Automatic positioning method for ear multi-class standard planes in temporal bone U-HRCT image

Abstract

The invention discloses a temporal bone U-HRCT image ear multi-class standard plane automatic positioning method, and relates to the field of medical image processing, and the method comprises the steps: obtaining temporal bone U-HRCT image data, and carrying out the preprocessing of the image data, and obtaining an enhanced ossicular chain region of interest; defining the enhanced ossicular chain region of interest as an environment of a reinforcement learning framework, and obtaining respective states from the environment by the intelligent agents through respective plane parameters; calculating an average normal vector and an average intercept of each standard observation surface based on the training set label data, and taking the average normal vector and the average intercept as initial plane parameters of the intelligent agent; a multi-agent strategy network is adopted, a differentiated action space is customized for each standard observation surface, and plane positioning is guided through a dual reward function; evaluating a plane positioning result in real time by using a termination discrimination network, and determining whether to terminate the search process of the multi-agent strategy network according to the evaluation result; therefore, end-to-end automatic positioning of the standard observation surface of the ossicular chain is realized, and the positioning efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an automatic localization method for multiple standard planes of the middle ear in temporal bone U-HRCT images. Background Technology

[0002] The ossicular chain, the most intricate bony conduction system in the human body, consists of three tiny ossicles: the malleus, incus, and stapes. Through mechanical leverage, it transmits sound wave vibrations from the tympanic membrane to the oval window of the inner ear, completing the gas-fluid conversion of sound energy. Ossicular chain disorders, such as deformities, fractures caused by trauma, and joint dislocations, are common causes of conductive hearing loss. However, the ossicular chain is extremely small; the thinnest stapes footplate is only about 0.1 mm thick, making it difficult to image clearly using commonly used high-resolution CT equipment. Ultra-high-resolution computed tomography (U-HRCT) is an advanced medical imaging technique with a voxel spatial resolution of 0.1 mm isotropic. It provides a high-quality data foundation for medical research applications such as bone microstructure parameter analysis, diagnosis of coronary artery calcification, and limb skeletal imaging, and also provides a data foundation for the imaging diagnosis of ossicular chain disorders. U-HRCT significantly improves image resolution through optimized hardware design and imaging algorithms, and has important clinical application value in fields such as otology, dentistry, and head and neck surgery. In particular, it shows significant advantages in diagnosing minute lesions, complex anatomical abnormalities, and detecting early lesions.

[0003] However, diagnosing ossicular chain injuries based on temporal bone U-HRCT imaging requires radiologists to possess thorough knowledge of temporal bone anatomy and extensive clinical experience. Ossicular chain injuries can lead to ossicular displacement, fracture separation, or dislocation. The most common types are fractures of the three ossicles and dislocation of both joints, resulting in conductive hearing loss. The key anatomical structures of the temporal bone are characterized by their small size, intricate structure, and numerous diagnostic points. Specialized temporal bone radiologists typically diagnose ossicular chain fractures and dislocations by locating a set of standard observation planes (SOPs) on the images, based on the ossicular chain anatomy. These standard observation planes are the most clearly visible surfaces for ossicular chain lesions. Locating the SOPs reflects the physician's experience and is also the most time-consuming and labor-intensive process in image interpretation.

[0004] Therefore, an automatic localization method for the middle ear in multiple standard planes in temporal bone U-HRCT images is provided to solve the above problems. Summary of the Invention

[0005] To address the aforementioned challenges, this invention provides an automatic localization method for multiple standard planes of the middle ear in temporal bone U-HRCT images, achieving end-to-end automatic localization of the standard observation planes of the ossicular chain, thereby further improving localization efficiency.

[0006] To achieve the above objectives, this invention provides an automatic localization method for multiple standard planes of the middle ear in temporal bone U-HRCT images, comprising the following steps: S1: Acquire temporal bone U-HRCT image data, divide it into training set and validation set; and preprocess the temporal bone U-HRCT image data to obtain the enhanced ossicular chain region of interest; S2: Define the enhanced ossicular chain region of interest as the environment of the reinforcement learning framework, and describe the plane equations through standard formal parameterization; each agent obtains its own state from the environment through its own plane parameters; S3: Calculate the average normal vector and average intercept of each standard observation surface based on the training set label data, and use them as the initial plane parameters of the agent; S4: A multi-agent policy network is adopted to customize a differentiated action space for each standard observation surface and guide planar localization through a dual reward function; the multi-agent policy network is trained on the training set data and validated on the validation set; S5: The termination discriminant network is used to evaluate the planar localization results in real time, and the search process of the multi-agent policy network is terminated based on the evaluation results; the termination discriminant network is trained on the training set data and validated on the validation set. S6: Output the final localization result.

[0007] Preferably, the preprocessing in S1 specifically includes: S11: A deep learning-based segmentation network is used to segment the ossicular chain in the temporal bone U-HRCT image data to obtain the segmentation results; S12: Extract the region of interest of the ossicular chain based on the segmentation results; S13: Target saturation enhancement is performed based on the region of interest of the ossicular chain to obtain the enhanced region of interest of the ossicular chain.

[0008] Preferably, S11 specifically includes: The method using a deep learning model first performs coarse segmentation on the temporal bone U-HRCT image data to complete the localization of key points and regions of interest. Based on the localization points and regions of interest, fine segmentation is then performed, and the results are mapped back to the temporal bone U-HRCT image data to obtain the segmentation results within the regions of interest and the segmentation results of the complete data. The fine segmentation uses the TransUnet model.

[0009] Preferably, S2 specifically includes: S21: Define the enhanced ossicular chain region of interest as the environment of the reinforcement learning framework, and construct a Cartesian coordinate system with the center point of the ossicular chain segmentation result as the origin. S22: The plane equation is described by standard formal parameterization, specifically as follows: ; in, a , b , c Let be the unit normal vector of the plane. d The intercept in the direction of the normal vector is... a 2 + b 2 + c 2 =1, and d >=0; the value of a planar node is set to a quadruple { a , b , c , d} represents the parameters of the standard observation plane equation; S23: Each agent obtains its own state from the environment through its own planar parameters. In each iteration, the image of the current plane is extracted from the temporal bone U-HRCT data, the gray value range of the image is normalized to 0 to 1, the image is centered on the physical center point of the malleus segmentation result, and the image size is cropped to 224×224 voxels. S24: Three temporally adjacent optimization planes are stitched together into a three-channel image as the state input of the agent. Channel 1 is the plane information of the previous iteration step, channel 2 is the optimization result of the current iteration step, and channel 3 is the update region of the subsequent iteration.

[0010] Preferably, the dual reward function in S4 includes individual rewards and collaborative rewards; Individual rewards are used to guide the current plane to adjust towards the target plane, reducing the angle between the normal vectors and the intercept distance; collaborative rewards are used to constrain the relative positions between agents and optimize the positioning results by utilizing spatial positional relationships.

[0011] Preferably, the individual reward is represented as follows: ; Collaborative rewards are represented as follows: ; The double reward function is expressed as: ; in, sgn For symbolic functions, The function represents the distance between two planes. P1 ( a 1, b 1, c 1, d 1) and plane P 2 ( a 2, b 2, c 2, d 2) The function represents the Euclidean norm of the difference between two planar vectors. , P i For the first i The plane equation of the agent's current plane at the next iteration; P t No. i The plane equation of the target plane at the next iteration; P i-1 For the first i- In the first iteration, the plane equation of the agent's current plane; P j i For the first i During the nth iteration, the 1st j The plane equation of the current plane for each agent; P j t For the first i During the nth iteration, the 1st j The plane equation of the target plane for each agent; the abs function represents the absolute value of the distance.

[0012] Preferably, the policy network in S4 specifically includes: An improved VGG13 deep convolutional feature network is used as the planar localization agent. A 128×128 input size is adopted and five spatial downsampling operations are performed to generate 512-channel 7×7 high-order semantic features. A 512-dimensional compact feature is constructed through flattening and fully connected layers, which bifurcate into an advantage stream and a state value stream. The advantage stream calculates the relative advantage value of actions through a 128-dimensional hidden layer, while the state value stream evaluates the global state value. The Q-value is generated by fusing through a centralized advantage function. The policy network outputs the Q-value corresponding to each action. The action with the largest Q-value is selected as the predicted action. Based on the Q-value, the agent selects its optimal action parameters and obtains new planar parameters.

[0013] Preferably, the termination discrimination network in S5 is the EfficientNetv2 image classifier.

[0014] Preferably, the training process of the EfficientNetv2 image classifier specifically includes: Multiple standard observation surfaces are used as different categories, plus one category that is not any of the standard observation surfaces, forming a multi-class network. First, the network is trained using three standard observation surfaces and other planes as the initial training set. The trained planes are then manually evaluated. When the evaluation accuracy is below 95%, the terminated plane is added to the training set, and the network is retrained. This process is repeated until the evaluation accuracy is not lower than 95%, at which point the training ends, and the trained EfficientNetv2 image classifier is obtained.

[0015] Preferably, in S5, a decision is made based on the evaluation results to determine whether to terminate the search process for the multi-agent reinforcement learning network, specifically including: In each iteration of the deep reinforcement learning framework, after the agent obtains the optimal action and generates a new plane, the image information of the current plane is extracted and input into the trained EfficientNetV2 image classifier. If the classifier outputs that a certain plane belongs to the target plane, the agent terminates the iteration for this round, and other agents continue to iterate until the states of all agents are found. The three planes are then used as the localization results, and the search process of the multi-agent policy network is terminated.

[0016] Therefore, the present invention employs the above-mentioned method for automatic localization of the middle ear using multiple standard planes in temporal bone U-HRCT images, which has the following beneficial effects: (1) This invention utilizes three-dimensional temporal bone U-HRCT image data to segment the ossicular chain and extract the region of interest (ROI) of the ossicular chain. By eliminating differences in imaging equipment, spatial deformation of anatomical structures and noise interference, it achieves consistency and standardization of cross-modal and cross-scanning protocol data, thereby significantly improving the generalization ability and robustness of deep learning models in tasks such as lesion segmentation, anatomical landmark localization and quantitative analysis. (2) This invention proposes a hot start based on statistical data for reinforcement learning networks. By introducing this physical anatomy prior constraint, it effectively overcomes the problem of large ineffective oscillations of plane parameters caused by random strategies in the early stage of traditional reinforcement learning exploration, and significantly improves the search efficiency and stability in the initial stage.

[0017] (3) The present invention sets up intelligent agents with different action spaces for each SOP plane in the three-dimensional temporal bone U-HRCT image, so that the intelligent agents can locate the SOP plane more efficiently and accurately.

[0018] (4) The present invention sets a special reward mechanism for the spatial positional relationship of each SOP plane, which effectively improves the success rate of SOP plane localization, alleviates the situation of the agent getting stuck in local optima, and enables the reinforcement learning network to locate multiple planes at the same time, thereby improving localization efficiency.

[0019] (5) The present invention can adaptively and dynamically adjust the termination conditions according to image features for the automatic termination module of the reinforcement learning network utilization, thereby improving the accuracy and intelligence of termination judgment and effectively breaking through the dilemma of premature / late termination caused by the reliance on manual thresholds in traditional reinforcement learning.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the network architecture of an automatic localization method for multiple standard planes of the middle ear in temporal bone U-HRCT images according to the present invention; Figure 2 This is a schematic diagram of the training termination discrimination network in an embodiment of the present invention; Detailed Implementation

[0022] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] The terms "comprising" or "including" as used in this invention mean that the element preceding the term encompasses the element listed after the term, and do not exclude the possibility of encompassing other elements as well. Terms such as "inner," "outer," "upper," and "lower" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In this invention, unless otherwise explicitly specified and limited, the term "attached" and similar terms should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication of two elements or the interaction relationship between two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] Example An automatic localization method for the middle ear in multiple standard planes on temporal bone U-HRCT images, such as Figures 1-2As shown, the reinforcement learning framework comprises six parts: data preprocessing, bounded environment definition, SOP model warm start based on historical data, multi-agent policy network and action space selection module, reward module, and automatic termination module. Specifically, it includes: S1: Acquire temporal bone U-HRCT image data, divide it into training set and validation set; and preprocess the temporal bone U-HRCT image data to obtain the enhanced ossicular chain region of interest; The data preprocessing module includes three steps: segmentation of the ossicular chain in U-HRCT images, extraction of the ossicular chain ROI, and target saturation enhancement. First, the ossicular chain segmentation is performed on the input 3D temporal bone U-HRCT image data. A deep learning model is used for initial coarse segmentation, completing keypoint localization and region of interest (ROI) identification. Then, fine segmentation is performed based on the localization points and ROIs. Finally, the results are mapped back to the original data. After obtaining the ROI from the temporal bone U-HRCT, the ROI is input into the fine segmentation module for further segmentation. The network used in the fine segmentation module is the TransUnet model trained on 55 labeled data examples. This model plays a crucial role in the ossicular chain segmentation step, combining the attention mechanism of the Transform encoding part with the advantages of the U-net network's interrelationship between low-level and high-level information, demonstrating excellent performance in medical image segmentation. After segmentation, the segmentation results within the ROI and the complete data segmentation results are obtained.

[0026] Then, the region of interest of the ossicular chain is extracted based on the segmentation results.

[0027] Finally, to address the automatic termination issue in planar localization and improve localization accuracy, target saturation enhancement is performed on the ossicular chain. The original CT image corresponding to the target region of the ossicular chain segmentation template is saturated and enhanced; that is, based on the segmentation template, the corresponding voxels in the original CT image are highlighted. For example, in the voxel HU-normalized image, the voxels within the ossicular chain target region are set to the maximum value to enhance the edge contrast of the ossicular chain target. Specifically, the pixel value corresponding to the ossicular chain pixel position in the segmentation result is uniformly set to 3000, creating a sharp contrast with the pixel value in the actual image, thereby achieving the effect of target saturation enhancement and guiding the model to focus more on the target structure. Through this strategy, the model can more effectively capture ossicular chain-related information and improve the accuracy of automatic termination.

[0028] S2: Define the enhanced ossicular chain region of interest as the environment of the reinforcement learning framework, and describe the plane equations through standard formal parameterization; each agent obtains its own state from the environment through its own plane parameters; The ROI region of the ossicular chain in the 3D temporal bone U-HRCT data is defined as the environment of this reinforcement learning framework. A Cartesian coordinate system is constructed with the center point of the ossicular chain segmentation result as the origin. The plane equation is described using standard formal parameterization, specifically as follows: ; in, a , b , c Let be the unit normal vector of the plane. d The intercept of the normal vector is the distance from the origin to the plane. a 2 + b 2 + c 2 =1, and d >=0. The value of a planar node is set to a quadruple { a , b , c , d} represents the parameters of the SOP plane equation. All three agents operate in the same environment, each corresponding to its own plane equation. In each iteration, the image of the current plane is extracted from the temporal bone U-HRCT data, and the grayscale value range of the image is normalized to 0 to 1. The image is centered at the physical center point of the malleus segmentation result, and the image size is cropped to 224×224 voxels.

[0029] The composite plane of fused temporal information generated by the current environmental feedback is used to represent the state of each agent. Specifically, unlike the method of using a single current plane as the state, this application uses three temporally adjacent optimization planes ( P t-1 , P t , P t+1 The images are stitched together to form a three-channel image (3×H×W) as the state input for the agent. This design has a clear physical meaning: Channel 1 (Historical State): Records the planar information of the previous iteration step (t-1), preserving the initial starting point and evolution baseline of the structural optimization.

[0030] Channel 2 (Current State): Precisely presents the optimization result of the current iteration step (t), forming the most direct basis for decision-making.

[0031] Channel 3 (Buffer / Prediction Space): Reserved for the update area of ​​subsequent iterations (t+1), implicitly guiding the dynamic direction of the learning process.

[0032] S3: Calculate the average normal vector and average intercept of each standard observation surface based on the training set label data, and use them as the initial plane parameters of the agent; Specifically, the hot start module based on statistical data will pre-calculate the average normal vector (for orientation calibration) and average intercept (for spatial positioning) of the SOP based on historical clinical data during the first iteration (t=0) of the algorithm, thereby initializing the initial state plane of the agent.

[0033] S4: A multi-agent policy network is adopted to customize a differentiated action space for each standard observation surface and guide planar localization through a dual reward function; the multi-agent policy network is trained on the training set data and validated on the validation set; Different reinforcement learning network action spaces are defined based on the spatial location of different SOPs; Using a multi-agent reinforcement learning network enables simultaneous localization of multiple planes, improving efficiency. In multi-agent action decision-making, a dual reward function is integrated: individual reward, which guides adjustment toward the target plane to reduce the angle between the normal vectors of the current plane and the target plane and the distance between the intercept of the current plane and the intercept of the target plane; and cooperative reward, which constrains the relative positions between agents and makes full use of spatial information.

[0034] In the reward module, each action taken by the multi-agent unit earns a reward. The reward consists of two parts: one part is determined based on the spatial position between the agents on each plane, and the other part is determined based on the target position within each agent's plane.

[0035] Specifically, the reward function is crucial for the training of the agent. It is a key mechanism used to evaluate whether the result obtained by the agent after performing an action contributes to achieving the goal. Within each plane, we focus on the spatial distance error between the current plane and the target plane, expressed as: ; Between planes, focusing on the spatial relationships between them, it can be represented as: ; The total reward function during training is set as follows: ; in, sgn For symbolic functions, The function represents the distance between two planes. P 1 ( a 1, b 1, c 1, d 1) and plane P 2 ( a 2,b 2, c 2, d 2) The function represents the Euclidean norm of the difference between two planar vectors. , P i For the first i The plane equation of the agent's current plane at the next iteration; P t No. i The plane equation of the target plane at the next iteration; P i-1 For the first i- In the first iteration, the plane equation of the agent's current plane; P j i For the first i During the nth iteration, the 1st j The plane equation of the current plane for each agent; P j t For the first i During the nth iteration, the 1st j The plane equation of the target plane for each agent; the abs function represents the absolute value of the distance.

[0036] The reward function reflects whether the current plane is closer to the target plane and whether it better conforms to the spatial relationships between the planes compared to the previous iteration. If the current observed plane is closer to the target plane and better conforms to the spatial characteristics, the reward is positive; if it does neither, the reward is negative; otherwise, the reward is zero. This design effectively guides the agent's optimization direction, enabling it to gradually approach the target plane while also focusing on spatial relationships, thereby improving training efficiency and optimization accuracy. By learning the relationship between states and actions to train the policy network, and through continuous iteration, the agent can dynamically optimize the plane equations, thus achieving efficient approximation of the target plane.

[0037] Based on the spatial characteristics of the SOP (Standard Operating Position), a unique action space was designed for each agent by inputting their initial state St (t=0) into their respective policy networks. The policy network uses an improved VGG13 deep convolutional feature network as the planar localization agent, employing a 128×128 input size and undergoing five spatial downsampling operations to generate 512-channel 7×7 high-order semantic features. Subsequently, a 512-dimensional compact feature network is constructed through flattening and fully connected layers, then bifurcates into an advantage stream and a state value stream: the advantage stream calculates the relative advantage value of actions through a 128-dimensional hidden layer, while the state value stream evaluates the global state value. Finally, a centralized advantage function is used to fuse the features and generate a Q-value. This approach retains the advantages of VGG's hierarchical feature extraction while enhancing the stability of policy decisions by decoupling action selection and state evaluation. The policy network outputs the Q-value for each action, selecting the action with the largest Q-value as the predicted action. Based on the Q-value, the agent selects its optimal action parameters to obtain new planar parameters.

[0038] S5: The termination discriminant network is used to evaluate the planar localization results in real time, and the search process of the multi-agent policy network is terminated based on the evaluation results; the termination discriminant network is trained on the training set data and validated on the validation set. The automatic termination condition determination in reinforcement learning employs an EfficientNetv2 image classifier to determine whether the search process terminates. Its core advantage lies in evaluating the difference between the current plane and the target plane by extracting image features from the current slice, thus achieving automatic termination of the search process based on image features. This strategy not only adaptively adjusts the termination condition dynamically according to image features, improving the accuracy and intelligence of termination judgment, but also effectively overcomes the premature / late termination dilemma caused by relying on manual thresholds in traditional reinforcement learning. Figure 2 The diagram shows the training and termination process of EfficientNetv2. The following is its specific implementation: Classifier Construction and Training: First, multiple SOPs are treated as different categories, plus one category that is not any SOP, i.e., a multi-class network. The network is trained using three types of SOPs and other planes (collecting some planes near the three types of SOPs) as the initial training set. Then, it is applied to the automatic termination module, and the terminated planes are manually evaluated. When the evaluation accuracy is lower than 95%, the terminated planes are added to the training set, and the network is retrained. This process is repeated until the evaluation accuracy is not lower than 95%. At this point, EfficientNetv2 is considered to be capable of using the current plane image information and identifying the target plane.

[0039] Termination Decision Process: In each iteration of the deep reinforcement learning framework, after the agent obtains the optimal action and generates a new plane, the image information of the current plane is extracted and input into the trained EfficientNetV2 image classifier. If the classifier outputs that a certain plane belongs to the target plane, the iteration of that agent terminates, while other agents continue iterating until the states of all agents are found. These three planes are then used as the localization results. At this point, the reinforcement learning network automatically terminates, and the multi-class SOP auto-localization is considered successful.

[0040] S6: Output the final localization result.

[0041] Therefore, this invention employs the aforementioned automatic localization method for multiple standard planes of the middle ear in temporal bone U-HRCT images. It aims to leverage the extensive experience of radiologists in interpreting middle ear lesions to achieve end-to-end automatic localization of the standard observation planes of the ossicular chain, thereby improving localization efficiency. By preprocessing the input three-dimensional temporal bone U-HRCT image data and extracting ROIs and target saturation enhancement, the reinforcement learning network and automatic termination network can focus their attention on key structures. Based on the spatial location of different SOPs, this method customizes differentiated action spaces (translation, rotation) to adapt to the specific localization of each standard plane. The automatic termination network can promptly terminate the search of the reinforcement learning network, saving time. Compared with traditional heuristic algorithm localization methods, it is faster and has a higher success rate. Furthermore, the training strategies for the target saturation enhancement method and the automatic termination network can be extended to other similar SOP plane localization tasks.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatic localization of the middle ear in multiple standard planes on temporal bone U-HRCT images, characterized in that: Includes the following steps: S1: Acquire temporal bone U-HRCT image data, divide it into training set and validation set; and preprocess the temporal bone U-HRCT image data to obtain the enhanced ossicular chain region of interest; S2: Define the enhanced ossicular chain region of interest as the environment of the reinforcement learning framework, and describe the plane equations through standard formal parameterization; each agent obtains its own state from the environment through its own plane parameters; S3: Calculate the average normal vector and average intercept of each standard observation surface based on the training set label data, and use them as the initial plane parameters of the agent; S4: A multi-agent policy network is adopted to customize a differentiated action space for each standard observation surface and guide planar localization through a dual reward function; the multi-agent policy network is trained on the training set data and validated on the validation set; S5: The termination discriminant network is used to evaluate the planar localization results in real time, and the search process of the multi-agent policy network is terminated based on the evaluation results; the termination discriminant network is trained on the training set data and validated on the validation set. S6: Output the final localization result.

2. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 1, characterized in that: Preprocessing in S1 specifically includes: S11: A deep learning-based segmentation network is used to segment the ossicular chain in the temporal bone U-HRCT image data to obtain the segmentation results; S12: Extract the region of interest of the ossicular chain based on the segmentation results; S13: Target saturation enhancement is performed based on the region of interest of the ossicular chain to obtain the enhanced region of interest of the ossicular chain.

3. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 2, characterized in that: S11 specifically includes: The method using a deep learning model first performs coarse segmentation on the temporal bone U-HRCT image data to complete the localization of key points and regions of interest. Based on the localization points and regions of interest, fine segmentation is then performed, and the results are mapped back to the temporal bone U-HRCT image data to obtain the segmentation results within the regions of interest and the segmentation results of the complete data. The fine segmentation uses the TransUnet model.

4. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 1, characterized in that: S2 specifically includes: S21: Define the enhanced ossicular chain region of interest as the environment of the reinforcement learning framework, and construct a Cartesian coordinate system with the center point of the ossicular chain segmentation result as the origin. S22: The plane equation is described by standard formal parameterization, specifically as follows: ； in, a , b , c Let be the unit normal vector of the plane. d The intercept in the direction of the normal vector is... a 2 + b 2 + c 2 =1, and d >=0; the value of a planar node is set to a quadruple { a , b , c , d } represents the parameters of the standard observation plane equation; S23: Each agent obtains its own state from the environment through its own planar parameters. In each iteration, the image of the current plane is extracted from the temporal bone U-HRCT data, the gray value range of the image is normalized to 0 to 1, the image is centered on the physical center point of the malleus segmentation result, and the image size is cropped to 224×224 voxels. S24: Three temporally adjacent optimization planes are stitched together into a three-channel image as the state input of the agent. Channel 1 is the plane information of the previous iteration step, channel 2 is the optimization result of the current iteration step, and channel 3 is the update region of the subsequent iteration.

5. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 1, characterized in that: The dual reward function in S4 includes individual rewards and collaborative rewards; Individual rewards are used to guide the current plane to adjust towards the target plane, reducing the angle between the normal vectors and the intercept distance; collaborative rewards are used to constrain the relative positions between agents and optimize the positioning results by utilizing spatial positional relationships.

6. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 5, characterized in that: Individual rewards are represented as follows: ； Collaborative rewards are represented as follows: ； The double reward function is expressed as: ； in, sgn For symbolic functions, The function represents the distance between two planes. P 1 ( a 1, b 1, c 1, d 1) and plane P 2 ( a 2, b 2, c 2, d 2) The function represents the Euclidean norm of the difference between two planar vectors. , P i For the first i The plane equation of the agent's current plane at the next iteration; P t No. i The plane equation of the target plane at the next iteration; P i-1 For the first i- In the first iteration, the plane equation of the agent's current plane; P j i For the first i During the nth iteration, the 1st j The plane equation of the current plane for each agent; P j t For the first i During the nth iteration, the 1st j The plane equation of the target plane for each agent; the abs function represents the absolute value of the distance.

7. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 1, characterized in that: The policy network in S4 specifically includes: An improved VGG13 deep convolutional feature network is used as the planar localization agent. A 128×128 input size is adopted and five spatial downsampling operations are performed to generate 512-channel 7×7 high-order semantic features. A 512-dimensional compact feature is constructed through flattening and fully connected layers, which bifurcate into an advantage stream and a state value stream. The advantage stream calculates the relative advantage value of actions through a 128-dimensional hidden layer, while the state value stream evaluates the global state value. The Q-value is generated by fusing through a centralized advantage function. The policy network outputs the Q-value corresponding to each action. The action with the largest Q-value is selected as the predicted action. Based on the Q-value, the agent selects its optimal action parameters and obtains new planar parameters.

8. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 1, characterized in that: The termination discriminant network in S5 is the EfficientNetv2 image classifier.

9. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 8, characterized in that: The training process of the EfficientNetv2 image classifier specifically includes: Multiple standard observation surfaces are used as different categories, plus one category that is not any of the standard observation surfaces, forming a multi-class network. First, the network is trained using three standard observation surfaces and other planes as the initial training set. The trained planes are then manually evaluated. When the evaluation accuracy is below 95%, the terminated plane is added to the training set, and the network is retrained. This process is repeated until the evaluation accuracy is not lower than 95%, at which point the training ends, and the trained EfficientNetv2 image classifier is obtained.

10. The method for automatic localization of the middle ear in multiple standard planes in temporal bone U-HRCT images according to claim 9, characterized in that: In S5, the search process for a multi-agent reinforcement learning network is terminated based on the evaluation results. Specifically, this includes: In each iteration of the deep reinforcement learning framework, after the agent obtains the optimal action and generates a new plane, the image information of the current plane is extracted and input into the trained EfficientNetV2 image classifier. If the classifier outputs that a certain plane belongs to the target plane, the agent terminates the iteration for this round, and other agents continue to iterate until the states of all agents are found. The three planes are then used as the localization results, and the search process of the multi-agent policy network is terminated.

Images