CN115713661B - Scoliosis Lenke parting system - Google Patents
Scoliosis Lenke parting system Download PDFInfo
- Publication number
- CN115713661B CN115713661B CN202211507294.8A CN202211507294A CN115713661B CN 115713661 B CN115713661 B CN 115713661B CN 202211507294 A CN202211507294 A CN 202211507294A CN 115713661 B CN115713661 B CN 115713661B
- Authority
- CN
- China
- Prior art keywords
- block
- spine
- point
- image
- representing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a scoliosis Lenke typing system which comprises a Lenke typing database unit, a full-spine X-ray image input unit, a full-spine X-ray image segmentation unit, a full-spine X-ray image matching unit and a scoliosis Lenke typing unit. According to the invention, the APSegmenter segmentation method is optimized through the self-adaptive post-processing module, so that the accuracy of the image extraction of the spine block is effectively improved, the main key nodes are matched by adopting a similarity matching algorithm of key points, the calculated amount is effectively reduced, the matching efficiency and accuracy are improved, and the problem of inaccurate matching of the spine block due to the problem of placement angle is solved through the shape representation of the symmetrical distance of the spine.
Description
Technical Field
The invention relates to the field of spinal diagnosis, in particular to a scoliosis Lenke typing system.
Background
With the development of the scientific and technical level and the popularization of electronic products, related health problems are endless. Myopia, obesity and scoliosis have become three major health problems for people today.
Scoliosis refers to abnormal curvature of the spine, and is the most common spine deformity disease. The cervical spondylosis is usually manifested as scoliosis, unequal heights of shoulders on two sides, backward protruding of the single-sided scapula during bending, and abnormal postures such as bending neck.
The impact of scoliosis on the human body is enormous, and in most patients, their physical health is affected by scoliosis, not only is the torso and chest contours deformed, but also the spinal cord and spinal nerves of severe persons are damaged, and respiratory and cardiac functions are affected. Particularly, in the teenager period, the patient is in the growth period of the spine, the scoliosis has serious influence on the patient, and if the patient is not treated effectively, the scoliosis condition can be worsened, and the scoliosis angle is increased. In daily life, a patient with serious scoliosis cannot walk, go upstairs and run, sometimes accompanied by heart pressure, and normal life is seriously affected.
The scoliosis parting is based on the morphological characteristics of the spine of the patient, and provides important value for determining the morphological characteristics of the scoliosis deformity of the patient. The scoliosis Lenke typing divides scoliosis into 6 types of scoliosis, is a common and well-known typing system for the prior spinal surgery, and is one of the standard typing methods of international idiopathic scoliosis. In scoliosis diagnosis, how to realize correct Lenke typing of scoliosis has important significance for correction, rehabilitation, operation strategies and the like of scoliosis, and is also a difficult problem in the current clinical diagnosis. Traditional Lenke typing is mainly carried out by traditional methods such as Cobb angle measurement, and the traditional measurement method is that doctors manually measure the Cobb angle on an X-ray film by using a pencil and a protractor for calculation. However, such a method has certain drawbacks: (1) The slower manual measurement speed is susceptible to personal subjective factors; (2) The measurement relies on the knowledge and experience of the doctor, and the diagnosis is slow and inaccurate. In practice, the scoliosis is mainly based on the shape characteristics, main bending positions and the like of the spine and the vertebral block, and aiming at the requirement of scoliosis parting, an automatic Lenke parting system can be realized by combining methods of deep learning, shape characteristic characterization and the like through a computer vision technology, and assisted scoliosis auxiliary diagnosis is realized.
Noun interpretation:
transformer encoder: each encoder consists of n encoders, wherein each encoder is provided with two sub-layers, one sub-layer is connected with a multi-head self-attention normalization layer and the other sub-layer is connected with a feed-forward layer normalization level residual unit.
MSA block: multi-head self-attention block.
Point-wise MLP block: the sensor modules are layered point by point.
Point-wise linear layer: the linear layer is pointed.
Disclosure of Invention
In order to solve the problems, the invention provides a scoliosis Lenke typing system.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
a scoliosis Lenke typing system comprises a Lenke typing database unit, a full spine X-ray image input unit, a full spine X-ray image segmentation unit, a full spine X-ray image matching unit and a scoliosis Lenke typing unit;
the Lenke typing database unit comprises 6 types of standard pictures and 1 type of normal spine pictures without lateral curvature, wherein the number of each type of standard pictures is m1, m2, m3, m4, m5, m6 and m7;
the full spine X-ray image input unit is used for inputting a full spine X-ray image to be detected;
the full spine X-ray image segmentation unit is used for segmenting a full spine X-ray image to be detected to obtain a spine block image in the full spine X-ray image to be detected;
the full spine X-ray image matching unit is used for comparing the full spine X-ray image to be detected with standard pictures in the Lenke typing database unit to obtain K standard pictures which are the closest to the image to be detected;
and the scoliosis Lenke typing unit judges that the K images have the most types of Lenke typing categories, namely the types of the Lenke typing of the full-spine X-ray images to be detected belong to.
Further improvement, the whole spine X-ray image segmentation unit segments the whole spine X-ray image by an APSegmenter method to obtain a spine block image:
1.1 Decomposing the input full spine X-ray image X into a sequence of blocks
Each image block is then stretched into a one-dimensional vector, and finally projected to a block embedding by linear transformation to produce
Representing a projection operation; then, by embedding a position of a position which can be learned, it is encoded with pos= [ pos ] 1 ，...，pos N ]∈R N×D To obtain an input sequence z of final results 0 ＝x 0 +pos；
Wherein x is N Represents the Nth block, R D A D-dimensional space representing a real number,
representing N×P 2 X C dimensional space, N represents the number of blocks, P x P represents the size of a block, C represents the number of channels, pos represents the sequence of position embedding points, pos N Represents the N-th position embedded point, R N×D Representing an N x D dimensional space; z is Z 0 Position marker, x, representing a sequence of blocks 0 Representing a block embedding sequence;
1.2 A transducer encoder consisting of L layers is provided, each layer comprising a multi-head self-attention block and a point-by-point multi-layer perceptron module, the sequence Z 0 Generating context encoding z L ∈R N×D ：
a i-1 ＝MSA(LN(z i-1 ))+z i-1 ， [1]
z i ＝MLP(LN(a i-1 ))+a i-1 ， [2]
Where i e {1,., L }, and by self-attention operation, to more effectively utilize Z L A sequence;
a i-1 representing intermediate results through MSA, MSA () representing processed through a multi-headed self-care block, LN () representing processed through LayerNorm, i.e., normalized for all features of each sample, z i-1 Representing the i-1 st context code, the MLP () representation being a point-by-point multi-layer perceptron module representation;
1.3 Decoding the image block code sequence z L ∈R N×D Decoding into split map s E R H×W×K S is segmentation mapping, H and W are image block height and width, and K is category number; decoder learning from encoderA class score mapped to the image block level; the class scores at the image block level are upsampled to pixel level scores by bilinear interpolation;
1.4 Z of input L ∈R N×D First through a point-to-line layer transformation to z lin ∈R N×K Then from z lin ∈R N×K Remodelling to s lin ∈R H/P×W/P×K ；s lin ∈R H/P×W/P×K Then performing bilinear upsampling to the original image size; z lin Represents the logarithm of block-level class, K represents the number of classes, s lin Representing a 2D feature map;
1.5 Introduction of a learnable class embedding, cls= [ cls ] 1 ，...，cls K ]∈R K×D Random initialization is allocated to a single semantic class, and the same image block codes z L ∈R N×D Processing together; cls represents class embedding, cls K Representing a kth class embedding, K representing a class number;
next, a transform encoder consisting of M layers is defined, and z is embedded by computing the normalized image block output by the decoder mask ∈R N×D And class embedding c E R K×D Generating K mask maps, the set of class masks is calculated as follows: masks= (z) mask ，c)＝z mask c T Wherein masks= (z) mask C) represents a set of block sequences; further, each mask block sequence is remodelled into a two-dimensional mask, denoted S mask ∈R H/P×W/P×K Obtaining class classification at pixel level by upsampling layer and then applying Argmax function to form final pixel segmentation map I 1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein z is mask Representing image block embedding, c representing class embedding, K representing class number, T representing transpose operation, masks representing class mask set.
Further improvement, further comprising the steps of:
2.1 Dividing the full spine X-ray image to obtain a final pixel division diagram I 1 Using the formula [3 ]]Firstly, performing open operation on the background, wherein the size of a convolution kernel B is 4 multiplied by 30, then performing open operation on the vertebral block region, the size of the convolution kernel B is 10 multiplied by 40, and recording the processed imageIs I 2 ；
2.2 Using equation [4 ]]Remove image I 2 Calculating the length and width of each divided vertebral block in the area with the middle area smaller than 500, wherein the length of the adhesion area is smaller than the width, and the length of the normal vertebral block is larger than the width, and respectively storing the normal vertebral block and the abnormal vertebral block into an image I 3 And I 4 ；
Wherein width represents the width high of each cone block in the segmentation result, and height area of each cone block in the segmentation result represents the ridge block area; connour represents the segmented cone region;
2.3 Acquiring image I) 4 The boundary information of each region in the map is traversed to the left boundary, the step length is 3, and the formula [5 ] is utilized]Solving for left_point i-1 、Left_point i 、Left_point i+1 The included angle theta between the three points is considered as the point left_point when the cos theta is more than or equal to 0 i For adhering the Left boundary point of the vertebral block segmentation, traversing the right boundary, and searching for the left_point i-1 、Left_point i+1 Right boundary point right_point at the same horizontal plane i-1 、Right_point i+1 Closing the left_point i-1 、Left_point i+1 、Right_point i-1 、Right_point i+1 For the boundary of the vertebral block, the saved image is I 5 ；
Left_point i-1 、Left_point i 、Left_point i+1 Respectively representing the left lower vertex of the cone block with the bit sequence before the current cone block, the left lower vertex of the current cone block and the left lower vertex of the cone block with the bit sequence behind the current cone block; right_Point i-1 、Right_point i+1 Respectively representing the right lower vertex of the cone block with the bit sequence before the current cone block and the right lower vertex of the cone block with the bit sequence after the current cone block;
2.4 Merging I) 3 And I 5 Obtaining a vertebral block image I in the full-spine X-ray image to be detected after final self-adaptive optimization processing 6 。
Further improvement, the full-spine X-ray image matching unit determines the category to which the Lenke classification of the spine block image in the full-spine X-ray image to be detected belongs by the following method:
3.1 A spine image after the segmentation in the step one is I 6 ，s＝{p 0 ，p 1 ，...，p m [ means I ] 6 The end point sequence of the left outer contour of the middle vertebral block is that of the vertebral block image I 6 The method comprises the steps of including m/2 vertebra blocks, wherein m endpoints are included in the endpoints of the left outer contour of the vertebra, setting a shape coding template, namely, equally dividing a circular area into n areas along the circumferential direction, coding according to the sequence of 1,2 and 3.n from the top area along the clockwise direction, sequentially connecting the m endpoints from top to bottom to form m-1 line segments, arranging the circle center of the circular area at the endpoint of the head end of each line segment, coding the shape of the corresponding line segment by the number of the area where the corresponding line segment passes, and coding all the line segments by using 1 to n to obtain a group of coding sequences CS= { x 1 ，x 2 ，...，x m-1 X, where x i Is the shape coding of the line segment between the i-1 th endpoint and the i-th endpoint in one vertebra image, and finally, the coding sequence C of two vertebra block images 1 ＝{x 1 ，x 2 ，...，x m-1 Sum C 2 ＝{y 1 ，y 2 ，...，y m-1 Dissimilarity DisSim of } 1 (C 1 ，C 2 ) The definition is as follows:
where n represents the number of cyclic partition templates and m represents the number of spinal blocks; p (P) m The mth endpoint, y, representing the right outer contour of the spinal block i Is the shape code between the i-1 th endpoint and the i-th endpoint in another spine image;
3.2 Let s= { p 1 ，p 2 ，...，p m Sum S '= { p' 1 ，p′ 2 ，...，p′ m Respectively, is the outer contour endpoint set, C, of the two segmented vertebral block images X and Y 1 ＝{x 1 ，x 2 ，...，x m-1 Sum C 2 ＝{y 1 ，y 2 ，...，y m-1 The shape coding sequences corresponding to X and Y are PL respectively 1 ＝{x 1 ，x 2 ，...，x m-1 The coordinate of the left lower corner point of the first block to the m-1 th block of the X cone block of the vertebra is set as PB= { y 2 ，y 3 ，...，y m Coordinates of upper left corner points of the 2 nd to the m th cone blocks of the vertebral cone blocks are provided with PK 1 ＝{x 1 ，x 2 ，...，x m-1 Is PL } 1 With PB 1 The slope formed by the corresponding points,
set PL 2 ＝{x 1 ，x 2 ，...，x m-1 The sign of the Y-shaped cone block is the coordinates of the left lower corner point of the first to m-1 th cone blocks of the vertebra, PB 2 ＝{y 2 ，y 3 ，...，y m Setting PK for coordinates of upper left corner points of the 2 nd to the m th cone blocks of the Y-cone blocks of the vertebra 2 ＝{x 1 ，x 2 ，...，x m-1 Is PL } 2 With PB 2 The slope formed by the corresponding points,
PL i the calculation method of (1) is that
x i [1]The ordinate information of the left lower corner point of the cone block and x are stored i [0]Stored is the abscissa information of the left lower corner point of the cone block
y i+1 [1]The ordinate information of the right lower corner point of the cone block, y is stored i+1 [0]Storing abscissa information of right lower corner point of cone block
PL n-i Wherein n takes the values of X and Y and corresponds to the finger cone block X and the cone block Y
For PL n-i If positive and negative changes of the slope occur, namely the maximum change slope occurs, more than or equal to 1 local maximum change slope exists in the spine; for s and s', if the boundary line segment (p i-1 ，p i ) If there is a slope of local maximum change, then define p i (i > 0) as key point, (p) i-1 ，p i ) Given the number of keypoints of any segmented spine image, the keypoints in s and s' are denoted as { kp }, respectively j J=1,.. j * J=1,..k; next, each key point p is calculated using equation (5) i Matching weight w of (i > 0) i ：
Wherein the operation represents a separation distance between two points on the outer contour of the spinal block;
finally, for two sets of code sequences C 1 ＝{x 1 ，x 2 ，...，x m Sum C 2 ＝{y 1 ，y 2 ，...，y m The degree of dissimilarity between } is further optimized as:
3.3 Given the left outer contour endpoint set s= { p for the segmented spinal block X 0 ，p 1 ，...，p m Respectively across each end point p i Make horizontal line l i In l i Calculating the sum p in the clockwise direction as a reference i Angle theta of corresponding upper or lower boundary i Then define equation [8 ]]To calculate the horizontal inclination angle alpha of the spine i In practice, equation [8]Implying the horizontal angle inclination direction of the vertebral block if theta i Less than or equal to 90 degrees, meaning alpha i Taking l i The angle at which the clockwise direction intersects the upper and lower boundaries, i.e. alpha i Is a positive number; if theta is i > 90 DEG, meaningTaste alpha i Taking l i The angle at which the inverted needle direction intersects the upper and lower boundaries, i.e. alpha i Is a negative number;
let X and Y be the two groups of spines to be matched after segmentation, S= { p 0 ,p 1 ,...,p m Sum S '= { p' 0 ,p′ 1 ,...,p′ m Respectively their outer contour endpoint sets, C 1 ＝{x 1 ,x 2 ,...,x m Sum C 2 ＝{y 1 ,y 2 ,...,y m The shape coding sequences corresponding to X and Y are respectively, and alpha= { alpha } is 0 ,α 1 ,...,α m Sum beta= { beta } 1 ,β 2 ,...,β m And the horizontal inclination angles of X and Y are respectively, the dissimilarity degree of X and Y is further optimized as:
further improvement, let X and Y be two vertebrae to be matched after segmentation, (1) calculate X and Y's outer package rectangle respectively first, and calculate the distance from all outline endpoints on the left side of the vertebra block to their outer package rectangle right boundary, record as respectively: h is a X ＝{h 0 ,h 1 ,...,h m And d y ＝{d 0 ,d 1 ,...,d m "wherein the mark with the greatest distance is h max And d max ；h X Representing the set of distances from all the outline points on the left side of the vertebral block on the vertebra X to the right boundary of the outsourcing rectangle, starting with the number 0, d y Representing the distance set from all the outline points on the left side of the vertebral block on the vertebra Y to the right boundary of the outsourcing rectangle, starting from 0 and h m Representing the distance set from all the outer contour points on the left side of the vertebral block with the sequence number m to the right boundary of the outer rectangle of the vertebral block with the sequence number m on the vertebra X, and dm represents the distance set from all the outer contour points on the left side of the vertebral block with the sequence number m to the right boundary of the outer rectangle of the vertebral block with the sequence number m on the vertebra Y;
(2) turning Y along a vertical line to obtain a mirror image of Y, marking as Y ', further calculating an outsourcing rectangle of Y ', and similarly calculating the distance from all outer contour endpoints on the left side of the vertebra block in Y ' to the right boundary of the outsourcing rectangle, wherein the distance is marked as: d, d y ′＝{d 0 ′,d 1 ′,...,d m ′}；
(3) Calculating the optimal symmetrical matching distance of X and Y as dis (X, Y) according to a formula [11 ];
the general formula [10] and the formula [11] are defined as the shape difference degree between X and Y:
wherein epsilon, lambda, delta are weight factors; the smaller the Dissimilar (X, Y) is, the closer the two spine block images are, the scoliosis Lenke typing unit obtains K standard pictures closest to the full spine X-ray image to be detected in the Lenke typing database unit according to a formula [12], and the K images have the most types of Lenke typing categories, so that the Lenke typing of the spine block images in the full spine X-ray image to be detected is judged to belong to the categories.
The invention has the following beneficial effects:
1. the segmentation method of the APsegment (APsegment is the name compiled by us) is optimized through the self-adaptive post-processing module, and the original method is called segment), so that the accuracy and the completeness of segmentation of the vertebral blocks in the X-ray image of the vertebral column are effectively improved.
2. And a similarity matching algorithm of key points is adopted to carry out weighted matching on main key nodes, so that the matching accuracy is improved, the maximum bending part is emphasized, and the Lenke typing standard is more met.
3. Through the description of the horizontal inclination angle of the vertebral block, the boundary trend distribution depiction of the vertebral block is thinned, and the accuracy of discriminating the scoliosis Lenke typing is improved.
4. The shape representation of the symmetrical distance of the spine solves the problem of inaccurate matching of the spine block caused by the problem of placement angle.
5. The invention can realize Lenke typing by only front X-ray chest vertebra images based on a deep learning and shape description method, greatly improves the convenience of diagnosis, and has the advantages of convenience, rapidness, accuracy and the like.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention. In the drawings:
FIG. 1 is a block diagram of post-processing of the results of a preliminary segmentation;
FIG. 2 is a diagram of a shape coding template and shape coding example;
FIG. 3 is an exemplary diagram of key points and key line segments;
FIG. 4 is a graph illustrating an example of the calculation of the horizontal inclination angle of a spinal block;
FIG. 5 is an exemplary view of a spinal vertical distance;
FIG. 6 is a graph comparing accuracy of the system of the present invention with that of manual diagnosis;
FIG. 7 is a graph comparing the efficiency of the system of the present invention with that of manual diagnosis.
Detailed Description
Embodiments of the invention are described in detail below with reference to the attached drawings, but the invention can be implemented in a number of different ways, which are defined and covered by the claims.
A scoliosis Lenke typing system and method comprises the following steps:
step one: the spine block in the spine X-ray image is segmented by an APsegment method, and mainly comprises two technical points:
1. the segment method is used for carrying out preliminary segmentation on the vertebral blocks, and the specific flow is as follows:
(1) First, an input image x is decomposed into
Each patch is then stretched into a 1D vector εR D Finally, through linear transformation to ebedding E R D It is encoded by position embedding of a leachable position pos= [ pos ] 1 ，...，pos N ]∈R N×D To obtain an input sequence z of final results 0 ＝x 0 +pos。
(2) Transformer encoder consisting of L layers, each layer comprising an MSA block and a point-wise MLP block, sequence Z 0 Generating contextualization Z L Sequence z L ∈R N×D Each block is preceded by an LN layer, and each block is followed by a residual connection for output and input.
a i-1 ＝MSA(LN(z i-1 ))+z i-1 ， [1]
z i ＝MLP(LN(a i-1 ))+a i-1 ， [2]
Where i e {1,., L }; by a self-attribute, Z is utilized more effectively L Sequence.
(3) Patch coding sequence z L ∈R N×D Decoded into a split map s e R H×W×K . The decoder learns to map the patch-level code from the encoder to the patch-level class score. The class score of the patch-level is upsampled to the pixel level score by bilinear interpolation.
(4) Input z L ∈R N×D First transformed to z through point-wise linear layer lin ∈R N×K ，z lin ∈R N ×K Remodelling to s lin ∈R H/P×W/P×K 。s lin ∈R H/P×W/P×K Then the original image size is obtained through bilinear upsampling to obtain the final segmented image s epsilon R H×W×K 。
(5) Introducing a leachable class ebedding, class= [ class ] 1 ，...，cls K ]∈R K×D Random initialization is assigned to a single semantic class, as is patch encodings z L ∈R N×D Processing together.
(6) Mask transducer passing meterCalculation patch encodings z mask ∈R N×D And class embeddings c E R K×D Scalar product of (c) masks= (z' M ，c)＝z′ M c T K mask graphs are generated. And then through a Upsample and Argmax layer, the up-sampled image is applied to Argmax to obtain a single class of each pixel, the size of the original image is restored, and a final pixel segmentation map is output through a classification layer.
2. Method for optimizing APSegmenter segmentation by self-adaptive post-processing module
(1) Dividing the full spine X-ray image by using segment network model to obtain an image I 1 Using the formula [3 ]]The background is first subjected to an open operation, and the convolution kernel B has a size of 4×30. Then the vertebral area is subjected to open operation, the size of the convolution kernel B is 10 multiplied by 40, and the processed image is recorded as I 2 。
(2) Using equation [4 ]]Remove image I 2 Calculating the length and width of each divided vertebral block in the area with the middle area smaller than 500, wherein the length of the adhesion area is smaller than the width of the divided vertebral block, and the length of the normal vertebral block is larger than the width of the adhesion area, and respectively storing the normal vertebral block and the abnormal vertebral block into an image I 3 And I 4 。
(3) Acquiring image I 4 The boundary information of each region in the map is traversed to the left boundary, the step length is 3, and the formula [5 ] is utilized]Solving for left_point i-1 、Left_point i 、Left_point i+1 The included angle between the three points is regarded as the point left_point when cos theta is more than or equal to 0 i Left boundary point for the segment of the adhesion cone. Traversing the right boundary, looking for a vector associated with left_point i-1 、Left_point i+1 Right boundary point right_point at the same horizontal plane i-1 、Right_point i+1 . Closed left_point i-1 、Left_point i+1 、Right_point i-1 、Right_point i+1 For the boundary of the vertebral block, the saved image is I 5 。
(4) Merge I 3 And I 5 Obtaining a final adaptive optimization processed result I 6 。
Step two: the adaptive shape description operator is provided for carrying out shape description and matching on the segmented spine image, and the adaptive shape description operator mainly comprises the following technical points:
(1) Unified representation and matching of spine outline
In order to effectively describe different bending degrees of scoliosis and extract shape characteristics of the scoliosis for Lenke typing, a unified representation frame of the outer contour of the spine and a similarity matching method are designed. In particular, taking into account the nature of the imaging of the spine x-ray image, directional cyclic encoding is used to characterize the change in the image. Let the segmented spine image of step one be I, s= { p 0 ，p 1 ，...，p m The "end point sequence of the right outer contour of the vertebral block" in I "indicates that if t vertebral blocks are included in the vertebral image I, the end point of the left outer contour of the vertebral block should include 2t end points. A shape coding template as shown in fig. 2 is provided, the directions are subdivided into n directions, and coding is performed using 1 to n. Further, all endpoints of the left outer contour are traversed under self-orientation, and directions between adjacent endpoints are encoded by using a shape encoding template, so that a group of encoding sequences CS= { x is obtained 1 ，x 2 ，...，x m X, where x i Is the shape coding between endpoint i-1 and endpoint i. Finally, for two sets of code sequences C 1 ＝{x 1 ，x 2 ，...，x m Sum C 2 ＝{y 1 ，y 2 ，...，y m -its dissimilarity is defined as:
where n represents the number of cyclic partition templates and m represents the number of spinal blocks.
(2) Adaptive similarity match weight based on key segments adaptive weight similarity matching based on keypoints
Those portions of the scoliosis that are most curved play a critical role in Lenke typing. Therefore, we have devised a similarity matching algorithm based on key points. Specifically, assume that s= { p 0 ,p 1 ,...,p m Sum S '= { p' 0 ,p′ 1 ,...,p′ m Respectively, is the outer contour endpoint set, C, of two segmented vertebral blocks X and Y 1 ＝{x 1 ,x 2 ,...,x m Sum C 2 ＝{y 1 ,y 2 ,...,y m And the shape coding sequences corresponding to X and Y are respectively. First, the definition of the keypoints is given, i.e. for S and S', if the boundary line segment (p i-1 ,p i ) If there is a slope of local maximum change, then define p i (i > 0) as key point, (p) i-1 ,p i ) Is a critical boundary line segment. The segmented spinal block, as shown in fig. 3, is marked with critical border segments that are found to tend to represent the most strongly varying portions of the segment. Given the number of keypoints for any segmented spine image, the keypoints in S and S' can be represented as { kp }, respectively j J=1,.. j * J=1,..k. Next, it is proposed to calculate each point p using equation (5) i Matching weight w of (i > 0) i 。
Where the operation represents the separation distance between two points on the outer contour of the spinal block.
Finally, for two sets of code sequences C 1 ＝{x 1 ,x 2 ,...,x m Sum C 2 ＝{y 1 ,y 2 ,...,y m The degree of dissimilarity between } may be further optimized as:
it should be noted that the number k of key points is optimized in the experiment, and the similarity matches the weight a i And further performing normalization operation. In fact, by this weight assignment strategy, the shape matching process can be made more focused on where the spinal block is curved the most, thus facilitating Lenke typing.
(3) improving shape representation by inner boundary shape representation and matching optimization based on horizontal inclination angle of vertebral block
The two steps respectively represent the scoliosis shape from the two aspects of the overall profile and the local maximum bending condition of the spine, and in practice, for the scoliosis Lenke typing, the horizontal inclination of each spine block, namely the inclination angle of the upper boundary and the lower boundary, plays an important role in judging the scoliosis Lenke typing. Therefore, we further devised a shape representation and matching method based on the horizontal inclination angle of the spinal block, so as to optimize the feature extraction of scoliosis Lenke typing. Specifically, as shown in fig. 4, the left outer contour endpoint set s= { p for a given segmented spinal block X 0 ,p 1 ,...,p m Respectively across each end point p i Make horizontal line l i In l i Calculating the sum p in the clockwise direction as a reference i Angle theta of corresponding upper or lower boundary i We define equation (6) to calculate the horizontal inclination angle α of the vertebra i . In fact, equation (6) implies the horizontal angular tilt direction of the spinal block if θ i Less than or equal to 90 degrees, meaning alpha i Taking l i The angle at which the clockwise direction intersects the upper and lower boundaries, i.e. alpha i Is a positive number; if theta is i > 90 °, then means α i Taking l i The angle at which the inverted needle direction intersects the upper and lower boundaries, i.e. alpha i And is negative.
Assuming that X and Y are the two sets of spines to be matched after segmentation, s= { p 0 ,p 1 ,...,p m Sum S '= { p' 0 ,p′ 1 ,...,p′ m Respectively their outer contour endpoint sets, C 1 ＝{x 1 ,x 2 ,...,x m Sum C 2 ＝{y 1 ,y 2 ,...,y m The shape coding sequences corresponding to X and Y are respectively, and alpha= { alpha } is 0 ,α 1 ,...,α m Sum beta= { beta } 1 ,β 2 ,...,β m The degree of dissimilarity of X and Y may be further optimized as:
(4) Shape representation based on spinal symmetric distance
(1) In the spine shape representation and matching, there is also a case of symmetrical mirror image matching, for example, as shown in fig. 5, where the left and right are the same spine segments, respectively, and are inverted, and if the two are represented and matched using only the above formulas (5), (6), and (8), there will be a great difference between them, and they are in fact identical. Therefore, we further devised a method of symmetric distance characterization to optimize the representation and matching of the spine. Specifically, let X and Y be two vertebrations to be matched after segmentation, (1) respectively calculating the outer wrapping rectangles of X and Y, and calculating the distances from all outer contour endpoints on the left side of the vertebration block to the right boundary of the outer wrapping rectangles, and respectively marking as: h is a X ＝{h 0 ,h 1 ,...,h m And d y ＝{d 0 ,d 1 ,...,d m "wherein the mark with the greatest distance is h max And d max . (2) And turning Y along the vertical line to obtain a mirror image of Y, which is denoted as Y'. And further calculating an outsourcing rectangle of the Y ', and likewise calculating the distances from all outer contour endpoints on the left side of the vertebral block in the Y' to the right boundary of the outsourcing rectangle, wherein the distances are recorded as follows: d, d y ′＝{d 0 ′,d 1 ′,...,d m '}. (3) Calculating the optimal symmetrical matching distance of X and Y according to the formula (9)Dis (X, Y).
In fact, the formula ensures that X and Y are matched in an optimal symmetrical mode, so that the influence caused by mirror image overturning is eliminated, and meanwhile, the bending degree of each spinal block of the scoliosis is also described in terms of the distance from the symmetry axis, so that the correct Lenke typing is facilitated.
Finally, in combination with the several shape representation optimization steps described above, we present the final spine shape matching formula. Namely, for two vertebrations to be matched after division, the shape difference degree between the two vertebrations can be defined by integrating the formula (8) and the formula (9):
wherein epsilon, lambda and delta are weight factors, and strategies used for representing the matching of the shapes represent the importance degree of matching for the final shape.
Step three: the Lenke typing method based on the segment post-processing and self-adaptive shape description method is provided and mainly comprises the following technical points:
(1) Let the data set contain n scoliosis medical samples, respectively belong to 7 categories, the first six categories are the 6 kinds of scoliosis categories of Lenke typing, the 7 th category is the normal spine image sample without scoliosis, and the number of each category is m1, m2, m3, m4, m5, m6, m7.
(2) Let X be the spine image to be classified and let y be i I=1, 2,..n, for any sample in the database. First, the method of the step one is used for X and y i Performing quick segmentation of the spine, and extracting x and y by using the method of the second step (namely formula (10)) i According to the similarity, finding out K images most similar to x in the database, wherein the K images have the Lenke classification type (including normal type) with the most types, namely, the Lenke classification of x is determinedThe model belongs to the category. Particularly, if the K images have the most normal categories, the full spine X-ray image to be detected is judged to be the normal categories, namely no Lenke scoliosis occurs.
As shown in fig. 6 and 7, the diagnostic accuracy of the present invention can reach 98.5%, and the detection efficiency is about 7.2 times of the manual detection efficiency.
The above is only a preferred embodiment of the present invention, and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (2)
1. The scoliosis Lenke typing system is characterized by comprising a Lenke typing database unit, a full-spine X-ray image input unit, a full-spine X-ray image segmentation unit, a full-spine X-ray image matching unit and a scoliosis Lenke typing unit;
the Lenke typing database unit comprises 6 types of standard pictures and 1 type of normal spine pictures without lateral curvature, wherein the number of each type of standard pictures is m1, m2, m3, m4, m5, m6 and m7;
the full spine X-ray image input unit is used for inputting a full spine X-ray image to be detected;
the full spine X-ray image segmentation unit is used for segmenting a full spine X-ray image to be detected to obtain a spine block image in the full spine X-ray image to be detected;
the full spine X-ray image matching unit is used for comparing the full spine X-ray image to be detected with the standard images in the Lenke typing database unit to obtain K' standard images which are the closest to the image to be detected;
the scoliosis Lenke typing unit judges that the K' images have the most types of Lenke typing categories, namely the types of the Lenke typing of the full-spine X-ray images to be detected;
the full spine X-ray image segmentation unit segments the full spine X-ray image by an APSegmenter method to obtain a spine block image:
1.1 Decomposing the input full spine X-ray image X into a sequence of blocks
Each image block is then stretched into a one-dimensional vector, which is finally projected by linear transformation into a block embedding, yielding +.>
Representing a projection operation; then, by embedding a position of a position which can be learned, it is encoded with pos= [ pos ] 1 ,...,pos N ]∈R N×D To obtain an input sequence z of final results 0 ＝x 0 +pos；
Wherein x is N Represents the Nth block, R D A D-dimensional space representing a real number,
representing N×P 2 X C dimensional space, N represents the number of blocks, P x P represents the size of a block, C represents the number of channels, pos represents the sequence of position embedding points, pos N Represents the N-th position embedded point, R N×D Representing an N x D dimensional space; z is Z 0 Position marker, x, representing a sequence of blocks 0 Representing a block embedding sequence;
1.2 A transducer encoder consisting of L layers is provided, each layer comprising a multi-head self-attention block and a point-by-point multi-layer perceptron module, the sequence Z 0 Generating context encoding z L ∈R N×D ：
a i-1 ＝MSA(LN(z i-1 ))+z i-1 , [1]
z i ＝MLP(LN(a i-1 ))+a i-1 , [2]
Where i e {1,., L }, and by self-attention operation, to more effectively utilize Z L A sequence;
a i-1 representing intermediate results through MSA, MSA () representing processed through a multi-headed self-care block, LN () representing processed through LayerNorm, i.e., normalized for all features of each sample, z i-1 Representing the i-1 st context code, the MLP () representation being a point-by-point multi-layer perceptron module representation;
1.3 Decoding the image block code sequence z L ∈R N×D Decoding into split map s E R H×W×K S is segmentation mapping, H and W are image block height and width, and K is category number; the decoder learns class scores that map image block level encodings from the encoder to image block levels; the class scores at the image block level are upsampled to pixel level scores by bilinear interpolation;
1.4 Z of input L ∈R N×D First through a point-to-line layer transformation to z lin ∈R N×K Then from z lin ∈R N×K Remodelling to s lin ∈R H/P×W/P×K ；s lin ∈R H/P×W/P×K Then performing bilinear upsampling to the original image size; z lin Represents the logarithm of block-level class, K represents the number of classes, s lin Representing a 2D feature map;
1.5 Introduction of a learnable class embedding, cls= [ cls ] 1 ,...,cls K ]∈R K×D Random initialization is allocated to a single semantic class, and the same image block codes z L ∈R N×D Processing together; cls represents class embedding, cls K Representing a kth class embedding;
next, a transform encoder consisting of M layers is defined, and z is embedded by computing the normalized image block output by the decoder mask ∈R N×D Class-sum embedded cls e R K×D Generating K mask maps, the set of class masks is calculated as follows: masks= (z) mask ,cls)＝z mask cls T Wherein masks= (z) mask Cls) represents a set of block sequences; further, each mask block sequence is remodelled into a two-dimensional mask, denoted S mask ∈R H/P×W/P×K Obtaining class classification at pixel level by upsampling layer and then applying Argmax function to form final pixel segmentation map I 1 The method comprises the steps of carrying out a first treatment on the surface of the Wherein z is mask Representing image block embedding, T representing transpose operation, masks representing class mask sets;
2.1 Dividing the full spine X-ray image to obtain a final pixel division diagram I 1 Using the formula [3 ]]Firstly, performing open operation on the background, wherein the size of a convolution kernel B is 4 multiplied by 30, then performing open operation on the vertebral block region, the size of the convolution kernel B is 10 multiplied by 40, and marking the processed image as I 2 ；
2.2 Using equation [4 ]]Remove image I 2 Calculating the length and width of each divided vertebral block in the area with the middle area smaller than 500, wherein the length of the adhesion area is smaller than the width, and the length of the normal vertebral block is larger than the width, and respectively storing the normal vertebral block and the abnormal vertebral block into an image I 3 And I 4 ；
Wherein width represents the width high of each cone block in the segmentation result, and height area of each cone block in the segmentation result represents the ridge block area; connour represents the segmented cone region;
2.3 Acquiring image I) 4 The boundary information of each region in the map is traversed to the left boundary, the step length is 3, and the formula [5 ] is utilized]Solving for left_point i-1 、Left_point i 、Left_point i+1 The included angle theta between the three points is considered as the point left_point when the cos theta is more than or equal to 0 i For adhering the Left boundary point of the vertebral block segmentation, traversing the right boundary, and searching for the left_point i-1 、Left_point i+1 Right boundary point right_point at the same horizontal plane i-1 、Right_point i+1 Closing the left_point i-1 、Left_point i+1 、Right_point i-1 、Right_point i+1 For the boundary of the vertebral block, the saved image is I 5 ；
Left_point i-1 、Left_point i 、Left_point i+1 Respectively representing the left lower vertex of the cone block with the bit sequence before the current cone block, the left lower vertex of the current cone block and the left lower vertex of the cone block with the bit sequence behind the current cone block; right_Point i-1 、Right_point i+1 Respectively representing the right lower vertex of the cone block with the bit sequence before the current cone block and the right lower vertex of the cone block with the bit sequence after the current cone block;
2.4 Merging I) 3 And I 5 Obtaining a vertebral block image I in the full-spine X-ray image to be detected after final self-adaptive optimization processing 6 The method comprises the steps of carrying out a first treatment on the surface of the The full spine X-ray image matching unit determines the category to which the Lenke classification of the spine block image in the full spine X-ray image to be detected belongs by the following method:
3.1 A spine image after the segmentation in the step one is I 6 ，S＝{p 0 ,p 1 ,...,p m [ means I ] 6 The end point sequence of the left outer contour of the middle vertebral block is that of the vertebral block image I 6 The method comprises the steps of including m/2 vertebra blocks, wherein m endpoints are included in the endpoints of the left outer contour of the vertebra, setting a shape coding template, namely, equally dividing a circular area into n areas along the circumferential direction, coding according to the sequence of 1,2 and 3.n from the top area along the clockwise direction, sequentially connecting the m endpoints from top to bottom to form m-1 line segments, arranging the circle center of the circular area at the endpoint of the head end of each line segment, coding the shape of the corresponding line segment by the number of the area where the corresponding line segment passes, and coding all the line segments by using 1 to n to obtain a group of coding sequences CS= { x 1 ，x 2 ，...，x m-1 X, where x i Is the shape coding of the line segment between the i-1 th endpoint and the i-th endpoint in one vertebra image, and finally, the coding sequence C of two vertebra block images 1 ＝{x 1 ，x 2 ，...，x m-1 Sum C 2 ＝{y 1 ，y 2 ，...，y m-1 Dissimilarity DisSim of } 1 (C 1 ，C 2 ) The definition is as follows:
where n represents the number of cyclic partition templates and m represents the number of spinal blocks; p (P) m The mth endpoint, y, representing the right outer contour of the spinal block i Is the shape code between the i-1 th endpoint and the i-th endpoint in another spine image;
3.2 Let s= { p 1 ，p 2 ，...，p m Sum S '= { p' 1 ，p′ 2 ，...，p′ m Respectively, is the outer contour endpoint set, C, of the two segmented vertebral block images X and Y 1 ＝{x 1 ，x 2 ，...，x m-1 Sum C 2 ＝{y 1 ，y 2 ，...，y m-1 The shape coding sequences corresponding to X and Y are respectively PL 1 ＝{x 1 ，x 2 ，...，x m-1 Setting PB for the coordinates of the left lower corner point of the first to m-1 th cone blocks of the X cone blocks of the vertebra 1 ＝{y 2 ，y 3 ，...，y m Setting PK for coordinates of upper left corner points from 2 nd to m nd cone blocks of the X cone blocks of the vertebra 1 For PL 1 With PB 1 Slope formed by the corresponding points;
set PL 2 ＝{x 1 ，x 2 ，...，x m-1 The sign of the Y-shaped cone block is the coordinates of the left lower corner point of the first to m-1 th cone blocks of the vertebra, PB 2 ＝{y 2 ，y 3 ，...，y m Setting PK for coordinates of upper left corner points of the 2 nd to the m th cone blocks of the Y-cone blocks of the vertebra 2 For PL 2 With PB 2 The slope formed by the corresponding points,
PK n-i the calculation method of (1) is that
x i [1]The ordinate information of the left lower corner point of the cone block and x are stored i [0]Stored is the abscissa information y of the left lower corner point of the cone block i+1 [1]The ordinate information of the right lower corner point of the cone block, y is stored i+1 [0]Storing abscissa information of right lower corner point of cone block
PK n-i Wherein n takes the values of X and Y and corresponds to the finger cone block X and the cone block Y
For PK n-i If positive and negative changes of the slope occur, local maximum change slopes occur, and more than or equal to 1 local maximum change slope exists in the spine; for S and S', if the boundary line segment p i-1 -p i If there is a slope of local maximum change, then define P i I > 0 is the key point, p i-1 -p i Given the number of keypoints of any segmented spine image, the keypoints in S and S' are denoted as { kp }, respectively j J=1,.. j * J=1,..k; next, each key point P is calculated using equation (5) i Matching weight w of i > 0 i ：
Wherein the operation represents a separation distance between two points on the outer contour of the spinal block;
finally, for two sets of code sequences C 1 ＝{x 1 ,x 2 ,...,x m Sum C 2 ＝{y 1 ,y 2 ,...,y m The degree of dissimilarity between } is further optimized as:
3.3 Set of left outer contour endpoints s= { p for given segmented spinal block X 0 ,p 1 ,...,p m Respectively across each end point p i Make horizontal line l i In l i Calculating the sum p in the clockwise direction as a reference i Angle theta of corresponding upper or lower boundary i Then define equation [8 ]]To calculate the horizontal inclination angle alpha of the spine i In practiceAbove, equation [8 ]]Implying the horizontal angle inclination direction of the vertebral block if theta i Less than or equal to 90 degrees, meaning alpha i Taking l i The angle at which the clockwise direction intersects the upper and lower boundaries, i.e. alpha i Is a positive number; if theta is i > 90 °, then means α i Taking l i The angle at which the inverted needle direction intersects the upper and lower boundaries, i.e. alpha i Is a negative number;
let X and Y be the two groups of spines to be matched after segmentation, S= { p 0 ,p 1 ,...,p m Sum S '= { p' 0 ,p′ 1 ,...,p′ m Respectively their outer contour endpoint sets, C 1 ＝{x 1 ,x 2 ,...,x m Sum C 2 ＝{y 1 ,y 2 ,...,y m The shape coding sequences corresponding to X and Y are respectively, and alpha= { alpha } is 0 ,α 1 ,...,α m Sum beta= { beta } 1 ,β 2 ,...,β m And the horizontal inclination angles of X and Y are respectively, the dissimilarity degree of X and Y is further optimized as:
2. the scoliosis Lenke typing system according to claim 1, wherein X and Y are set as two vertebrae to be matched after segmentation, (1) first, the wrapping rectangles of X and Y are calculated respectively, and the distances from all outer contour endpoints on the left side of the vertebrae to the right boundary of the wrapping rectangles are calculated respectively as: h is a X ＝{h 0 ,h 1 ,...,h m And d y ＝{d 0 ,d 1 ,...,d m "wherein the mark with the greatest distance is h max And d max ；h X Representing the set of distances from all outline points on the left side of the vertebral block on vertebra X to the right boundary of their bounding rectangle, numbered from0 start, d y Representing the distance set from all the outline points on the left side of the vertebral block on the vertebra Y to the right boundary of the outsourcing rectangle, starting from 0 and h m Representing the distance set from all the outer contour points on the left side of the vertebral block with the sequence number m to the right boundary of the outer rectangle of the vertebral block with the sequence number m on the vertebra X, and dm represents the distance set from all the outer contour points on the left side of the vertebral block with the sequence number m to the right boundary of the outer rectangle of the vertebral block with the sequence number m on the vertebra Y;
(2) turning Y along a vertical line to obtain a mirror image of Y, marking as Y ', further calculating an outsourcing rectangle of Y ', and similarly calculating the distance from all outer contour endpoints on the left side of the vertebra block in Y ' to the right boundary of the outsourcing rectangle, wherein the distance is marked as: d, d y ′＝{d 0 ′,d 1 ′,...,d m ′}；
(3) Calculating the optimal symmetrical matching distance of X and Y as dis (X, Y) according to a formula [11 ];
the general formula [10] and the formula [11] are defined as the shape difference degree between X and Y:
wherein epsilon, lambda, delta are weight factors; the smaller the Dissimilar (X, Y) is, the closer the two spine block images are, the scoliosis Lenke typing unit obtains K 'standard pictures closest to the full spine X-ray image to be detected in the Lenke typing database unit according to a formula [12], the K' images have the most types of Lenke typing categories, and the Lenke typing categories of the spine block images in the full spine X-ray image to be detected are determined.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211507294.8A CN115713661B (en) | 2022-11-29 | 2022-11-29 | Scoliosis Lenke parting system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211507294.8A CN115713661B (en) | 2022-11-29 | 2022-11-29 | Scoliosis Lenke parting system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115713661A CN115713661A (en) | 2023-02-24 |
| CN115713661B true CN115713661B (en) | 2023-06-23 |
Family
ID=85235181
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211507294.8A Active CN115713661B (en) | 2022-11-29 | 2022-11-29 | Scoliosis Lenke parting system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115713661B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117392120A (en) * | 2023-12-07 | 2024-01-12 | 北京维卓致远医疗科技发展有限责任公司 | Spine parameter processing method and system based on sagittal spine image |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108320288A (en) * | 2017-12-08 | 2018-07-24 | 李书纲 | A kind of data processing method of idiopathic scoliosis image |
| CN113205535A (en) * | 2021-05-27 | 2021-08-03 | 青岛大学 | X-ray film spine automatic segmentation and identification method |
| CN113888466A (en) * | 2021-09-03 | 2022-01-04 | 武汉科技大学 | Pulmonary nodule image detection method and system based on CT image |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108053394A (en) * | 2017-12-08 | 2018-05-18 | 李书纲 | A kind of Medical Imaging System |
| CN111374670A (en) * | 2020-05-01 | 2020-07-07 | 上海健康医学院 | Intelligent detection device and detection method for scoliosis of human body |
| CN111671454B (en) * | 2020-06-08 | 2023-10-31 | 平安科技（深圳）有限公司 | Spine bending angle measuring method, device, computer equipment and storage medium |
| CN112802019B (en) * | 2021-04-01 | 2021-06-29 | 成都成电金盘健康数据技术有限公司 | Leke typing method based on spine AIS image |
| CN114240893A (en) * | 2021-12-18 | 2022-03-25 | 长春理工大学 | Method for measuring and calculating spinal Cobb angle in external image |
| CN115100182A (en) * | 2022-07-19 | 2022-09-23 | 成都泽康智骨科技有限公司 | Method for typing vertebral column vertebral body fracture |
| CN115222937A (en) * | 2022-07-21 | 2022-10-21 | 北京深睿博联科技有限责任公司 | Method and device for detecting scoliosis |
| CN115063891B (en) * | 2022-08-19 | 2022-11-04 | 山东新蓝海科技股份有限公司 | Human body abnormal sign data monitoring method |
-
2022
- 2022-11-29 CN CN202211507294.8A patent/CN115713661B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108320288A (en) * | 2017-12-08 | 2018-07-24 | 李书纲 | A kind of data processing method of idiopathic scoliosis image |
| CN113205535A (en) * | 2021-05-27 | 2021-08-03 | 青岛大学 | X-ray film spine automatic segmentation and identification method |
| CN113888466A (en) * | 2021-09-03 | 2022-01-04 | 武汉科技大学 | Pulmonary nodule image detection method and system based on CT image |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115713661A (en) | 2023-02-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| El-Feghi et al. | Automatic localization of craniofacial landmarks for assisted cephalometry | |
| Yue et al. | Automated 2-D cephalometric analysis on X-ray images by a model-based approach | |
| WO2019232960A1 (en) | Automatic bone age prediction method and system, and computer device and storage medium | |
| CN110458158B (en) | Text detection and identification method for assisting reading of blind people | |
| CN112215858A (en) | Method and system for image segmentation and recognition | |
| CN115713661B (en) | Scoliosis Lenke parting system | |
| Peng et al. | Hybrid automatic lung segmentation on chest ct scans | |
| CN115187606B (en) | Juvenile idiopathic scoliosis PUMC typing method | |
| US20220005222A1 (en) | Automatic recognition method for measurement point in cephalo image | |
| Liu et al. | Automatic assessment of facial paralysis based on facial landmarks | |
| Fatemizadeh et al. | Automatic landmark extraction from image data using modified growing neural gas network | |
| Gupta et al. | Iris recognition system using biometric template matching technology | |
| Learned-Miller et al. | Detecting acromegaly: screening for disease with a morphable model | |
| CN115602320B (en) | Difficult airway assessment method and system | |
| CN108537194A (en) | A kind of expression recognition method of the hepatolenticular degeneration patient based on deep learning and SVM | |
| Gao et al. | Learning and synthesizing MPEG-4 compatible 3-D face animation from video sequence | |
| CN115331153B (en) | Posture monitoring method for assisting vestibule rehabilitation training | |
| CN109886320B (en) | Human femoral X-ray intelligent recognition method and system | |
| CN113327274B (en) | Lung CT image registration method and system integrating segmentation function | |
| Pilgram et al. | Knowledge-based femur detection in conventional radiographs of the pelvis | |
| Johnston et al. | A Fully Automated System for Sizing Nasal PAP Masks Using Facial Photographs | |
| Li et al. | Nasolabial Folds Extraction based on Neural Network for the Quantitative Analysis of Facial Paralysis | |
| US11610305B2 (en) | Method and system for postural analysis and measuring anatomical dimensions from a radiographic image using machine learning | |
| CN115148341B (en) | AI structure sketching method and system based on body position recognition | |
| CN116310282B (en) | Method and system for identifying focus in thoracoscopic image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |