ui+m the distance to the line, dist denotes the vertex u i the distance to the line; denotes the scale m the average distance under the scale;

[0064] Pleural traction detection: extract the shortest distance from the nodule contour to the pleura D ple , obtain the nearest point of the node u min , so as to obtain the traction direction and calculate the directional gradient consistency of the traction direction;

[0065] The density feature module includes multi-modal density entropy and calcification specificity detection:

[0066] Multi-modal density entropy fusion lung window F lu and mediastinal window F me The density distribution is obtained, and the joint entropy of the nodule region is obtained:

[0067] ;

[0068] In the formula: p(i, j) denotes the joint probability of the lung window gray scale i and the mediastinal window gray scale j in the nodule region;

[0069] If MDE>1.2, it indicates that the density is uneven;

[0070] Calcification specificity detection: adopt a lightweight decoder and combine calcification morphological features (such as circularity, high density, etc.) to realize calcification judgment;

[0071] The dynamic feature module includes core dynamic index acquisition and dynamic risk index fusion:

[0072] The core dynamic index includes the volume change rate, the density change rate and the morphological change rate of the previous frame (i.e. the previous stage) t 0 and the current t 1 two time points:

[0073] ;

[0074] In the formula: △t represents the interval time of two frames of detected CT images, V t0 , V t1 respectively represent the volume of the previous frame and the current nodule; respectively represent the average density of the previous frame and the current nodule; MI t0, MI t1 These represent the contour fluctuations of the previous frame and the current nodule, respectively.

[0075] Dynamic risk index fusion:

[0076] ;

[0077] In the formula: These represent the weighting coefficients for changes in volume, density, and morphology, respectively.

[0078] Based on further optimization of the above scheme, step S6 specifically includes:

[0079] First, combining the features of morphology, density, and dynamics from step S5, a feature vector for a single nodule is constructed:

[0080] ;

[0081] in: TS This indicates pleural traction. CS Indicates calcification specificity;

[0082] Then, a trained binary classification model (such as logistic regression, support vector machine, etc.) is used for prediction. F JJ-k Taking the input as input and benign / malignant labels as output, a predicted probability is generated. Based on the predicted probability and the true label, the AUC value of the corresponding feature is calculated. AUC k ;

[0083] Then, based on clinical benchmark weights w cilm-k ( k = MI, TS, MDE, CS, DRI ) and feature discrimination AUC k Obtain the attention weights for each feature. W k :

[0084] ;

[0085] Obtain fusion feature values:

[0086] ;

[0087] The clinical diagnostic model is embedded in the detection network module in step S2. Its feature extraction backbone reuses the nodule feature branch of the detection network module. Simultaneously, the clinical diagnostic model adds two output heads to the detection network module: a new classification head and a feature annotation head. The new classification head outputs the probability of "malignancy" through a sigmoid activation function.P mal ; the feature annotation head outputs the contribution degrees of five key features (i.e. morphological irregularity MI , pleural traction TS , density entropy MDE , calcification CS , dynamic risk DRI ) GX k GX k ∈[0, 1]; at the same time, the clinical diagnosis model adopts a multi-task loss function Loss dia

[0088]

[0089] , wherein: w 1, w 2 respectively represent corresponding weight coefficients; CE represents a cross-entropy loss, y mal represents a benign or malignant label; represents a feature contribution degree label; and MSE represents a mean square error loss.

[0090] If P mal <0.1, it is represented as extremely low risk; if 0.1≤ P mal <0.3, it is represented as low risk; if 0.3≤ P mal <0.6, it is represented as medium risk; if 0.6≤ P mal <0.8, it is represented as high risk; and if P mal ≥0.8, it is represented as extremely high risk.

[0091] The technical effects of the scheme are as follows:

[0092] The scheme fuses multi-modal features and adopts a lightweight network (i.e. based on RetinaNet ​​​The cooperation of the improved detection network module and the targeted loss function effectively solves the problems of single feature, large background interference, incomplete lung area coverage, and fuzzy lung area edge in the existing CT image segmentation, and realizes high integrity and high precision segmentation of the lung area. Meanwhile, the dynamic adaptive dilated convolution module is embedded in the detection network module, the dilated rate is adjusted in real time according to the local feature complexity, and the problems such as'small nodule missing and large nodule feature breaking' are effectively avoided. Through the feature pyramid network, shallow fine-grained features and deep semantic features are fused to generate multi-scale feature maps, so as to cover the full-size detection requirements of nodules (such as 2-3mm small nodules and 10mm or more large nodules), and further avoid the problems of missed detection or feature loss. The improved Snakes model is used to realize the outlining and quantification of the 3D contour of the nodule, which can accurately capture the 3D geometric shape of the nodule (including lobulated, irregular edge, etc.), so that the coincidence degree of the outlined 3D contour and the real nodule is more than 90%, and the nodule volume and equivalent diameter are obtained by converting the 3D contour into voxel representation, which reduces the calculation amount and effectively avoids the problems of contour deviation or large volume calculation error in the nodule quantification process.

[0093] The spatial and temporal dynamic tracking module is used for dynamic tracking of nodule changes, which effectively solves the problems that the existing static detection cannot track the changes of nodules, and it is difficult to identify the newly added or disappeared nodules, so as to effectively deal with the problems of nodule position deviation caused by CT image shooting angle, respiratory motion, and large prediction nodule position error. Through the multi-feature module, the shape, density and dynamic three-dimensional features of the nodule are extracted, and the attention weight and risk grading are combined to realize the cooperation of intelligent auxiliary decision and professional judgment of doctors, accurately output the detection and diagnosis results of CT images, and effectively reduce the risk of misdiagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0094] Figure 1 The structure block diagram of the lung lesion automatic detection and diagnosis method in the embodiment of the application. DETAILED DESCRIPTION

[0095] The technical solutions in the embodiments of the application will be described clearly and completely below. In the following description, specific details such as specific system structures, technologies, etc. are proposed for explanation but not for limitation, so as to thoroughly understand the embodiments of the application.

[0096] Embodiment 1:

[0097] A lung lesion automatic detection and diagnosis method based on CT images, comprising:

[0098] Step S1, multi-modal feature fusion and segmentation are performed on the lung area of the input CT image (the input CT image includes multi-modal data of lung window, mediastinum window and thin layer 3D sequence);

[0099] Multi-modal feature fusion is directed to lung window (clear edge) and mediastinum window (density sensitive) features, i.e., lung window feature map and mediastinum window feature map wherein, C represents the number of channels, H , W , D respectively represent the height, width and depth of the feature map), through a cross-window attention mechanism, dynamically allocate feature weights of different window widths:

[0100] ;

[0101] In the formula: w lu , w me respectively represent the feature weights of the lung window and the mediastinum window; Conv represents a 1x1x1 convolutional dimension reduction; represents a Sigmoid function;

[0102] Then the lung window feature map and the mediastinum window feature map are fused to obtain a fused feature map:

[0103] .

[0104] The encoder of the segmentation network structure takes MobileNetV3 as the feature extraction backbone, and its internal structure includes an initial feature layer, a group of intermediate bottleneck modules, and a feature compression layer. The initial feature layer performs preliminary feature extraction on the input lung image (such as a CT image) through a Conv3x3 convolution. The group of intermediate bottleneck modules (bneck) includes multiple 3x3 and 5x5 separated convolution blocks (the number of separated convolution blocks is determined according to actual conditions), which are matched with batch normalization (BN) and H-Switch activation functions, etc. The deeper the level, the more bneck modules there are, which are used to extract high-level semantic features (such as the outline and regional pattern of lung tissue). The feature compression layer uses a Conv1x1 convolution to replace the fully connected layer, and combines a pooling operation to compress the extracted features into a form suitable for the decoder. The decoder of the segmentation network structure adopts the architecture of "down-sampling-up-sampling-skip connection" of U-Net, and its core is cross-level feature fusion. The encoder generates multi-scale feature maps through "down-sampling and bneck modules", and the decoder restores the feature map size through up-sampling (such as bilinear interpolation and transposed convolution), and fuses the "feature map of the corresponding level of the encoder" with the "feature map after up-sampling of the decoder" through skip connections;

[0105] For the imbalance between lung tissue and background categories (e.g., lung tissue accounts for a much smaller proportion of CT images than background), the loss function of the segmentation network structure adopts a hybrid loss of weighted Dice (based on region overlap calculation, which can effectively optimize whether the segmentation region completely covers the real lung tissue) and cross-entropy (based on probability distribution optimization, which can improve the accuracy of whether each pixel is correctly classified as lung tissue or background):

[0106] ;

[0107] ;

[0108] In the formula: w CE , w Dice are weight coefficients of cross-entropy and Dice , respectively (generally w CE = 0.3, w Dice = 0.7); p represents the prediction output of the model (i.e., the prediction probability of outputting lung parenchyma category), y represents the true label (for example: y= 1 indicates that the position is lung parenchyma, y = 0 indicates that the position is background).

[0109] Step S2, based on dynamic adaptive dilated convolution, realizes the detection of lung nodules, specifically:

[0110] The dynamic adaptive dilated convolution (DADC) module adjusts the dilated rate in real time according to the local feature complexity, including: obtaining the lung region feature map segmented in step S1 The local entropy and gradient norm of each voxel ( x, y, z ):

[0111] ;

[0112] In the formula: K represents the number of gray levels (generally K = 16); p i represents the probability of gray i in a 3x3x3 neighborhood;

[0113] The dynamic dilated rate r(x, y, z) is obtained by local entropy and gradient norm to avoid feature breakage caused by excessive dilation:

[0114] ;

[0115] In the formula: , respectively represent the entropy weight coefficient and the gradient weight coefficient (generally ); r 0 represents the base inflation rate (generally r 0=1); clip represents a truncation function (1 and 5 are the upper and lower limits of the truncation function, that is, the constraint of the dynamic inflation rate is between 1 and 5);

[0116] The detection network module is improved based on RetinaNet , which integrates lightweight, adaptive feature extraction and complex sample optimization, including 3D MobileNetV3 as the backbone network (i.e. Backbone) for feature extraction. 3D MobileNetV3 uses depth separable convolution to greatly reduce the parameter amount, and at the same time it adapts to the 3D body data of the lung CT (including height, width and depth dimensions), extracts multi-scale basic features from the 3D image, and provides initial feature support for subsequent nodule detection. The feature pyramid network (FPN) and the dynamic adaptive dilated convolution (DADC) module are used as the neck structure (i.e. Neck) of the network, which is used for multi-scale fusion and adaptive complex morphology capture; among them, FPN fuses "shallow fine-grained features (for small nodules)" and "deep semantic features (for large nodules)", generates multi-scale feature maps, and solves the detection needs of different size lung nodules; DADC adjusts the inflation rate in real time according to the "feature complexity" of the local nodule. The double-branch head structure of RetinaNet is used as the detection head (i.e. Heads) of the network, including the classification head and the regression head, and the classification head uses the improved Focal Loss to predict the probability of each anchor box outputting lung nodules or background:

[0117] ;

[0118] In the formula: p t represents the prediction probability of the model that the sample belongs to the lung nodule; represents the class weight (generally 0.25); represents the focus parameter (generally 2).

[0119] Step S3, combine multi-modal features and improved Snakes model to realize 3D contouring of lung nodules, so as to quantify the size of lung nodules; Specifically:

[0120] The energy function of the improved Snakes model includes multi-modal external energy (fusion of lung window edge and mediastinal window density), that is, internal energy E int , external energy E ext and constraint energy E con :

[0121] ;

[0122] ;

[0123] In the formula: a(s) 、 b(s) represents the smoothness coefficient, respectively, for controlling the bending smoothness and stretching smoothness of the contour; U represents the contour vertex, i.e. the point on the 3D active contour, for describing the geometric position of the active contour; k 1、 k 2 respectively represent the corresponding weight coefficients (generally k 1=0.7、 k 2=0.3); represents the average density of the nodule (with the density reference value of the nodule tissue being pre-statistically or calculated); C z represents the centroid of the candidate nodule; R yue represents the constraint radius, avoiding the contour from overflowing the candidate nodule area;

[0124] After the improved Snake model converges, a continuous 3D contour is obtained, which is converted into a voxel representation of the 3D image (if it is a parameterized Snake contour, a 3D surface composed of a group of control points is fitted by using a triangular patch, a discrete geometric representation of the contour is obtained, and the spatial range of the surface in the 3D voxel grid is determined; if it is a zero level set of the level set method, the 0 surface of the level set function, i.e. the target contour, is found, and the relative positions of the voxels and the contour are directly associated by using the characteristics of the signed distance function, i.e. the function value is negative, indicating that the voxel is inside the contour), the voxels inside the Snake contour are obtained by traversing the voxels of the 3D image N ts (for example: using bounding box acceleration: first determine the axial bounding box of the contour, only traverse the voxels inside the bounding box to reduce the calculation amount. Voxel internal judgment: ray method: from the center of the voxel, a ray is emitted in any direction, the number of intersection points with the contour surface is counted, if the number of intersection points is odd, the voxel is inside the contour; level set method: if the "level set function value of the voxel <0", it indicates that the voxel is inside the contour); and the volume of the nodule is calculated by the number of voxels inside the Snake contour V F and the equivalent diameter d F :

[0125] ;

[0126] In the formula: d x 、 d y ,d z This indicates the voxel resolution of CT.

[0127] Step S4: Employ a spatiotemporal dynamic tracking module to achieve dynamic tracking of lung nodules across time frames; specifically:

[0128] First, construct the feature vector for each nodule. F n :

[0129] ;

[0130] in, f DADC This represents the output features of the Dynamically Adaptive Dilated Convolution (DADC) module; cx , cy , cz Indicates the coordinates of the centroid;

[0131] Then, Kalman filtering is used for state prediction and observation updates:

[0132] State prediction:

[0133] ;

[0134] In the formula: Indicates the current number t Prior state estimation of the frame; A represents the state transition matrix (describing the system state from...) t -1 frame to t Regarding the frame transition pattern, in this application, frame t-1 can be a previously captured CT image, obtained by calling historical data or providing historical data from the patient, and frame t is the currently captured CT image (the same applies below). C t-1 Indicates the first t Posterior state estimation for frame -1; B This represents the control input matrix (usually 0, indicating no control input). Represents the control input vector; Indicates the current number t The prior estimate of the covariance matrix of the frame; P t-1 Indicates the first t The posterior estimated covariance matrix of frame -1; Q Representing the process noise covariance matrix:

[0135] ;

[0136] In the formula: I Represents the identity matrix;

[0137] Observation Update:

[0138] ;

[0139] In the formula: K t Indicates Kalman gain; O Represents the observation matrix ( O = I ); Z r Observation noise covariance matrix:

[0140] ;

[0141] Then, by fusing features and positional similarity, feature matching of nodes is performed across time frames:

[0142] ;

[0143] In the formula: S(n t-1,i ,n t,j ) Indicates the ( t -1) frame i The nodule and the first t The first frame j The fusion similarity of individual nodules; Indicates the ( t -1) frame i Nodule characteristics With the t The first frame j Nodule characteristics Cosine similarity; Represents the t-th frame i Predicted location of each nodule; C t,j Indicates the first t Frame number j The actual detection location of each nodule; Indicates the position weight (usually 2);

[0144] If a nodule is matched through feature matching, the trend of the nodule's change (such as the volume growth rate) is output; if no nodule is matched, the nodule is marked as newly added and the nodule in the previous frame is marked as disappeared.

[0145] Step S5: Based on the specific differences between benign and malignant nodules, extract the morphological, density, and dynamic features of lung nodules through the morphological feature module, density feature module, and dynamic feature module, respectively, to provide input for diagnostic decision-making;

[0146] The morphological feature module includes edge irregularity calculation and pleural traction sign detection:

[0147] Edge irregularity calculation:

[0148] Nodule contour obtained by improved Snake model U ={ u 1, u 2,…, u N}, u i denotes the contour vertex coordinates, and contour fluctuations under multiple scales (i.e., 2 mm, 4 mm, 6 mm) are obtained:

[0149] ;

[0150] In the formula: L i+m denotes the vertex u i and the line connecting the vertex u i+m , and dist denotes the distance from the vertex u i to the line; denotes the average distance under the scale m ( MI The higher the value, such as greater than 0.8 mm, indicates that the edge is irregular, and has a malignant tendency);

[0151] Pleural traction sign detection: extract the shortest distance from the nodule contour to the pleura D ple , obtain the nearest point u min of the node, so as to obtain the traction direction and calculate the directional gradient consistency of the traction direction; specifically:

[0152] First, a deep learning method (such as 3D DeepLab ) is used to segment the lung parenchyma, and a three-dimensional body data of the lung parenchyma is obtained, and the Marching Cubes algorithm is used to extract the outer surface to obtain a triangular mesh representation of the pleura; traverse the nodule vertex, and for each vertex u i ∈ U , first calculate the shortest distance from the vertex to the pleura, and for the vertex u i , find the nearest point u i in the pleural mesh: zj i

[0153] ;

[0154] In the formula: x ui ​,y ui ,z ui x zji ,y zji ,z zji represent the three-dimensional coordinates of the vertex u i and the nearest point zj i respectively;

[0155] The minimum distance in all vertices, i.e. the shortest distance D ple is obtained:

[0156]

[0157] Then, the nodule nearest point D ple and the pleura nearest point u min (i.e. the two end points of the shortest distance) are obtained, so as to obtain the pulling direction vector zj min : Xu

[0158]

[0159] Taking the nodule nearest point u min as the center, a small three-dimensional neighborhood (such as a 3x3x3 voxel block) is taken, and a three-dimensional Sobel operator is used to obtain the gradient of each voxel :

[0160]

[0161] In the formula: represent the gradients in x, y and z directions (obtained by convolution of Sobel kernel and image);

[0162] Finally, the consistency of the gradient and the pulling direction is calculated: for the gradient of each voxel in the neighborhood, the cosine similarity is used to measure the consistency of the direction with the pulling direction vector Xu , and finally the neighborhood mean value is taken as the feature:

[0163]

[0164] In the formula: M represents the number of neighborhood voxels; J con represents the neighborhood mean value. ​​​​​​

[0165] The density feature module includes multimodal density entropy and calcification-specific detection:

[0166] Multimodal density entropy fusion lung window F lu With lattice windows F me The density distribution is used to obtain the joint entropy of the nodule region:

[0167] ;

[0168] In the formula: p(i, j) Indicates the gray level of the lung window within the nodule area. i grayscale of longitudinal diaphragm windows j The joint probability;

[0169] If MDE > 1.2, it indicates uneven density (malignant tendency).

[0170] Calcification-specific detection: A lightweight decoder combined with calcification morphology features (such as circularity, high density, etc.) is used to determine calcification. This includes a lightweight calcification decoder, calcification morphology feature extraction, and multi-feature fusion for calcification determination. The lightweight calcification decoder is based on the encoder output of the original segmentation network structure. A U-shaped upsampling structure with depthwise separable convolution kernels reduces the number of parameters. The decoder uses a U-shaped symmetrical structure and multi-scale feature skip connections with the encoder. Each layer uses depthwise separable 3D convolution (i.e., channel-wise convolution). The last layer of the decoder uses a sigmoid function to output a probability map of each voxel or pixel belonging to calcification. P calc ;

[0171] Calcification morphology feature extraction combines the high-density attribute of CT images with the circular shape of calcifications to extract high-density and circularity features:

[0172] High-density features: First, extract the HU value (Hunsfield Unit, i.e., the unit of CT value) feature map from the CT image. F HU And set a density threshold. T HU Generate high-density region mask M high :

[0173] ;

[0174] Circularity feature extraction: for probability maps P calc and M high The fusion mask is used to extract each connected region Ω, and the area S of the connected regions is calculated. Ω and perimeter CΩ , then the circularity is:

[0175] ;

[0176] The closer the shape is to irregular, the closer the circularity is to 0;

[0177] The multi-feature fusion of calcification is a combination of calcification probability, high density and circularity, and the probability threshold P prob , such as 0.5, the density threshold T HU , such as 100HU, and the circularity threshold CIR c , such as 0.7, to obtain the final calcification region:

[0178] ;

[0179] In the formula: Ω( x, y, z ) represents a connected region containing coordinates ( x, y, z ).

[0180] The dynamic feature module includes core dynamic index acquisition and dynamic risk index fusion:

[0181] The core dynamic index includes the volume change rate, the density change rate and the shape change rate of the previous frame (i.e. the previous period) t 0 and the current t 1 two time points:

[0182] ;

[0183] In the formula: △t represents the interval time of two frames of detected CT images, V t0 , V t1 respectively represent the volume of the previous frame and the current nodule (the volume change rate greater than 10% / month indicates malignant growth); respectively represent the average density of the previous frame and the current nodule (the density change rate greater than 5% indicates that the density is not uniform and intensifies); MI t0 , MI t1 respectively represent the contour fluctuation of the previous frame and the current nodule (the shape change rate greater than 20% indicates that the edge is irregular and intensifies);

[0184] Dynamic risk index fusion:

[0185] ;

[0186] In the formula: respectively represent the weight coefficients of volume change, density change and shape change (obtained according to clinical experience, generally ).

[0187] Step S6, using the clinical diagnosis model, output the benign / malignant diagnosis result, specifically:

[0188] First, combined with the features of the three dimensions of shape, density and dynamics in step S5, the feature vector of a single nodule is constructed:

[0189] ;

[0190] Among them: TS represents the pleural traction, CS represents the calcification specificity;

[0191] Then, using the trained binary classification model (such as logistic regression, support vector machine, etc.) for prediction, taking F JJ-k as input and benign / malignant label as output, the prediction probability is generated, and the AUC value of the corresponding feature is calculated based on the prediction probability and the real label, that is AUC k ;

[0192] Then, based on the clinical benchmark weight w cilm-k ( k = MI, TS, MDE, CS, DRI , for example: w cilm-MI =0.2, w cilm-TS =0.25, w cilm-MDE =0.15, w cilm-CS =0.1, w cilm-DRI =0.3) and the feature discriminability AUC k , the attention weight W k of each feature is obtained:

[0193] ;

[0194] The fusion feature value is obtained:

[0195] ;

[0196] The clinical diagnostic model is embedded in the detection network module in step S2. Its feature extraction backbone reuses the nodule feature branch of the detection network module. Simultaneously, the clinical diagnostic model adds two output heads to the detection network module: a new classification head and a feature annotation head. The new classification head outputs the probability of "malignancy" through a sigmoid activation function. P mal The feature annotation header outputs five key features (i.e., morphological irregularity). MI pleural traction TS Density entropy MDE Calcification CS Dynamic risks DRI ) contribution GX k ( GX k (∈[0,1]); Meanwhile, the clinical diagnostic model employs a multi-task loss function. Loss dia :

[0197] ;

[0198] In the formula: w 1. w 2 represents the corresponding weight coefficients ( w 1 = 0.6 w 2=0.4); CE represents the cross-entropy loss. y mal Indicates benign or malignant (if benign, then...). y mal =0, if it is malignant, then y mal =1); The feature contribution label (pre-labeled by professionals) represents the mean squared error loss.

[0199] like P mal <0.1 indicates extremely low risk; if 0.1≤ P mal <0.3 indicates low risk; if 0.3≤ P mal <0.6 indicates medium risk; if 0.6≤ P mal <0.8 indicates high risk; if P mal A value of ≥0.8 indicates an extremely high risk.

[0200] Furthermore, it can be combined with clinical practice to correspond key characteristics of risk levels (facilitating doctors' observation and comparison, further pinpointing the location of lesions), and providing clinical recommendations:

[0201] For example:

[0202] The key feature condition corresponding to very low risk is CS=1 and DRI<1%, and the clinical recommendation is annual follow-up; the key feature condition corresponding to low risk is MI<0.5mm and DRI<3%, and the clinical recommendation is half-year follow-up; the key feature condition corresponding to medium risk is single feature abnormality (such as TS=1 or DRI not less than 5%), and the clinical recommendation is three-month follow-up+enhanced CT; the key feature condition corresponding to high risk is two feature abnormalities (such as TS=1 and MDE>1.2), and the clinical recommendation is biopsy+PET-CT review; and the key feature condition corresponding to very high risk is three or more feature abnormalities (such as TS=1 and DRI=8% and MI>1mm), and the clinical recommendation is expert consultation evaluation surgery.

[0203] Embodiment 2:

[0204] As another preferred embodiment of the scheme of the present application, on the basis of the scheme of embodiment 1, after outputting the diagnosis result of benign / malignant in step S6, the consistency of the current diagnosis result with the historical trend is verified in combination with the historical data of the spatiotemporal dynamic tracking module in step S4, to exclude the interference of accidental errors (for example, density unevenness caused by single CT artifact), assuming that the diagnosis results of T times of history are P mal,1 , P mal,2 ,…, P mal,T} and the current diagnosis result is P mal,now , then the consistency is judged as:

[0205] ;

[0206] In the formula: represents the historical result weight (set according to the time interval, recent result is 1.2 and long-term result is 0.8);

[0207] If Consist>0.7, the result is consistent;

[0208] If Consist<0.7 and P mal >0.6, output the result that needs to be reviewed (through manual review or reanalysis of the result);

[0209] If Consist>0.7 and , in the formula, represents the average value of the "malignant prediction probability" of the nodules in the historical multiple examinations, reflecting the historical level of the malignant tendency of the nodules; then output the suggestion of "increased malignant risk, suggesting shortening the follow-up interval".

[0210] Embodiment 3:

[0211] As another preferred embodiment of the present application, on the basis of the scheme of Embodiment 1, a "double-teacher-single-student" distillation architecture is adopted to distill the detection in step S2 and the quantization model in step S3, so as to balance the accuracy and speed. The specific distillation architecture is as follows:

[0212] The teacher model adopts a 3D ResNET50 model for detection and a 3D U-Net model for quantization.

[0213] The student model adopts a 3D MobileNetV3 for detection and a lightweight Snakes model for quantization.

[0214] The total loss function is as follows:

[0215] ;

[0216] In the formula: respectively represent the weight coefficients of the corresponding loss functions (in this embodiment, ); Loss har represents a hard loss, which is used to supervise the classification or segmentation accuracy of the student model on the real label; Loss sof represents a soft loss, which is used to let the student model learn the feature distribution knowledge (i.e. soft label) of the teacher model; Loss qua represents a quantization loss, which is used to supervise the consistency of the student model with the teacher model in size quantization;

[0217] ;

[0218] In the formula: S cls represents the classification output of the student model, y cls represents the real classification label (such as benign / malignant classification); S seg represents the segmentation output of the student model, y seg represents the real segmentation label (such as segmentation of lung nodule region and background); KL represents the KL divergence (Kullback-Leibler Divergence), which is used to measure the difference between the student feature distribution and the teacher feature distribution; S feat represents the intermediate feature (for example: the feature map of a certain layer) of the student model, T featrepresents the intermediate feature corresponding to the teacher model, Tc represents the temperature parameter (generally 10), and the student model is easier to learn the fine-grained distribution of the teacher feature; CC s represents the target size (for example: lung nodule diameter, etc.) predicted by the student model; CC t represents the target size predicted by the teacher model.

[0219] Through the "double teacher-single student" distillation architecture, the model is lightened, the parameter quantity is reduced by about 75% (the parameter quantity is reduced from 28M to about 7M), the inference speed is increased by more than 3.5 times (the 3D sequence processing time is reduced from 12s to 3.4s in the GPU environment), and the inference precision loss is not more than 1%.

Claims

1. A method for automatic detection and diagnosis of lung lesions based on CT images, characterized in that: Comprise: Step S1, multi-modal feature fusion and segmentation are performed on the input CT image lung region; Step S2, lung nodule detection is realized based on dynamic adaptive dilated convolution, specifically: The dynamic adaptive dilated convolution module adjusts the dilated rate in real time according to the local feature complexity, comprising: acquiring the segmented lung area feature map in step S1 The local entropy and the gradient norm of each voxel ( x, y, z ): ； where: K denotes the number of gray levels; p i denotes the probability of a gray level i within a 3x3x3 neighborhood; Dynamic dilation rate is obtained by local entropy and gradient norm r(x, y, z) Avoid over-dilation leading to feature breakage: ； In the formula, , respectively represent the entropy weight coefficient and the gradient weight coefficient; r 0 represents the basic expansion rate; clip represents a truncation function; The detection network module is based on RetinaNet Improvements are made to integrate lightweight, adaptive feature extraction and complex sample optimization, including 3D MobileNetV3 as the backbone network for feature extraction; a feature pyramid network and a dynamic adaptive dilated convolution module are used as the neck structure of the network for multi-scale fusion and adaptive complex morphology capture; a double-branch head structure of RetinaNet is used as the detection head of the network, including a classification head and a regression head, wherein the classification head uses an improved FocalLoss to predict the probability of each anchor box outputting a lung nodule or background: ； wherein: p t denotes the predicted probability of the model that the sample belongs to lung nodules; denotes the class weight; denotes the focus parameter; Step S3, the 3D contour of the lung nodule is realized by combining multi-modal features and an improved Snakes model, thereby quantifying the size of the lung nodule; specifically: Improving the energy function of the Snakes model includes a multi-modal external energy, i.e. including an internal energy E int , an external energy E ext and a constraint energy E con : ； ； In the formula: a(s) , b(s) These represent the smoothness coefficients, used to control the curvature smoothness and stretch smoothness of the contour, respectively. U Represents the contour vertex, that is, a point on the 3D active contour, used to describe the geometric position of the active contour; k 1. k 2 represents the corresponding weight coefficients; This represents the average density of the nodules; C z Indicates the centroid of the candidate nodule; R yue Indicates the constraint radius to prevent the contour from overflowing the candidate nodule region; A continuous 3D contour is obtained after convergence of the improved Snake model, and is converted into a voxel representation of the 3D image, and the number of voxels inside the Snake contour is obtained by traversing the voxels of the 3D image N ts ; and the volume of the nodule is calculated by the number of voxels inside the Snake contour V F and the equivalent diameter d F : ； wherein: d x , d y , d z denotes the voxel-wise rate of CT; Step S4, a space-time dynamic tracking module is used to realize dynamic tracking of the lung nodule across time frames; specifically: First, a feature vector of each nodule is constructed F n : ； wherein, f DADC denote the output features of the dynamic adaptive dilated convolution module; cx 、 cy 、 cz denote the centroid coordinates; Then, Kalman filtering is used for state prediction and observation update: State prediction: ； where: represents the prior state estimate of the current t frame; A represents the state transition matrix; C t-1 represents the posterior state estimate of the t -1 frame; B represents the control input matrix; represents the control input vector; represents the prior estimate covariance matrix of the current t frame; P t-1 represents the posterior estimate covariance matrix of the t -1 frame; Q represents the process noise covariance matrix: ； In the formulae: I denotes the identity matrix; Observation update: ； wherein: K t denotes the Kalman gain; O denotes the observation matrix; Z r observation noise covariance matrix: ； Then, the features and position similarity are fused to perform feature matching of the nodule across time frames: ； In the formula: S(n t-1,i ,n t,j ) Indicates the ( t -1) frame i The nodule and the first t The first frame j The fusion similarity of individual nodules; Indicates the ( t -1) frame i Nodule characteristics With the t The first frame j Nodule characteristics Cosine similarity; Represents the t-th frame i Predicted location of each nodule; C t,j Indicates the first t Frame number j The actual detection location of each nodule; Indicates positional weight; Through feature matching, if the nodule is matched, the change trend of the nodule is output; if the nodule is not matched, a new nodule is marked and the previous frame nodule is marked as disappeared; Step S5, for the specific differences between benign and malignant nodules, the shape, density, and dynamic features of the lung nodule are extracted through the morphological feature module, the density feature module, and the dynamic feature module, respectively, to provide input for diagnosis decision; Step S6, a clinical diagnosis model is used to output the benign / malignant diagnosis result.

2. The method of claim 1, wherein the method further comprises: The multi-modal feature fusion is for lung window and mediastinum window features, that is, lung window feature maps and mediastinum window feature maps By means of a cross-window attention mechanism, feature weights of different window widths are dynamically allocated: ； In the formula: w lu , w me respectively represent the feature weights of the lung window and the mediastinum window; Conv represents 1x1x1 convolution dimension reduction; represents a Sigmoid function; The lung window feature map and the mediastinal window feature map are fused to obtain a fused feature map: 。 3.The CT image-based lung lesion automatic detection and diagnosis method according to claim 1 or 2, characterized in that: The encoder of the segmentation network structure in step S1 uses MobileNetV3 as the feature extraction backbone, and its internal structure includes an initial feature layer, a middle bottleneck module group, and a feature compression layer. The initial feature layer performs preliminary feature extraction on the input lung image through Conv3x3 convolution. The middle bottleneck module group includes multiple 3x3 and 5x5 separated convolution blocks, which are matched with batch normalization and H-Switch activation function. The feature compression layer uses Conv1x1 convolution to replace the fully connected layer and combines with the pooling operation to compress the extracted features into a form suitable for the decoder. The decoder of the segmentation network structure adopts the architecture of "downsampling-upsampling-jump connection" of U-Net, and its core is cross-level feature fusion. The encoder generates multi-scale feature maps through "downsampling and bneck module", and the decoder restores the feature map size through upsampling and fuses the "encoder corresponding level feature map" and "decoder upsampled feature map" through jump connection; The loss function of the segmentation network structure uses a hybrid loss of weighted Dice and cross-entropy: ； ； wherein: w CE , w Dice denote the weight coefficients of cross-entropy and Dice , respectively; p denotes the predicted output of the model, y denotes the true label.

4. The method of claim 3, wherein the method further comprises: The morphological feature module includes edge irregularity calculation and pleural traction sign detection: Edge irregularity calculation: Nodule contours obtained by improved snake model U ={ u 1, u 2,…, u N}, u i representing the contour vertex coordinates, obtaining contour fluctuations at multiple scales: ； wherein: L i+m denotes a vertex u i a line connecting the vertices u i+m dist denotes a vertex u i a distance to the line; denotes a measure m of the average distance;​ Pleural traction detection: shortest distance of nodule contour to pleura D ple obtaining the closest point of the node u min obtaining the traction direction and calculating the directional gradient consistency of the traction direction; The density feature module includes multi-modal density entropy and calcification specificity detection: Multimodality density entropy fusion lung window F lu with mediastinal window F me obtaining joint entropy of the nodule region: ； wherein: p(i, j) represents the joint probability of the lung window gray scale within the nodule region i and the mediastinum window gray scale j ; If MDE>1.2, it indicates that the density is uneven; Calcification specificity detection: a lightweight decoder is used to realize calcification judgment combined with calcification morphological features; The dynamic feature module includes core dynamic indicator acquisition and dynamic risk index fusion: Core dynamic indicators include the volume change rate, density change rate and morphology change rate at the previous frame t 0 and the current t 1 two time points ； wherein: Δt represents the interval time of two frames of detected CT images, V t0 , V t1 respectively represent the volume of the previous frame, the current nodule; respectively represent the average density of the previous frame, the current nodule; MI t0 , MI t1 respectively represent the contour fluctuation of the previous frame, the current nodule; Dynamic risk index fusion: ； In the formula: respectively represent the weight coefficients of the volume change, the density change, and the morphology change.

5. The method of claim 4, wherein the method further comprises: The step S6 specifically comprises: Firstly, the feature vector of single nodule is constructed by combining the features of shape, density and dynamics in step S5: ； wherein: TS represents pleural traction, CS represents calcification specificity; After that, the trained binary classification model is used for prediction, to F JJ-k the input, the benign and malignant label as the output, the prediction probability is generated, and the AUC value of the corresponding feature is calculated based on the prediction probability and the true label, that is AUC k ; Then, based on the clinical benchmark weights w cilm-k with feature discriminativeness AUC k , obtain attention weights for each feature W k : ； The fusion feature value is obtained: ； The clinical diagnosis model is embedded into the detection network module in step S2, the feature extraction backbone reuses the nodule feature branch of the detection network module, and meanwhile, the clinical diagnosis model adds two output heads, including a classification head and a feature labeling head, on the basis of the detection network module. The classification head outputs the probability of "nodule being malignant" through a Sigmoid activation function P mal ; the feature labeling head outputs the contribution degrees of five key features GX k ; meanwhile, the clinical diagnosis model adopts a multi-task loss function Loss dia : ； In the formula: w 1、 w 2 respectively represent corresponding weight coefficients; CE represents a cross-entropy loss, y mal indicates a benign or malignant label; indicates a feature contribution degree label; MSE represents a mean square error loss; If P mal <0.1 indicates very low risk; if 0.1 P mal <0.3 indicates low risk; if 0.3 P mal <0.6 indicates moderate risk; if 0.6 P mal <0.8 indicates high risk; if P mal ≥0.8 indicates very high risk.

Images