Cochlea volute tube helix automatic extraction method based on temporal bone CT (Computed Tomography) image

Abstract

The invention discloses a cochlea volute tube helix automatic extraction method based on a temporal bone CT image, and belongs to the technical field of medical image processing. Through cochlea segmentation, cochlea tube helix extraction and cochlea tube helix nonlinear fitting, full-automatic analysis of the cochlea tube helix is achieved, the problems that traditional cochlea measurement is low in efficiency and large in error are solved, a high-precision anatomical evaluation tool is provided for artificial cochlea implantation, and the cochlea tube helix extraction method is suitable for cochlea implantation. The clinical practicability and scientific research value are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and in particular to a cochlea duct spiral automatic extraction method based on temporal bone CT images. BACKGROUND

[0002] The cochlea, as the core of the human auditory system, plays an indispensable role in the accurate perception and processing of sound due to its unique spiral shape. Cochlear implantation, as a clinical method for treating patients with severe and profound sensorineural hearing loss, is closely related to the anatomical structure of the cochlea. Individual differences in cochlea size and shape significantly affect the implantation position of cochlear electrodes and the cochlea coverage, which in turn affects the final pitch discrimination ability. In particular, in surgeries aimed at preserving residual hearing, individualized selection of electrode length is of great significance. Short electrode insertion may result in poor functional outcomes.

[0003] Therefore, during the surgical planning phase, accurate knowledge of the cochlea's geometric shape, especially the relevant parameters of the cochlea spiral, is crucial for determining the optimal cochlea coverage for cochlear stimulation and improving surgical outcomes.

[0004] Currently, clinical practice relies on ultra-high resolution cone beam computed tomography imaging (U-HRCT) to image the cochlea of patients. Based on this, cochlear implant clinicians or radiologists need to measure the geometric morphological parameters of the cochlea, such as A-diameter, B-diameter length, cochlea height H, base circumference, duct length, etc., for preoperative assessment. However, existing cochlea measurement methods have many problems. On the one hand, on the U-HRCT image, the clinician needs to carefully outline the cochlea region, determine the key points, and then perform three-dimensional reconstruction based on the outlined results. Finally, a three-dimensional spiral is used to fit the three circles to obtain various morphological parameters. Due to the high resolution of U-HRCT imaging, a large number of tomographic sections are required for clinical cochlea region outlining, resulting in a time-consuming outlining process and repetitive labor, which greatly increases the workload of clinicians. On the other hand, due to the very fine structure of the cochlea on U-HRCT imaging, manual outlining and measurement errors are large, and the measurement results of different qualified doctors have high differences and low consistency, which greatly disturbs cochlear implant planning.

[0005] In addition, although the automatic segmentation method based on 3D UNET combined with patch algorithm and the like has appeared in the prior art, attempts are made to replace manual operation, but such method still has significant limitations: due to the close proximity of the cochlear basal turn and the middle turn, the segmentation result is prone to adhesion in the U-HRCT image, and the prior art relies on ideal segmentation results, and the conventional segmentation model and simple processing flow (such as three-dimensional refinement, two-dimensional mapping combined with spiral line extraction mode of statistical intersection) are difficult to cope with such adhesion problems, often leading to mixing of redundant trajectories during spiral line extraction, which cannot accurately reflect the real anatomical structure of the cochlea, and further affects the accuracy of geometric parameter measurement, and it is difficult to meet the accuracy requirements of the clinic.

[0006] In view of the above status, there is an urgent practical need to develop a method for automatically extracting cochlear spiral lines. SUMMARY

[0007] The purpose of the present application is to provide a cochlear spiral line automatic extraction method based on temporal bone CT images to solve the problems in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides a cochlear spiral line automatic extraction method based on temporal bone CT images, comprising the following steps:

[0009] S1, using a TransUnet network to process the cochlear medical image to obtain a cochlear segmentation result; constructing a voxel model, combining the cochlear structure with the voxel model, and generating a three-dimensional point cloud model of the cochlear;

[0010] S2, using morphological operation, principal component analysis, spatial topology analysis algorithm, minimum spanning tree algorithm and depth-first search algorithm to extract the cochlear spiral line from the three-dimensional point cloud model to obtain a cochlear spiral line coordinate set;

[0011] S3, using a logarithmic spiral equation to nonlinearly fit the results of S2 to output a cochlear spiral line simulation result.

[0012] Preferably, in S1, the specific steps of cochlear segmentation using the TransUnet network are as follows:

[0013] (1) using CNN to extract features from the cochlear medical image to obtain feature maps at different scales;

[0014] (2) remodeling the feature maps into a sequence form and inputting them into the Transformer for long-distance dependency relationship capture;

[0015] (3) combining the features output by the Transformer with the feature maps based on up-sampling and feature fusion strategy, and outputting the cochlear segmentation mask after training and evaluation.

[0016] Preferably, the specific process of the training and evaluation is: adopting cross-entropy loss function and Dice loss function to adjust network parameters and hyperparameters, introducing transfer learning strategy to accelerate convergence, and completing training; adopting multiple indicators to evaluate segmentation quality, and optimizing network performance through active learning.

[0017] Preferably, the specific steps of S2 are:

[0018] S21, based on the original cochlea medical image, the cochlea segmentation result is optimized and calibrated, and the standard observation surface of the cochlea is determined;

[0019] S22, the neighborhood relationship analysis in the depth direction is carried out on the calibrated cochlea segmentation result, and the smooth spiral trajectory of the cochlea edge is extracted;

[0020] S23, the basal region of the calibrated cochlea segmentation result is identified, the edge voxel coordinate set of the cochlea base is extracted by using the neighborhood relationship analysis in S22, and the edge voxel coordinate set is divided into cochlea base outer wall to cochlear axis extreme far point group and cochlea base inner wall to cochlear axis extreme near point group;

[0021] S24, the cochlea base inner wall to cochlear axis extreme near point group is removed from the result of S22, and the cochlea outer wall to cochlear axis extreme far point set is obtained;

[0022] S25, the minimum spanning tree algorithm and the depth first search algorithm are used to extract the trajectory of the cochlea outer wall to cochlear axis extreme far point set, and the cochlea spiral line is obtained.

[0023] Preferably, the specific steps of S21 are:

[0024] (1) based on the three-dimensional point cloud model, the original cochlea medical image is separated from the foreground and the background by using threshold segmentation algorithm;

[0025] (2) the mutually connected foreground pixels are identified as the same connected region, and each connected region is assigned a unique identifier;

[0026] (3) the largest connected region in volume is screened out, and the preliminary optimization of the segmentation result is completed after restoring the pixel value;

[0027] (4) the morphological opening operation is used to smooth the preliminary optimized segmentation result;

[0028] (5) for all coordinates of the cochlea voxels processed by step (4), principal component analysis is used for calibration, and the cochlea standard observation surface is determined by traversing each layer of the calibrated cochlea segmentation result.

[0029] Preferably, in S21, the calculation formula for calibration is:

[0030]

[0031] wherein B1 represents the coordinate matrix after rotation calibration, A represents the original coordinate matrix of the cochlea voxels, M1 represents the affine transformation matrix, x i , y i , z i represent the three-dimensional coordinates of the i-th voxel in the original space, respectively, m jk is the element of the j-th row and k-th column of the affine transformation matrix.

[0032] Preferably, the specific steps of S22 are as follows:

[0033] (1) surface reconstruction: performing surface reconstruction on the calibrated cochlea segmentation result to generate a surface model containing topological information;

[0034] (2) cochlea wall neighborhood relationship analysis: establishing a spatial neighborhood relationship network in the depth direction for all voxel coordinate points in the surface model, screening out a coordinate point set exceeding a neighborhood quantity threshold, grouping the coordinate points in the same vertical trajectory and different depths in the coordinate point set, calculating the average value of each group, and generating a smooth spiral trajectory of the cochlea edge;

[0035] The smooth spiral trajectory of the cochlea edge contains the cochlea outer wall to the cochlear pole far point and the cochlea partial basal inner wall to the cochlear pole near point.

[0036] Preferably, the specific steps of S23 are as follows:

[0037] (1) base identification: identifying two independent connected regions in the calibrated sagittal cochlea segmentation result, taking the smaller connected region as the cochlea base, integrating all sagittal cochlea bases, and generating a cochlea base model;

[0038] (2) edge grouping: extracting the edge voxel coordinate set of the base using the neighborhood relationship analysis in S22, and dividing the edge voxel coordinate set into two groups according to the characteristics of the coordinate points;

[0039] (3) cochlear axis positioning: observing the voxel distribution of the transverse cochlea segmentation result, determining the coordinates of the cochlear axis through statistical average calculation;

[0040] (4) cochlear pole near point screening: respectively calculating the spatial distance between the two groups of coordinate points in step (2) and the cochlear axis, and taking the group with smaller distance as the cochlear base inner wall to the cochlear pole near point group.

[0041] Preferably, the specific steps of S25 are as follows:

[0042] (1) calculating the Euclidean space distance between all coordinate points in the cochlea outer wall to cochlear pole far point set, and constructing a distance matrix;

[0043] (2) Adopting minimum spanning tree algorithm to optimize the distance matrix, generating the shortest path tree connecting all coordinate points and identifying the endpoints;

[0044] (3) Based on the result of step (2), adopting depth-first search to find the longest path connecting all endpoints, and defining the node coordinate set on the longest path as the spiral line coordinate set.

[0045] Preferably, the specific steps of S3 are: converting the spiral line coordinate set from Cartesian coordinate system to polar coordinate system, and simulating the spiral line by adopting logarithmic spiral equation; and adopting nonlinear least squares method to iteratively adjust the parameters in the logarithmic spiral equation to find the optimal parameter solution.

[0046] In the formula, r represents the polar radius, a is the initial radius and the scaling parameter, b is the spiral expansion rate parameter, θ is the polar angle, and θ0 is the initial angle offset.

[0047]

[0048] In the formula, r represents the polar radius, a is the initial radius and the scaling parameter, b is the spiral expansion rate parameter, θ is the polar angle, and θ0 is the initial angle offset.

[0049] Therefore, the cochlear duct spiral line automatic extraction method based on temporal bone CT image has the following beneficial effects:

[0050] (1) The TransUnet network realizes a better cochlear structure segmentation basis, and combined with the maximum connected region analysis introduced in the morphological operation, can effectively remove the noise not connected with the cochlea, and provide a clear segmentation premise for processing the adhesion structure; further refining the spiral line extraction into 5 sub-steps, through spatial topology and graph theory algorithm (including neighborhood relationship analysis, base identification and grouping, etc.), the adhesion area of cochlear bottom turn and middle turn can be distinguished, the adhesion structure can be accurately peeled off, the redundant trajectory can be avoided, and the spiral line closely fitted to the real anatomical morphology of the cochlea can be ensured, so that the robustness and accuracy of spiral line extraction under complex adhesion conditions are significantly improved.

[0051] (2) Relying on the segmentation capability of the TransUnet network and the multi-link automatic algorithm (morphological operation, spatial topology analysis, etc.), the whole process automation of cochlear structure segmentation and spiral line coordinate extraction is realized, completely replacing the traditional tedious operation of relying on doctors to manually delineate the cochlear region and label layer by layer, greatly reducing human error; the U-HRCT image can be efficiently processed, the segmentation adhesion problem can be accurately solved, the preoperative evaluation time can be significantly shortened, the doctor's repetitive labor burden can be reduced, and more reliable and efficient cochlear structure analysis results can be provided for clinical use.

[0052] (3) Based on the segmentation algorithm of deep learning and the automatic spiral line extraction process, the capture ability of the cochlea fine structure is enhanced, the subjective error of manual drawing is effectively overcome, the consistency of the measurement result is improved, and high-precision key parameters are provided for the artificial cochlea implant planning.

[0053] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is the overall block diagram of the embodiment of the present application.

[0055] Figure 2 It is the TransUnet network architecture diagram of the embodiment of the present application.

[0056] Figure 3 It is the cochlea schematic diagram after calibration of the embodiment of the present application.

[0057] Figure 4 It is the extracted cochlea standard observation surface of the embodiment of the present application.

[0058] Figure 5 It is the extraction schematic diagram of the cochlea outer wall to the cochlear pole far point and the cochlea partial basal inner wall to the cochlear pole near point of the embodiment of the present application.

[0059] Figure 6 It is the cochlea basal separation schematic diagram of the embodiment of the present application.

[0060] Figure 7 It is the cochlea basal outer wall to cochlear pole far point and cochlea basal inner wall to cochlear pole near point schematic diagram of the embodiment of the present application, wherein A represents the cochlea basal outer wall to cochlear pole far point; B represents the cochlea basal inner wall to cochlear pole near point.

[0061] Figure 8 It is the cochlea cochlear spiral line nonlinear fitting schematic diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0062] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples.

[0063] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with the help of the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments.

[0064] EMBODIMENT

[0065] As Figure 1As shown, the application provides an automatic cochlear duct spiral line extraction method based on temporal bone CT image, which is realized through three parts of cochlea segmentation, cochlear duct spiral line extraction and cochlear duct spiral line nonlinear fitting, including the following steps:

[0066] S1, using TransUnet network (or other deep network model or method capable of accurately segmenting the cochlea, such as U-net, U-net++, MedSam, etc.) to process the cochlear medical image, and obtaining the cochlea segmentation result;

[0067] As shown, Figure 2 The overall framework of TransUnet is mainly composed of an encoder and a decoder. The encoder part combines CNN and Transformer structure. First, the CNN is used as an initial feature extractor to extract features from the cochlear medical image (such as CT, MRI, etc. Cochlea image data). Through multi-layer convolution and pooling operation, local features at different scales are extracted, which contain basic information such as edges and textures of the cochlea, and feature maps at different scales are obtained.

[0068] Then the feature map extracted by CNN is reshaped into a sequence and input into Transformer. The multi-head self-attention mechanism in Transformer plays a key role, which can model the feature sequence from a global perspective and capture the long-distance dependency relationship between different position features, so as to accurately identify the spatial position and morphological features of the cochlea in complex anatomical structure.

[0069] The decoder part adopts up-sampling and feature fusion strategy. In the process of step-by-step up-sampling, the features output by the Transformer module are fused with the feature maps at the corresponding scale in the CNN stage, making full use of the local detailed features extracted by CNN and the global semantic information obtained by Transformer, restoring the spatial resolution of the image, and gradually generating accurate cochlea segmentation mask after training and evaluation.

[0070] During the training process, a hybrid loss function is used to optimize the network performance. The cross-entropy loss function and the Dice loss function are combined. The cross-entropy loss function focuses on the accuracy of classification, and the Dice loss function focuses on the overlap of the segmented area. The two work together to effectively improve the segmentation accuracy of the network for the boundaries and internal structure of the cochlea. At the same time, the transfer learning strategy is introduced, and the model parameters pre-trained on a large-scale medical image dataset are used to initialize the network, which accelerates the network convergence, reduces the amount of data required for training, and makes the model also achieve good segmentation effect on limited cochlea image data.

[0071] After training, the segmentation results are quality tested by various evaluation indicators such as Dice coefficient, Hausdorff distance, etc., to comprehensively measure the difference between the segmentation results and the true labels. At the same time, the segmentation results are superimposed and displayed with the original image using visualization technology, which is convenient for doctors to intuitively judge the accuracy of the segmentation.

[0072] For cases with poor segmentation results, they are fed back to the training process, and the data is revised and labeled by doctors through active learning, and the labeled data is re-input into the network for training, continuously optimizing the network model's ability to capture fine structures of the cochlea, and gradually improving the performance and reliability of TransUnet for cochlea segmentation.

[0073] In order to better adapt to the cochlea segmentation task, the network structure of TransUnet is improved in this embodiment: a special multi-class output head is designed at the output layer of the network to meet the segmentation needs of multiple sub-parts of the cochlea structure (such as cochlear axis, spiral canal, etc.).

[0074] In the technology of automatic cochlea segmentation, CT scanning technology is used to collect data; then, with the help of conventional image processing technology and three-dimensional reconstruction algorithm, each layer of the segmentation image generated by CT scanning is analyzed. By accurately identifying the boundary contour and geometric features of the voxel, combining with the two-dimensional image information for spatial integration, the transformation from two-dimensional tomographic data to complete three-dimensional point cloud data is finally completed, realizing the three-dimensional visualization reconstruction of the model.

[0075] S2, morphological operations, principal component analysis, spatial topology analysis algorithm, minimum spanning tree algorithm and depth-first search algorithm are used to extract the cochlear spiral line of the three-dimensional point cloud model, and the cochlear spiral line coordinate set is obtained, specifically:

[0076] S21, based on the original cochlear medical image, the cochlea segmentation result is optimized and calibrated to determine the standard observation surface of the cochlea, specifically:

[0077] (1) Based on the three-dimensional point cloud model, the threshold segmentation algorithm is applied to the original cochlear medical image, and the region in the image that meets the specific gray value range (corresponding to the cochlear structure) is set as the foreground value, and the rest is set as the background value, separating the foreground and background;

[0078] (2) Identify the connected foreground pixels as the same connected region, and assign a unique identifier to each connected region to generate a labeled image;

[0079] (3) The regions in the marked image are sorted based on the volume of the connected regions, and are numbered from large to small. The largest connected region with the largest volume is reserved as the main target structure (i.e., the cochlea), and the remaining regions are regarded as noise or non-target structures. The largest connected region selected is restored to the pixel value representation of the original segmentation result, and the remaining regions are kept as background values, so as to remove the redundant voxels, and finally obtain an accurate segmentation result containing only the target cochlea structure, thereby completing the preliminary optimization.

[0080] (4) The morphological opening operation is performed on the segmentation result after the preliminary optimization to achieve smoothing, and the specific operation is as follows:

[0081] According to the resolution of the cochlea segmentation result, an appropriate operation kernel is set for the morphological operation on the cochlea. First, the morphological erosion operation is performed on the cochlea segmentation result to eliminate small protrusions and noise points in the segmentation result, and to reduce the overall size of the target region. Then, the morphological dilation operation is performed to fill the small holes inside and connect the adjacent segmentation regions, so as to form a smoother and more continuous boundary. After the opening operation, the reconstructed cochlea segmentation result has a smooth surface.

[0082] (5) For all coordinates of the cochlea voxels after the step (4) processing, principal component analysis is performed for calibration, and the cochlea standard observation surface is determined by traversing each layer of the calibrated cochlea segmentation result, and the specific operation is as follows:

[0083] All coordinates of the cochlea voxels after the step (4) processing are traversed and recorded as coords1. The PCA (Principal Components Analysis) is performed on the cochlea voxel coordinates, and three eigenvectors of the principal component analysis are used to form a 3x3 affine transformation matrix M1. The cochlea segmentation result is rotated to achieve spatial position calibration by using the following calibration formula:

[0084]

[0085] In the formula, B1 represents the coordinate matrix after rotation and calibration, A represents the original coordinate matrix of the cochlea voxels, M1 represents the affine transformation matrix, x i , y i , and z i represent the three-dimensional coordinates of the i-th voxel in the original space, and m jk is the element of the j-th row and k-th column of the affine transformation matrix.

[0086] After calibration, the cochlear axis is parallel to the Z axis (depth direction) and perpendicular to the transverse plane. In the calibrated cochlea segmentation result, the transverse plane is traversed every slice, and the number of cochlea voxels is calculated. The section with the largest number of voxels is the standard observation surface of the cochlea.

[0087] S22, the neighborhood relationship analysis in the depth direction is performed on the calibrated cochlea segmentation result, and a smooth spiral trajectory of the cochlea edge is extracted, and the specific steps are as follows:

[0088] (1) Surface reconstruction: the surface reconstruction is performed on the calibrated cochlea segmentation result, and a three-dimensional surface model containing complete topological information is generated;

[0089] (2) Cochlea wall neighborhood relationship analysis: the spatial neighborhood relationship network in the depth direction (z-axis) is established for all voxel coordinate points in the surface model, and whether there is a neighboring point with the same x and y coordinates but different z coordinates by 1 is specifically analyzed for each coordinate point, and the number of neighboring points meeting the condition is counted. Set the neighborhood number threshold, and screen out the coordinate point set (with high spatial continuity) exceeding the neighborhood number threshold. Through threshold screening, noise interference is effectively excluded, and the significance of edge features is enhanced.

[0090] The screened coordinate points are spatially grouped, the coordinate points with the same x and y coordinates and different depths in the coordinate point set are regarded as sampling points on the same vertical trajectory for grouping, the average value of each group in the z-axis direction is calculated, and a new coordinate point set is generated. The set represents the smooth spiral trajectory of the cochlea duct edge, and contains the cochlea outer wall to the cochlea polar far point and the cochlea partial basal inner wall to the cochlea polar near point.

[0091] S23, the basal region of the calibrated cochlea segmentation result is identified, and the edge voxel coordinate set of the cochlea base is extracted by using the neighborhood relationship analysis in S22, and the edge voxel coordinate set is divided into a cochlea base outer wall to cochlea polar far point group and a cochlea base inner wall to cochlea polar near point group, and the specific steps are as follows:

[0092] (1) Basal identification: observe the calibrated cochlea segmentation result on the sagittal plane two-dimensional slice, identify the characteristic structure presenting as two independent connected regions; through area comparison, the smaller connected region is regarded as the cochlea base, the smaller connected region containing the cochlea base feature is extracted and retained by traversing all sagittal plane slices, and the two-dimensional slice information is integrated into a complete three-dimensional cochlea base model through conventional spatial superposition and three-dimensional reconstruction technology.

[0093] (2) Edge grouping: the edge voxel coordinate set of the three-dimensional cochlea base model is extracted by using the neighborhood relationship analysis (spatial topological analysis algorithm) in S22, and the edge voxel coordinate set is divided into two groups according to the characteristics of the coordinate points in the set;

[0094] (3) Cochlea axis positioning: observe the voxel distribution of the cochlea lower specific layer segmentation result on the transverse cross-section, and determine the accurate coordinates of the cochlea cochlea axis through statistical average calculation;

[0095] (4) The extreme near point screening: the total space distance between the two groups of coordinate points in step (2) and the cochlear axis is calculated respectively, and the group with smaller distance is defined as the cochlear basal inner wall to the cochlear axis extreme near point group.

[0096] S24, removing the cochlear basal inner wall to the cochlear axis extreme near point group from the results of S22 (the cochlear outer wall to the cochlear axis extreme far point and the cochlear basal inner wall to the cochlear axis extreme near point), to obtain the cochlear outer wall to the cochlear axis extreme far point set;

[0097] S25, using the minimum spanning tree algorithm and the depth-first search algorithm to extract the cochlear outer wall to the cochlear axis extreme far point set to obtain the cochlear spiral, the specific steps are as follows:

[0098] (1) Calculate the Euclidean space distance between all coordinate points in the cochlear outer wall to the cochlear axis extreme far point set, and construct a complete distance matrix, which reflects the spatial topological relationship between the points.

[0099] (2) Using the minimum spanning tree algorithm (Minimum Spanning Tree, MST) to optimize the distance matrix, generating a tree structure connecting all coordinate points with the shortest total path length. In the construction process, record the connection times of each node to form the node connection degree distribution. Based on the node connection degree analysis, identify the end points (leaf nodes) with connection times of 1, which are usually located at the starting or ending position of the spiral line.

[0100] (3) Using the adjacency list form to construct the graph data structure, which completely stores the connection relationship between nodes. Using the depth-first search algorithm to traverse the graph structure to find the longest path connecting all end points, which represents the main spiral trajectory of the cochlear spiral, effectively excluding noise interference and branch structure. The node coordinate set on the longest path is defined as the cochlear spiral line coordinate set.

[0101] S3, according to the extracted cochlear spiral and the standard observation surface of the cochlea, using the logarithmic spiral equation to nonlinearly fit the results of S2, and outputting the cochlear spiral simulation results, the specific steps are as follows:

[0102] Convert the spiral line coordinate set from Cartesian coordinate system to polar coordinate system, and based on the biological characteristics of the cochlea, use the logarithmic spiral equation to simulate the spiral line:

[0103]

[0104] In the formula, r represents the polar radius, parameter a determines the initial radius and overall scaling ratio of the spiral line, parameter b controls the exponential rate of the radius growth with angle, directly affecting the expansion strength and layer density of the spiral, θ is the polar angle, and θ0 is the initial angle offset.

[0105] The nonlinear least square method is applied to find the optimal parameter solution by iteratively adjusting the parameter combination to minimize the mean square error (MSE) between the model prediction value and the actual data point. The average distance square between the theoretical spiral line and the actual scatter point under different parameter combinations is calculated in each iteration, and gradually converges to the optimal solution.

[0106] The following uses the cochlear spiral automatic extraction method based on the temporal bone CT image proposed in the application to automatically extract the cochlear spiral of the NIfTI (Neuroimaging Informatics Technology Initiative) format data of the temporal bone U-HRCT. The specific steps are as follows:

[0107] A complete scan result is composed of bilateral ear data, and the axial single-sided data contains 370 images. The scan field of view is 65mmx65mmx37mm, the spatial resolution is 650x650x370, the slice thickness is 0.1mm, the slice interval is 0.1mm, the tube voltage is 100KV, and the tube current is 3.5mA.

[0108] S1, a deep learning method is used to complete the segmentation task of the cochlea. A pre-trained TransUnet network is selected as the core model, which combines the powerful global feature capture ability of Transformer and the efficient feature extraction and recovery structure of U-Net, and can effectively process complex structural information in medical images. The cochlea is segmented from each complete scan result.

[0109] S21, the BinaryThreshold function of the SimpleITK library is used to perform threshold processing on the original image, and the cochlea segmentation result gray pixel value is set to foreground value 1 and the rest is set to background value 0, to realize selective extraction of a specific gray value region; then the ConnectedComponent algorithm is used to label the connected region of the binary image, and the mutually connected foreground pixels are classified into the same region and assigned a unique identifier to form a labeled image;

[0110] The RelabelComponent function is used to relabel the labeled image and sort the regions by volume from large to small, and the largest region is kept as the main target structure; then the BinaryThreshold function is applied again to restore the pixel value of the largest connected region to the pixel value of the cochlea segmentation result, and the remaining regions remain 0, to obtain the final segmentation result containing only the largest target structure.

[0111] The BinaryErode and BinaryDilate functions in the SimpleITK library in Python are used to perform opening and closing operations on the cochlea segmentation result after preliminary optimization to smooth the cochlea surface, and the cochlea segmentation result voxel coordinates are recorded in coords1.

[0112] The PCA principal component analysis is performed on the cochlea segmentation result voxel coordinate set coords1, three characteristic vectors constitute a 3x3 affine transformation matrix M1, spatial position calibration is performed, and the cochlea voxel coordinates after calibration are recorded as coords2. The cochlea axis after calibration is parallel to the Z axis as shown in Figure 3 . At this time, the slice with the largest number of cochlea segmentation result voxels is traversed in the Z axis slice, and the slice is the standard observation surface of the cochlea as shown in Figure 4 .

[0113] S22, the surface of the cochlea segmentation result after S21 processing is extracted by using the Marching Cubes algorithm, the voxel coordinates of the cochlea surface are recorded as coords3, the threshold is set to 3, all the voxel coordinates of the cochlea surface are traversed, and it is checked whether they are adjacent in the z-axis direction, that is, the x and y coordinates of the coordinates are the same, the z coordinate difference is 1, and the number of adjacent points is more than three. If more than three, record these coordinates as coords4. The average value of the z coordinates of all x and y coordinates of coords4 is calculated, and the average value is recorded as the coordinate saved in coords5 to obtain the smooth spiral trajectory of the cochlear duct edge, that is, the cochlea outer wall to the cochlea farthest point and the cochlea partial basal wall to the cochlea closest point, as shown in Figure 5 .

[0114] S23, as shown in Figure 6 , the connected component analysis is performed on the sagittal slice, and the cochlea base is separated. The distance between the cochlea base and the cochlea axis farthest and closest points recorded as coords6 is extracted by using the algorithm in S22. According to the average value of the y-axis coordinates in coords6, two groups are divided, as shown in Figure 7 , which are the cochlea base outer wall to the cochlea axis farthest point group and the cochlea base inner wall to the cochlea axis closest point group. In the transverse section, all the cochlea voxels of the cochlea top 10 layers of the cochlea segmentation result after the fourth step processing are observed, and the average coordinates of these voxels are calculated, which are the cochlea axis coordinates (x1, y1). The group with smaller sum of distances from all coordinates in the two groups to the cochlea axis is the cochlea base inner wall to the cochlea axis closest point group, recorded as coords7.

[0115] S24, the cochlea base inner wall to the cochlea axis closest point coordinates are removed from the cochlea outer wall to the cochlea axis farthest point and the cochlea partial basal wall to the cochlea axis closest point set coords5 obtained in S22 and recorded as coords8. After that, coords8 only contains the cochlea outer wall to the cochlea axis farthest point coordinates.

[0116] S25, calculate the Euclidean distance between all coordinates in coords8, construct the minimum spanning tree (MST) to connect all points and ensure the lowest cost, and record the connection times of each point. Extract the leaf nodes with connection times of 1 in the bone line, and build a graph with an adjacency list to store the point connection relationship. Find the longest path by depth-first search (DFS) algorithm, and keep the edge line coordinates on the path as coords9 (i.e. cochlea spiral line coordinates) to realize denoising. The length of the longest path is the cochlea length.

[0117] S3, as shown in Figure 8 The cochlea spiral line coordinates coords 10 are converted into polar coordinates, and the fitting results of the cochlea spiral line are output by fitting the logarithmic spiral formula.

[0118] Finally, the length of the cochlea spiral line obtained by measurement can assist the electrode implantation of cochlear implant surgery, and the nonlinear parameters obtained can also be used for related cochlear scientific research. Not only can it liberate the clinical productivity and reduce the work pressure of doctors, but also can improve the measurement accuracy and provide more accurate and reliable preoperative evaluation basis for cochlear implant surgery, thereby promoting the development of the field of audiology and improving the quality of life of the general hearing-impaired patients.

[0119] Therefore, the cochlea spiral line automatic extraction method based on the temporal bone CT image can significantly eliminate the subjective error and noise interference of manual delineation by deep learning segmentation and multi-algorithm collaborative processing (principal component analysis spatial calibration, neighborhood relationship continuity screening, minimum spanning tree topology optimization, and logarithmic spiral equation nonlinear fitting, etc.), and ensure that the extracted spiral line trajectory closely fits the real cochlear anatomical structure, thereby providing high-precision key morphological parameters (such as cochlea length and spiral curvature) for cochlear implant surgery and improving the reliability of preoperative planning.

[0120] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for automatically extracting cochlear spiral in the cochlear duct based on a temporal bone CT image, characterized in that, The method comprises the following steps: S1, using a TransUnet network to process a cochlea medical image to obtain a cochlea segmentation result; constructing a voxel model, combining the cochlea structure with the voxel model, and generating a three-dimensional point cloud model of the cochlea; S2, using morphological operations, principal component analysis, spatial topology analysis algorithm, minimum spanning tree algorithm and depth-first search algorithm to extract the cochlea spiral line of the three-dimensional point cloud model to obtain a cochlea spiral line coordinate set; S3, using a logarithmic spiral equation to perform nonlinear fitting on the result of S2 to output a cochlea spiral line simulation result.

2. The method of claim 1, wherein the method is based on a CT image of a temporal bone. In S1, the specific steps of using the TransUnet network for cochlea segmentation are: (1) using CNN to extract features of the cochlea medical image to obtain feature maps at different scales; (2) remodeling the feature maps into a sequence form and inputting into the Transformer for long-distance dependence relationship capture; (3) combining the features output by the Transformer with the feature maps based on up-sampling and feature fusion strategy, and outputting the cochlea segmentation mask after training and evaluation.

3. The method of claim 2, wherein the method further comprises: The specific process of training and evaluation is: adjusting network parameters and hyperparameters using cross-entropy loss function and Dice loss function, introducing a transfer learning strategy to accelerate convergence, and completing training; evaluating the segmentation quality by using multiple indicators, and optimizing the network performance through active learning.

4. The method of claim 1, wherein the method further comprises: The specific steps of S2 are: S21, optimizing and calibrating the cochlea segmentation result based on the original cochlea medical image to determine a standard observation surface of the cochlea; S22, performing smoothing spiral trajectory extraction on the cochlea edge of the calibrated cochlea segmentation result; S23, performing base area identification on the calibrated cochlea segmentation result, using the neighborhood relationship analysis in S22 to extract the edge voxel coordinate set of the cochlea base, and dividing the edge voxel coordinate set into a cochlea base outer wall to cochlear axis extreme far point group and a cochlea base inner wall to cochlear axis extreme near point group; S24, removing the cochlea base inner wall to cochlear axis extreme near point group from the result of S22 to obtain a cochlea outer wall to cochlear axis extreme far point set; S25, using the minimum spanning tree algorithm and the depth-first search algorithm to extract the trajectory of the cochlea outer wall to cochlear axis extreme far point set to obtain the cochlea spiral line.

5. The method of claim 4, wherein the method further comprises: The specific steps of S21 are: (1) based on the three-dimensional point cloud model, using a threshold segmentation algorithm to separate the foreground and background of the original cochlea medical image; (2) identifying the mutually connected foreground pixels as the same connected region, and assigning a unique identifier to each connected region; (3) screening out the largest connected region, and completing the preliminary optimization of the segmentation result after restoring the pixel value; (4) performing smoothing processing on the preliminary optimized segmentation result using morphological opening operation; (5) for all coordinates of the cochlea voxel after step (4), using principal component analysis for calibration, traversing each layer of the calibrated cochlea segmentation result, and determining a standard observation surface of the cochlea.

6. The method of claim 4, wherein the method further comprises: In S21, the calculation formula for calibration is: where A represents a matrix of original coordinates of cochlea voxels, M1represents an affine transformation matrix, x i , y i , z i represent three-dimensional coordinates of the i-th voxel in the original space, respectively, and m jk is an element of the j-th row and the k-th column of the affine transformation matrix.

7. The method of claim 4, wherein the method further comprises: determining a center of the cochlea based on the cochlea center point; and determining a center of the cochlea based on the cochlea center point. The specific steps of S22 are: (1) Surface reconstruction: surface reconstruction is performed on the calibrated cochlea segmentation result to generate a surface model containing topological information; (2) Cochlea wall neighborhood relationship analysis: a spatial neighborhood relationship network in the depth direction is established for all voxel coordinate points in the surface model, a coordinate point set exceeding a neighborhood quantity threshold is screened out, coordinate points in the same vertical trajectory and different depths in the coordinate point set are grouped, the average value of each group is calculated, and a smooth spiral trajectory of the cochlea edge is generated; The smooth spiral trajectory of the cochlea edge contains the cochlea outer wall to the cochlea polar far point and the cochlea partial basal inner wall to the cochlea polar near point.

8. The method of claim 4, wherein the method further comprises: The specific steps of S23 are: (1) Basal identification: two independent connected regions in the calibrated sagittal plane cochlea segmentation result are identified, and the smaller connected region is taken as the cochlea basal, the cochlea basal of all sagittal planes is integrated to generate a cochlea basal model; (2) Edge grouping: the edge voxel coordinate set of the basal is extracted by using the neighborhood relationship analysis in S22, and the edge voxel coordinate set is divided into two groups according to the characteristics of the coordinate points; (3) Modiolus positioning: the voxel distribution of the transverse plane cochlea segmentation result is observed, and the coordinates of the cochlea modiolus are determined by statistical average calculation; (4) Polar near point screening: the spatial distances of the two groups of coordinate points in step (2) from the modiolus are calculated respectively, and the group with smaller distance is taken as the cochlea basal inner wall to the cochlea polar near point group.

9. The method of claim 4, wherein the method further comprises: determining a center of the cochlea based on the CT image; and determining a center of the cochlea based on the 3D model. The specific steps of S25 are: (1) Calculate the Euclidean space distance between all coordinate points in the cochlea outer wall to cochlea polar far point set, and construct a distance matrix; (2) Optimize the distance matrix by using the minimum spanning tree algorithm to generate the shortest path tree connecting all coordinate points and identify the end points; (3) Based on the result of step (2), the longest path connecting all end points is found by using depth-first search, and the node coordinate set on the longest path is defined as the spiral line coordinate set.

10. The method of claim 1, wherein the method is based on a CT image of a temporal bone. The specific steps of S3 are: converting the spiral line coordinate set from Cartesian coordinate system to polar coordinate system, and using logarithmic spiral equation to simulate the spiral line; using nonlinear least squares method to iteratively adjust the parameters in the logarithmic spiral equation to find the optimal parameter solution; Wherein, r represents the polar radius, a is the initial radius and scaling parameter, b is the spiral expansion rate parameter, θ is the polar angle, and θ0 is the initial angle offset. In the formula, r represents the polar radius, a is the initial radius and scaling parameter, b is the spiral expansion rate parameter, θ is the polar angle, and θ0 is the initial angle offset.

Images