Carotid artery OCT fiber cap segmentation method, device and equipment and storage medium

Abstract

The invention discloses a carotid artery OCT fiber cap segmentation method, device and equipment and a storage medium, and the method comprises the steps: carrying out the weighting of an original carotid artery OCT image, a user click feature graph and a segmentation probability graph, and obtaining a weighted input feature; performing feature extraction on the input features by using a KAN trunk feature extraction module to obtain deep semantic features, dynamically adjusting a segmentation probability graph according to newly added clicks of a user, obtaining a modulated target probability graph, and generating original interaction information; and performing attention feature fusion on the deep semantic features and the original interaction information through a double-branch post-fusion strategy to obtain a segmentation prediction result of the current round, determining whether to iterate the segmentation prediction result according to user feedback information, and outputting a final segmentation result. The method can improve the boundary definition and precision of carotid artery OCT image fiber cap segmentation, supports real-time user interactive iterative correction, and improves the speed and efficiency of carotid artery OCT fiber cap segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, apparatus, device, and storage medium for segmenting the carotid artery OCT fiber cap. Background Technology

[0002] Most ischemic strokes are caused by the rupture of vulnerable plaques in the carotid artery; fibrous cap thickness is the most critical indicator for assessing plaque vulnerability; and intravascular optical coherence tomography (OCT) is currently the only imaging technique that can accurately measure fibrous cap thickness in vivo.

[0003] However, existing clinical methods for assessing fibrous caps have significant limitations: 1) Manual measurement method: Experts need to manually locate and measure the thickness of the thinnest fiber cap in hundreds of OCT images. This process is extremely time-consuming and labor-intensive, and highly dependent on the operator's subjective experience. The results of different measurement operators or multiple measurements by the same measurement operator are inconsistent. 2) Semi-automatic segmentation algorithm: Existing research (such as dynamic programming-based methods) can assist in segmentation, but users still need to manually label the approximate range of the lipid core or fibrous cap in advance, which fails to completely free up manpower. Moreover, the algorithm is easily affected by image noise and artifacts, and its generalization ability is limited. 3) Fully automatic deep learning segmentation algorithm: When faced with common problems such as blood artifacts, rotation artifacts and blurred fibrous cap boundaries in carotid artery OCT images, the segmentation accuracy is difficult to meet clinical needs, and doctors cannot intervene or correct unsatisfactory results. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, device, and storage medium for segmenting carotid artery OCT fibrous caps, aiming to solve the technical problems of poor segmentation accuracy and lack of effective interaction in existing carotid artery OCT images.

[0005] In a first aspect, the present invention provides a method for segmenting the carotid artery OCT fibrous cap, the method comprising the following steps: The original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round are obtained. The original carotid artery OCT image, the user click feature map, and the segmentation probability map are weighted to obtain the weighted input features. The input features are extracted using the KAN backbone feature extraction module to obtain deep semantic features. The segmentation probability map is dynamically adjusted according to the user's new clicks to obtain the modulated target probability map. The original interaction information is generated based on the target probability map. The deep semantic features and the original interaction information are fused using a dual-branch fusion strategy to obtain the segmentation prediction result for the current round. Based on user feedback, it is determined whether to iterate the segmentation prediction result and output the final segmentation result.

[0006] Optionally, the step of acquiring the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round, and weighting the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain weighted input features, includes: Raw carotid artery optical coherence tomography (OCT) images are acquired via a medical imaging system interface. A user click feature map is obtained by using an interactive user interface to generate user click marks on the original carotid artery OCT image; Load the segmentation probability map predicted by the previous round of carotid artery OCT fibrous cap segmentation; The HICAM information preprocessing module adaptively assigns weights to the original carotid artery OCT image, the user-clicked feature map, and the segmentation probability map to obtain weighted input features.

[0007] Optionally, the step of adaptively weighting the original carotid artery OCT image, the user-clicked feature map, and the segmentation probability map using the HICAM information preprocessing module to obtain weighted input features includes: The first-level channel attention modulation module of the information preprocessing module HICAM obtains the global descriptor of each channel through global average pooling, and dynamically learns the activation weight of each channel using gate control. The original carotid artery OCT image, the user click feature map, and the segmentation probability map are processed in parallel using convolutional kernels of different sizes through HICAM's multi-scale feature extraction module to obtain contextual information at different scales. The HICAM's two-level channel attention modulation module fuses contextual information at different scales to generate an adaptive weight map. Based on the global descriptor and the activation weight, the confidence scores of the original carotid artery OCT image, the user click feature map, and the segmentation probability map are weighted and fused to obtain the weighted input features.

[0008] Optionally, the step of using the KAN backbone feature extraction module to extract features from the input features to obtain deep semantic features, dynamically adjusting the segmentation probability map based on user clicks to obtain a modulated target probability map, and generating original interaction information based on the target probability map includes: The KAN backbone feature extraction module is used to achieve flexible nonlinear mapping through multiple KAN linear transformation layers. Wavelet convolution spatial enhancement is introduced to decompose the input features at multiple scales to obtain deep semantic features. The modulation radius is adaptively calculated based on the minimum Euclidean distance between the current new click point in the new user clicks and the opposite category click point in the previous segmentation. A linear distance mask is generated based on the modulation radius, and the segmentation probability map is nonlinearly modulated by combining gamma correction and additive masking to obtain the modulated target probability map. Original interaction information is generated based on the target probability map, the original carotid artery OCT image, and the user click feature map.

[0009] Optionally, generating the original interaction information based on the target probability map, the original carotid artery OCT image, and the user click feature map includes: The target probability map, the original carotid artery OCT image, and the user click feature map are stitched together at the channel level to obtain a stitched image; The stitched image is input into a lightweight residual network, and the local details and global contextual semantics of the newly clicked area by the user are captured through convolutional layers to generate original interaction information that integrates user interaction intent and area confidence.

[0010] Optionally, the step of fusing the deep semantic features and the original interaction information using a dual-branch fusion strategy to obtain the segmentation prediction result for the current round, and deciding whether to iterate the segmentation prediction result based on user feedback information, and outputting the final segmentation result, includes: The importance weights of each channel in the deep semantic feature map corresponding to the deep semantic features and the interaction feature map corresponding to the original interaction information are calculated by using a channel-level weighting mechanism after the dual-branch fusion strategy. The deep semantic feature map and the interaction feature map are scaled channel by channel according to the importance weights to enhance the key texture and boundary features of the fibrous cap region in the deep semantic feature map and the interaction feature map, while suppressing the vascular background noise in the deep semantic feature map and the interaction feature map. The weighted deep semantic feature map and the interaction feature map are then fused element by element to obtain a fused feature map. The fused feature map is processed by the Sigmoid activation function to obtain the segmentation prediction result for the current round; Receive user feedback information, decide whether to iterate the segmentation prediction result based on the user feedback information, and output the final segmentation result.

[0011] Optionally, receiving user feedback and determining whether to iterate the segmentation prediction result based on the user feedback information, and outputting the final segmentation result, includes: Receive the segmentation prediction results from the user through an interactive interface. Receive user feedback information from users through an interactive interface regarding the segmentation prediction results; When the user feedback is satisfactory, no iteration is performed, and the segmentation prediction result is used as the final segmentation result and output. When the user feedback is unsatisfactory, the process is repeated according to a preset number of iterations until the user feedback is satisfactory.

[0012] Secondly, to achieve the above objectives, the present invention also proposes a carotid artery OCT fiber cap segmentation device, the carotid artery OCT fiber cap segmentation device comprising: The weighting module is used to acquire the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round, and to weight the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain the weighted input features; The feature extraction module is used to extract features from the input features using the KAN backbone feature extraction module to obtain deep semantic features, dynamically adjust the segmentation probability map according to the user's new clicks to obtain the modulated target probability map, and generate the original interaction information according to the target probability map. The fusion prediction module is used to fuse the deep semantic features and the original interaction information with attention features through a dual-branch fusion strategy to obtain the segmentation prediction result of the current round, and decide whether to iterate the segmentation prediction result based on user feedback information, and output the final segmentation result.

[0013] Thirdly, to achieve the above objectives, the present invention also proposes a carotid artery OCT fibrous cap segmentation device, the carotid artery OCT fibrous cap segmentation device comprising: a memory, a processor, and a carotid artery OCT fibrous cap segmentation program stored in the memory and executable on the processor, the carotid artery OCT fibrous cap segmentation program being configured to implement the steps of the carotid artery OCT fibrous cap segmentation method as described above.

[0014] Fourthly, to achieve the above objectives, the present invention also proposes a storage medium storing a carotid artery OCT fibrous cap segmentation program, wherein the carotid artery OCT fibrous cap segmentation program, when executed by a processor, implements the steps of the carotid artery OCT fibrous cap segmentation method as described above.

[0015] The carotid artery OCT fibrous cap segmentation method proposed in this invention obtains the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round. These three elements are then weighted to obtain weighted input features. A KAN backbone feature extraction module is used to extract deep semantic features from the input features. The segmentation probability map is dynamically adjusted based on new user clicks to obtain a modulated target probability map. Original interaction information is generated based on the target probability map. Finally, a dual-branch fusion strategy is used to integrate the deep semantic features and the original interaction information. Information is fused with attention features to obtain the segmentation prediction result for the current round. Based on user feedback, it is determined whether to iterate the segmentation prediction result and output the final segmentation result. This significantly improves the boundary clarity and pixel-level accuracy of fibrous cap segmentation in carotid OCT images. It also supports real-time interactive iterative correction, making the segmentation result highly compatible with clinical diagnostic needs. It effectively solves the problems of low segmentation accuracy, lack of interactivity, and poor efficiency of manual correction in traditional methods. It has high image segmentation accuracy, low interaction cost, strong targeting, good robustness, and good generalization ability, thus improving the speed and efficiency of carotid OCT fibrous cap segmentation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the first embodiment of the carotid artery OCT fiber cap segmentation method of the present invention; Figure 3 This is a flowchart illustrating the second embodiment of the carotid artery OCT fiber cap segmentation method of the present invention; Figure 4 This is a flowchart illustrating the third embodiment of the carotid artery OCT fiber cap segmentation method of the present invention; Figure 5 This is a flowchart illustrating the fourth embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention; Figure 6 This is a functional block diagram of the first embodiment of the carotid artery OCT fiber cap segmentation device of the present invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] The solution of this invention mainly involves: acquiring the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round; weighting the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain weighted input features; using the KAN backbone feature extraction module to extract features from the input features to obtain deep semantic features; dynamically adjusting the segmentation probability map based on new user clicks to obtain a modulated target probability map; generating original interaction information based on the target probability map; and fusing the deep semantic features and the original interaction information using a dual-branch fusion strategy to obtain the current round's segmentation probability map. The system segments the predicted results and, based on user feedback, determines whether to iterate the predicted results, outputting the final segmentation result. This significantly improves the boundary clarity and pixel-level accuracy of the segmentation of the fibrous cap in carotid OCT images. It also supports real-time interactive iterative correction, ensuring the segmentation results highly match clinical diagnostic needs. This effectively solves the problems of low segmentation accuracy, lack of interactivity, and poor efficiency of manual correction in traditional methods. The system offers high image segmentation accuracy, low interaction cost, strong targeting, good robustness, and excellent generalization ability, improving the speed and efficiency of carotid OCT fibrous cap segmentation and resolving the technical problems of poor segmentation accuracy and lack of effective interaction in existing carotid OCT image technology.

[0020] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0021] like Figure 1 As shown, the device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0023] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a network communication module, a user interface module, and a carotid OCT fiber cap segmentation program.

[0024] The device of the present invention calls the carotid artery OCT fibrous cap segmentation program stored in the memory 1005 through the processor 1001, and performs the operations described in the embodiment of the carotid artery OCT fibrous cap segmentation method below.

[0025] Based on the above hardware structure, an embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention is proposed.

[0026] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention.

[0027] In the first embodiment, the carotid artery OCT fibrous cap segmentation method includes the following steps: Step S10: Obtain the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round. Weight the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain the weighted input features.

[0028] It should be noted that the original carotid artery OCT image is a high-resolution cross-sectional image of the blood vessel acquired by an OCT scanning device, used to present the microscopic structural details of the carotid artery wall; the user click feature map is a heatmap generated by user interaction clicks on the image, encoding the location points of the fibrous cap region and the user's interaction intent; the segmentation probability map predicted in the previous round is a grayscale image output from the previous iteration, representing the segmentation confidence of the fibrous cap region with a pixel-level probability distribution; after obtaining the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round, the original carotid artery OCT image, the user click feature map, and the segmentation probability map can be weighted to obtain the weighted input features.

[0029] Step S20: Use the KAN backbone feature extraction module to extract features from the input features to obtain deep semantic features. Dynamically adjust the segmentation probability map according to the user's new clicks to obtain the modulated target probability map. Generate the original interaction information based on the target probability map.

[0030] It should be understood that the Kolmogorov-Arnold Network (KAN) backbone feature extraction module can be used to extract features from the input features to obtain deep semantic features, and dynamically adjust the segmentation probability map predicted in the previous round according to the user's new click to obtain the modulated target probability map. The original interaction information can be generated based on the modulated target probability map.

[0031] Step S30: The deep semantic features and the original interaction information are fused using a dual-branch fusion strategy to obtain the segmentation prediction result for the current round. Based on user feedback, it is decided whether to iterate the segmentation prediction result and output the final segmentation result.

[0032] It is understandable that the deep semantic features and the original interaction information can be fused by attention features through the dual-branch fusion strategy to obtain the segmentation prediction result of the current round. Furthermore, it is determined whether the segmentation prediction result needs to be iterated based on user feedback information, and then the final segmentation result is determined and output.

[0033] This embodiment, through the above-described scheme, acquires the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round. It then weights these three elements to obtain weighted input features. The KAN backbone feature extraction module is used to extract features from the input features, obtaining deep semantic features. The segmentation probability map is dynamically adjusted based on new user clicks to obtain a modulated target probability map. Original interaction information is generated based on the target probability map. Finally, a dual-branch fusion strategy is used to combine the deep semantic features and the original interaction information. Attention feature fusion is used to obtain the segmentation prediction result for the current round. Based on user feedback, it is determined whether to iterate the segmentation prediction result and output the final segmentation result. This significantly improves the boundary clarity and pixel-level accuracy of fibrous cap segmentation in carotid OCT images. It also supports real-time interactive iterative correction, making the segmentation result highly matched with clinical diagnostic needs. It effectively solves the problems of low segmentation accuracy, lack of interactivity and poor efficiency of manual correction in traditional methods. It has high image segmentation accuracy, low interaction cost, strong targeting, good robustness and good generalization ability, and improves the speed and efficiency of carotid OCT fibrous cap segmentation.

[0034] Furthermore, Figure 3 This is a schematic flowchart of the second embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention, as shown below. Figure 3 As shown, based on the first embodiment, a second embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention is proposed. In this embodiment, step S10 specifically includes the following steps: Step S11: Obtain the original carotid artery optical coherence tomography (OCT) image through the medical imaging system interface.

[0035] It should be noted that the raw carotid artery OCT image is obtained through the medical imaging system interface (generally, high-resolution cross-sectional images of blood vessels can be acquired by OCT scanning equipment, which include microstructural details of the carotid artery wall).

[0036] Understandably, the original carotid artery OCT image is the raw medical image data to be segmented and remains unchanged throughout the iteration process.

[0037] Step S12: Obtain the user click feature map generated by the user marking clicks on the original carotid artery OCT image through the interactive user interface.

[0038] It should be understood that user click feature maps (i.e., user-specified fiber cap region locations and interaction intentions encoded in the form of heatmaps) can be generated by capturing user click points marked on images through interactive user interfaces.

[0039] In the specific implementation, the user clicks on the feature map, that is, the user evaluates the segmentation results of the previous round (round t-1). If there are oversegmented regions, a negative click is added; if there are undersegmented regions, a positive click is added. These newly added click information, together with the clicks of all previous rounds, constitute the click set of round t. This set is converted into a two-dimensional or multi-channel feature map through spatial location encoding (such as Gaussian distance transform), which serves as the guiding signal for the network.

[0040] Step S13: Load the segmentation probability map predicted by the previous round of carotid artery OCT fibrous cap segmentation.

[0041] It is understandable that the segmentation probability map of the carotid artery OCT fibrous cap segmentation prediction output by the model in the previous iteration is loaded (that is, the pixel-level probability distribution of the fibrous cap region in grayscale image form, which is used to reflect the segmentation confidence of the previous round).

[0042] It should be understood that the segmentation probability map of the previous round of prediction This is a two-dimensional tensor output by the network at the end of the (t-1)th iteration, representing the probability value of each pixel belonging to the fiber cap region. It contains the model's latest knowledge of the target region.

[0043] Understandably, the modulated probability map of the previous round... This input is obtained through... This invention utilizes the improved probabilistic graph modulation mechanism proposed in this paper; the core function of this mechanism is to leverage the newly added click information in the t-th round. This is used to "correct" and "guide" the probability graph from the previous round to better align with the user's latest intent. Specifically, for new positive clicks, the low probability values ​​near them are amplified (closer to 1); for new negative clicks, the high probability values ​​near them are suppressed (closer to 0). This modulation is adaptive, and the modulation range (radius R) is dynamically determined based on the distance between the current click and the opposite type of click to avoid over-modulation.

[0044] These four inputs are fed into the main branch and side branches of the framework for processing, and finally, a segmentation prediction result for the t-th round is generated through a dual-branch fusion strategy. The system will then prompt the user whether to proceed to the next iteration. If the user is satisfied with the current result, the iteration will stop and the final segmentation mask will be output. If there is still an error, the system will proceed to the (t+1)th iteration and repeat the above process until a satisfactory segmentation result is obtained.

[0045] In practical implementation, multimodal input representation can convert input data from different sources and with different properties (raw pixels, user interactions, historical predictions) into a standard tensor format that neural networks can process. 1. Input Acquisition X image : Original carotid artery OCT image, containing original texture and structural information, with dimensions [3, H, W].

[0046] C: The set of all clicks (coordinates, labels) provided by the user in the previous iteration.

[0047] X prev : The segmentation probability map predicted by the network in the previous round, with dimensions [1, H, W].

[0048] 2. Click on the code Objective: To transform discrete click locations into dense feature maps containing spatial information, enabling the network to perceive the location indicated by the user.

[0049] Method: Generate a two-dimensional Gaussian heatmap for each click and accumulate it to the corresponding channel (positive click channel or negative click channel).

[0050] Formula: For the click set Every click in ,in As coordinates, For each label (0 negative, 1 positive), calculate the click feature map. :

[0051] in, Representative channel, For positive click channel, This is a negative click channel.

[0052] It is an indicator function; when the i-th clicked label equals the current channel... The value is 1 if it is active, and 0 otherwise.

[0053] It is the Gaussian kernel standard deviation that controls the range of influence of a click; it is a hyperparameter.

[0054] The final result It is a 2-channel feature map [2, H, W].

[0055] 3. Input concatenation: Objective: To initially integrate information from different modalities along the channel dimension to form the direct input of the backbone network.

[0056] Method: Direct concatenation.

[0057]

[0058] in, Features are input to the backbone network. Indicates splicing, Original carotid artery OCT image, For users to click on the feature map, This is the segmentation probability map from the previous prediction.

[0059] spliced ​​tensors The dimensions are [3+2+1, H, W] = [6, H, W].

[0060] Step S14: The original carotid artery OCT image, the user click feature map, and the segmentation probability map are adaptively weighted by the information preprocessing module HICAM to obtain the weighted input features.

[0061] It should be understood that the information preprocessing module, namely the Hierarchical Information Channel Attention Module (HICAM), can process the original carotid artery OCT image, the user click feature map, and the segmentation probability. Figure 3 These components together constitute multimodal input data, which are then adaptively weighted to obtain weighted input features, thereby ensuring that subsequent iterative optimization can accurately integrate user interaction feedback and image semantic information.

[0062] Furthermore, step S14 specifically includes the following steps: The first-level channel attention modulation module of the information preprocessing module HICAM obtains the global descriptor of each channel through global average pooling, and dynamically learns the activation weight of each channel using gate control. The original carotid artery OCT image, the user click feature map, and the segmentation probability map are processed in parallel using convolutional kernels of different sizes through HICAM's multi-scale feature extraction module to obtain contextual information at different scales. The HICAM's two-level channel attention modulation module fuses contextual information at different scales to generate an adaptive weight map. Based on the global descriptor and the activation weight, the confidence scores of the original carotid artery OCT image, the user click feature map, and the segmentation probability map are weighted and fused to obtain the weighted input features.

[0063] It should be noted that the information preprocessing module HICAM uses its three-level structure (first-level channel attention modulation, multi-scale feature extraction, and second-level channel attention modulation) to adaptively assign weights to the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round: First, the first-level channel attention modulation obtains the global descriptors of each channel through global average pooling and dynamically learns the channel activation weights using a gating mechanism; second, the multi-scale feature extraction module uses multi-size convolutional kernels in parallel, such as 1×1, 3×3, and 5×5 convolutional kernels, and max pooling to process the input features, capturing local and global semantics at different scales; finally, the second-level channel attention modulation fuses the multi-scale features and generates an adaptive weight map, which weights and fuses the structural information of the original OCT image, the interactive intent of the user click feature map, and the confidence distribution of the segmentation probability map, outputting optimized input features. This provides the KAN backbone network with feature inputs with high semantic expressive power, ensuring the accuracy and robustness of subsequent segmentation iterations.

[0064] In the specific implementation, the information preprocessing module HICAM can solve the problem of information weight imbalance caused by simple splicing. This module uses a channel attention mechanism to enable the network to adaptively learn the importance of different input channels, so as to highlight the importance of different inputs and extract multi-scale spatial features for subsequent calculations.

[0065] in: 1. Primary channel attention modulation: Objective: To recalibrate the importance of each channel of the stitched 6-dimensional input features.

[0066] Formula (Squeeze-and-Excitation): Squeeze (Global Information Embedding): Obtains the global descriptor for each channel through Global Average Pooling (GAP).

[0067]

[0068] in, is the global average pooling output (i.e., the global descriptor) of channel c, representing the average activation intensity (scalar value) at all spatial locations of this channel, used for subsequent channel attention weight calculation; H is the height of the input feature map (number of pixel rows), representing the vertical size of the image; W is the width of the input feature map (number of pixel columns), representing the horizontal size of the image; i is the index in the height direction (from 1 to H), traversing each row of the feature map; j is the index in the width direction (from 1 to W), traversing each column of the feature map; For the input feature map at channel c, position The pixel value, i.e., the spatial coordinates of that channel. The characteristic response at that location.

[0069] Excitation (Adaptive Recalibration): Learns the activation weights of each channel through a simple gating mechanism.

[0070]

[0071]

[0072] in, For adaptive weights, It is the vector after Squeeze.

[0073] and These are the weights of the two fully connected layers, and r is the reduction ratio.

[0074] It is the ReLU activation function. It is the Sigmoid activation function, which maps weight values ​​to the range [0, 1].

[0075] This indicates channel-by-channel multiplication (broadcast mechanism).

[0076] 2. Multi-scale feature extraction (Inception Module): Objective: To extract features in parallel using convolutional kernels of different sizes to capture contextual information at multiple scales that may exist at the fiber cap boundary.

[0077]

[0078] in, The output feature map (the concatenated multi-scale features) contains local and global semantic information at different scales, which is used for subsequent second-level channel attention modulation. For feature splicing operations, It is a 1×1 convolution used for channel compression and feature transformation. A 3×3 convolution is used to capture local texture and edge features. A 5×5 convolution is used to capture a wider range of contextual information. 3×3 max pooling is used to reduce spatial resolution to enhance translation invariance and suppress noise. This is the input feature map.

[0079] The number of channels in the spliced ​​feature map Y becomes:

[0080] in, This is the fused multi-scale feature map (output feature). The feature map output by a 1×1 convolution. The feature map output by the 3×3 convolution. The feature map output by the 5×5 convolution. This is the feature map output by 3×3 max pooling.

[0081] 3. Secondary channel attention modulation: Objective: To recalibrate the channel importance of the multi-scale fused features output by the Inception module to further optimize the feature quality of the input backbone network.

[0082] Formula: It is exactly the same as the first-level modulation process, but it applies to feature Y.

[0083]

[0084] Output features It will be fed into the main encoder.

[0085] This embodiment acquires the original carotid artery optical coherence tomography (OCT) image through a medical imaging system interface; obtains a user-generated click feature map by user-marked clicks on the original carotid artery OCT image through an interactive user interface; loads the segmentation probability map from the previous round of carotid artery OCT fibrous cap segmentation prediction; and uses the information preprocessing module HICAM to adaptively assign weights to the original carotid artery OCT image, the user-generated click feature map, and the segmentation probability map to obtain weighted input features. This significantly improves the initial recognition accuracy of the segmentation model for the fibrous cap region and the efficiency of user interaction response, effectively solving the segmentation deviation problem caused by insufficient feature fusion in traditional methods.

[0086] Furthermore, Figure 4 This is a schematic flowchart of the third embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention, as shown below. Figure 4 As shown, based on the first embodiment, a third embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention is proposed. In this embodiment, step S20 specifically includes the following steps: Step S21: Utilize the KAN backbone feature extraction module to achieve flexible nonlinear mapping through multiple KAN linear transformation layers, and introduce wavelet convolution spatial enhancement to perform multi-scale decomposition of the input features to obtain deep semantic features.

[0087] It should be noted that the KAN backbone feature extraction module implements a flexible nonlinear mapping based on the Kolmogorov-Arnold representation theorem through multiple KAN linear transformation layers. At the same time, wavelet convolution spatial enhancement (using Daubechies wavelet basis for discrete wavelet transform) is introduced to decompose the input feature map into sub-band features of high-frequency details (such as edges and textures) and low-frequency structures (such as overall contours). This captures cross-scale semantic associations while maintaining spatial resolution, and finally outputs deep semantic features with strong discriminative and context-aware capabilities, providing high-precision feature input for subsequent dual-branch fusion strategies.

[0088] In the specific implementation, in order to improve the backbone network's ability to model the fibrous cap structure in carotid OCT images at multiple scales, especially to maintain the integrity of the fine spatial structure under challenging conditions such as high noise, low contrast and blurred boundaries, this embodiment uses the Kolmogorov-Arnold Network (KAN) module to replace the traditional MLP or convolutional feedforward block. It utilizes its learnable spline basis functions and innovatively introduces wavelet transform convolution (WTConv). By extracting features in the wavelet domain, it achieves explicit modeling of multi-scale spatial frequency information and realizes efficient capture of texture, edge and local frequency domain information.

[0089] 1. KAN Linear Transformation Layer (KANLinear) Objective: To enhance the model's ability to fit complex nonlinear mappings by using a learnable activation function based on piecewise polynomial splines, outperforming traditional linear layers with fixed activation functions.

[0090]

[0091] in, The output function of the KAN linear transform layer represents the input features. The mapping result after KAN transformation (i.e., the component of deep semantic features) is used to capture the nonlinear relationships of the input data.

[0092] The number of spline basis functions (a hyperparameter, typically 10~20) represents the total number of spline functions used in parallel in the KAN layer.

[0093] is the learnable weight coefficient (scalar) of the k-th spline function, which is optimized through backpropagation and used to dynamically adjust the contribution strength of each spline function to the final output. The spline order is 3 by default (cubic spline), the grid size is 5, and the coverage interval is [-1,1]. In this module, three serial KANLinear layers are used to complete the dimensionality increase and decrease of the channel dimension.

[0094] For the k-th spline function, take the input... As the independent variable, Use the parameters to generate a nonlinear response.

[0095] The parameters of the k-th spline (composed of grid points and coefficients) are dynamically adjusted through training to enable KAN to flexibly fit complex functions (such as the fibrous cap boundary features of carotid artery OCT images).

[0096] 2. Wavelet Convolution Spatial Enhancement (WTConv + BN + GELU) Objective: Traditional convolution operates directly in the spatial domain, making it difficult to effectively separate structural information from noise in OCT images. This invention proposes wavelet convolution (WTConv2d), whose core is feature extraction in the wavelet transform domain. Specifically, the input feature map is subjected to discrete wavelet transform, decomposing it into sub-bands of different scales and directions, and then fused using learnable weights to achieve accurate capture of high-frequency information such as edges and textures.

[0097] First, perform a first-order two-dimensional discrete wavelet transform (2D-DWT) on it, using the Daubechies wavelet basis (db1).

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] in: Low-frequency subband (approximate component) Horizontal high-frequency sub-band (vertical edge) Vertical high-frequency sub-band (horizontal edge) Diagonal high-frequency subband (texture and noise) and These are the low-pass and high-pass decomposition filters for the db1 wavelet, respectively. This indicates a product operation, and ↓2 indicates that the downsampling factor is 2. The wavelet convolution module is applied once between every two KANLinear layers, forming an alternating structure of "nonlinear transformation - wavelet feature extraction".

[0105] Finally, the main branch outputs high-level semantic features F. backbone This feature combines the strong nonlinear modeling capability of KAN with the explicit mathematical modeling of multi-scale edges and textures by wavelet convolution, making it particularly suitable for segmentation tasks of thin-layer, high-noise, and blurred-boundary targets such as fiber caps in OCT images.

[0106] Step S22: Adaptively calculate the modulation radius based on the minimum Euclidean distance between the current newly added click point in the user's new click and the click point of the opposite category in the previous segmentation.

[0107] Understandably, by dynamically integrating user interaction feedback to achieve accurate updates to the segmentation probability map, the modulation radius can be adaptively calculated based on the minimum Euclidean distance between the newly added click point and the opposite category click point in the previous segmentation.

[0108] Step S23: Generate a linear distance mask based on the modulation radius, and perform nonlinear modulation on the segmentation probability map by combining gamma correction and additive mask to obtain the modulated target probability map.

[0109] It should be understood that a linear distance mask can be generated based on the modulation radius, thereby quantifying the influence intensity of the click point on neighboring pixels. The segmentation probability map is then nonlinearly refined by combining gamma correction and additive masking, thereby outputting a modulated target probability map that incorporates the user's intent, ensuring that the segmentation boundary of the fiber cap region is highly consistent with the user interaction.

[0110] Step S24: Generate original interaction information based on the target probability map, the original carotid artery OCT image, and the user click feature map.

[0111] It should be noted that by processing the modulated segmentation probability map with the original image and the user click map, the original interaction information containing user interaction feedback can be extracted. This interaction information and the deep semantic features of the main branch are fused with attention features through a dual-branch fusion strategy, thereby directly integrating the dynamic adjustment result of the user's new clicks (i.e., the modulated probability map) into the current segmentation prediction, ensuring the continuity and accuracy of the iterative process, and avoiding logical breaks between steps.

[0112] Furthermore, step S24 includes the following steps: The target probability map, the original carotid artery OCT image, and the user click feature map are stitched together at the channel level to obtain a stitched image; The stitched image is input into a lightweight residual network, and the local details and global contextual semantics of the newly clicked area by the user are captured through convolutional layers to generate original interaction information that integrates user interaction intent and area confidence.

[0113] Understandably, the modulated segmentation probability map serves as the input component of the side branch (a lightweight residual network, i.e., a lightweight ResNet network). The side branch concatenates the original carotid artery OCT image, the user click feature map, and the modulated target probability map at the channel level, and inputs them into the lightweight ResNet network for multi-scale feature extraction. This network captures local details (such as edge gradients and neighborhood textures of the click point) and global contextual semantics (such as the continuity of the fibrous cap structure) of the newly clicked region through convolutional layers, generating original interaction information that integrates user interaction intent and region confidence. This provides high-precision interaction-aware features for the subsequent dual-branch fusion strategy, ensuring the dynamic consistency between the segmentation results and user feedback.

[0114] In its implementation, the purpose of Probability Map Modulation (PMR) is to fully utilize historical predictions. Information in the middle, and based on the latest user clicks By making localized and precise adjustments, strong prior user information is injected into the network, thereby accelerating convergence and improving accuracy.

[0115] 1. Adaptive modulation radius calculation: Objective: To dynamically determine the modulation range based on the distance between the current click and the opposite category click, so as to avoid over-modulation or under-modulation.

[0116] formula: For a positive click (label l(u) = 1):

[0117] For negative clicks (label l(u) = 0):

[0118] in, For the modulation radius, This is the upper limit of the modulation radius, i.e., the preset maximum radius hyperparameter. This is the lower limit of the modulation radius, i.e., the preset minimum radius hyperparameter. For the current click point, For the opposite category of clicks, Let N be the Euclidean distance, and let P be the set of coordinates of all negative and positive clicks in the current batch.

[0119] 2. Linear distance mask generation: Objective: To generate a weight mask that is inversely proportional to the distance from the click center, such that the modulation intensity is greater the closer the position is to the click center.

[0120] formula: For each position x in the image:

[0121] in, The linear distance mask value. Image position coordinates, Provide the coordinates of the user's click point. The Euclidean distance between the two coordinates is given. For the modulation radius, This is a hyperparameter for controlling the overall modulation intensity.

[0122] 3. Refined probability modulation: Objective: To improve the segmentation probability map of the previous prediction by combining gamma correction (global adjustment) and additive masking (local fine-tuning). Perform nonlinear transformation.

[0123] formula: For each position x within the modulation window, the probability value :

[0124] in, This is the modulated target probability map. This is the probability map of the previous round of segmentation. For adaptive gamma coefficients, The linear distance mask value. User interaction tags, These are the image position coordinates.

[0125] in, This is the gamma value modulated by the gamma. This formula ensures that: Positive click: Simultaneously via exponentiation Increase the probability value and use a positive mask +M(x) to further increase the probability value.

[0126] Negative click: Simultaneously through exponentiation The probability value is reduced further by reducing the probability value and using a negative mask -M(x).

[0127] This embodiment utilizes the above-described scheme, employing a KAN backbone feature extraction module to achieve flexible nonlinear mapping through multiple KAN linear transformation layers. Wavelet convolution spatial enhancement is introduced to perform multi-scale decomposition of the input features, obtaining deep semantic features. The modulation radius is adaptively calculated based on the minimum Euclidean distance between the currently added click point in the user's new click and the opposite category click point in the previous segmentation. A linear distance mask is generated based on the modulation radius, and combined with gamma correction and additive masking, nonlinear modulation is applied to the segmentation probability map to obtain a modulated target probability map. Original interaction information is generated based on the target probability map, the original carotid OCT image, and the user click feature map. This significantly improves the accuracy of the fibrous cap segmentation boundary and the efficiency of the interaction response, effectively solving the technical bottlenecks of large segmentation deviations and long user correction times in traditional methods, and improving the speed and efficiency of carotid OCT fibrous cap segmentation.

[0128] Furthermore, Figure 5 This is a schematic flowchart of the fourth embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention, as shown below. Figure 5 As shown, based on the first embodiment, a fourth embodiment of the carotid artery OCT fibrous cap segmentation method of the present invention is proposed. In this embodiment, step S30 specifically includes the following steps: Step S31: Using a dual-branch fusion strategy, a channel-level weighting mechanism is employed to calculate the importance weights of each channel in the deep semantic feature map corresponding to the deep semantic features and the interaction feature map corresponding to the original interaction information.

[0129] It should be noted that the attention feature fusion module using the dual-branch fusion strategy can employ a channel-level weighting mechanism to calculate the importance weight of each channel in the feature map. That is, the dual-branch fusion strategy uses a channel-level weighting mechanism to calculate the importance weight of each channel in the deep semantic feature map corresponding to the deep semantic features and the interaction feature map corresponding to the original interaction information.

[0130] Understandably, the deep semantic feature map output by the main branch is generated by the KAN main feature extraction module and contains high-level semantic information such as the global structure of the carotid OCT image and the boundary of the fibrous cap (e.g., the microscopic texture of the vessel wall and the distribution of plaques). The original interaction information extracted by the side branches is an interaction feature map generated by processing the original image, user click map and modulated probability map into a light-weight residual network and then splicing them together. It reflects the user's intention to add a click (e.g., the localization correction of the fibrous cap region).

[0131] Step S32: Scale the deep semantic feature map and the interaction feature map channel by channel according to the importance weights to enhance the key texture and boundary features of the fibrous cap region in the deep semantic feature map and the interaction feature map, while suppressing the vascular background noise in the deep semantic feature map and the interaction feature map. Perform element-level fusion on the weighted deep semantic feature map and the interaction feature map to obtain a fused feature map.

[0132] It should be noted that after obtaining the importance weights, the attention feature fusion module applies the weights to the deep semantic feature map (the global semantic features output by the KAN backbone network) and the original interaction information feature map (the user interaction features processed by the side branch lightweight residual network). The feature map is scaled channel by channel through a channel-level weighting mechanism (e.g., the weight vector is multiplied element by element with the corresponding channel feature), which significantly enhances the key texture and boundary features of the fibrous cap region while effectively suppressing the background noise of blood vessels. Subsequently, the two weighted feature maps are fused element by element to generate a fused feature map.

[0133] In the specific implementation, the purpose of dual-branch encoding and attention fusion is to prevent valuable user interaction information (clicks and modulated probability maps) from being diluted during the downsampling process of the backbone network; high-resolution original interaction information is preserved through side branches and adaptively fused with deep semantic features extracted from the backbone branches.

[0134] 1. Side branch input concatenation:

[0135] in, For side branch input features, Indicates splicing, Original carotid artery OCT image, For users to click on the feature map, This is the segmentation probability map from the previous prediction.

[0136] 2. Feature extraction: Main branch: Processes the HICAM-weighted feature Z, extracting high-level semantic features through convolution and KAN modules. .

[0137] Side branches: processing Features that retain a large amount of original spatial information are extracted using a lightweight ResNet network. .

[0138] Attention Feature Fusion (AFF): Objective: To adaptively determine, at each spatial location, whether to trust the deep features of the main branches or the detailed features of the side branches more.

[0139]

[0140] in, This is the global attention weight vector. This is the weight matrix for the second-level linear transformation. For activation function, This is the weight matrix for the first-level linear transformation. For global average pooling, The main feature map This is a feature map of the side branches.

[0141] Calculate the fusion weight α:

[0142]

[0143] in, This is the local attention weight vector. This is the weight matrix for the second-level linear transformation. For activation function, This is the weight matrix for the first-level linear transformation. For batch normalization, The main feature map For side branch feature maps, To integrate weights, It is the sigmoid activation function.

[0144] Fusion: The learned weight map α is used to perform a weighted summation of the features of the two branches.

[0145]

[0146] in, To fuse feature maps, To integrate weights, This is element-wise multiplication. The main feature map This is a feature map of the side branches.

[0147] Step S33: Process the fused feature map using the Sigmoid activation function to obtain the segmentation prediction result for the current round.

[0148] It should be understood that the fused feature map is processed by the Sigmoid activation function to output the segmentation probability map of the current round, which serves as an accurate segmentation prediction result, ensuring that the fiber cap boundary is highly consistent with the user's interaction intent, while providing optimized input for the next iteration.

[0149] Understandably, the fused feature map is used by the Sigmoid activation function to generate the segmentation probability map for the current round, outputting a high-precision segmentation prediction result. This result not only retains the semantic expression ability of the KAN backbone network, but also accurately incorporates user interaction feedback, providing optimized input for the next iteration and significantly improving the boundary clarity and robustness of fiber cap segmentation.

[0150] In the specific implementation, the fused high-level features Upsampling to the original image resolution and outputting a pixel-level classification probability map, the resolution can be gradually restored using a decoder (typically composed of transposed convolutions or interpolated upsampling layers and regular convolutional layers); and multi-scale features from corresponding layers in the encoder are fused through skip connections; the final output is a segmentation probability map for the current round obtained through a 1x1 convolutional layer and a sigmoid activation function. .

[0151] The training process is as follows: Simulated click sampling: Simulated clicks are generated using a fusion strategy.

[0152] First, generate random clicks: randomly sample from within the target mask. One positive click, ensuring the distance between clicks > And the distance from the boundary is greater than Random sampling from the background area A negative click ensures that its distance from the target boundary is within [a certain range]. Within the range.

[0153] Then, based on the current forecast Compared to real masks To identify the differences, an iterative sampling strategy is used to generate new clicks. The misclassified region E is found, and morphological erosion is applied to it. Within the largest connected component, the point furthest from the boundary is selected as the new click.

[0154] With probability Reset the click sequence to prevent overfitting.

[0155] Loss function and optimization: The network parameters are optimized round by round using the SoftIoU loss function.

[0156]

[0157] in, for Loss function value, To predict the segmentation probability map, For true segmentation labels, The pixel weight matrix This is the Hadamard product (element-wise multiplication). For the predicted and actual maximum values.

[0158] The Adam optimizer was selected, and the MultiStepLR learning rate scheduler was used. The initial learning rate was set to... , ( = (0.9, 0.999), .

[0159] The training batch size (BatchSize) can be set to 4, and an early stopping mechanism can be enabled when the performance on the validation set no longer improves.

[0160] In the specific implementation, the core of the framework of this embodiment is a backbone segmentation network with a U-shaped network and KAN modules as key components; the encoder is composed of alternating layers of convolutional modules and KAN modules to extract multi-level features; the KAN module replaces the traditional fixed activation function with a learnable activation function, thereby enhancing the expressive power and fitting accuracy of the model.

[0161] The framework comprises four key parts: Information Preprocessing Module (HICAM): This module first performs preliminary weighting on the concatenated multimodal input through a channel attention mechanism to highlight the features of important channels; then it extracts multi-scale spatial features through a parallel Inception module (using 1x1, 3x3, and 5x5 convolutional kernels); finally, it performs final calibration through channel attention again to provide the network with high-quality weighted input information.

[0162] KAN Backbone Feature Extraction Module: As the core computational unit of the backbone branch, this invention uses the Kolmogorov-Arnold Network (KAN) module to replace the traditional feedforward network structure to enhance the nonlinear modeling ability of complex tissue structures in OCT images. This module consists of three levels of KANLinear layers alternating with wavelet convolution (WTConv). The KANLinear layers realize dynamic nonlinear transformation of the channel dimension through learnable piecewise spline functions. The wavelet convolution introduces multi-scale frequency domain analysis capability in the spatial dimension. This design enables the network to explicitly distinguish between fiber cap edges (high frequency) and speckle noise in the wavelet domain, achieving synergistic optimization of edge enhancement and noise suppression.

[0163] Probabilistic graph modulation mechanism: This mechanism adaptively adjusts the probability values ​​of pixels near the click in the previous probabilistic graph based on the user's new click. Its innovations are: a) dynamically calculating the modulation radius based on the minimum Euclidean distance between the current click and the opposite type of click to avoid over-modulation; b) introducing a distance-based linear mask to further strengthen or suppress the probability value of the click center area, making the modulation effect more accurate and robust.

[0164] Dual-branch fusion strategy: Two parallel branches are set up. The main branch processes the HICAM-weighted input and extracts deep semantic features; the side branch consists of a lightweight ResNet, which is dedicated to processing the input concatenated from the original image, click map, and modulated probability map, aiming to preserve rich original interaction information; finally, the features of the two branches are adaptively fused using an attention feature fusion module (AFF) to make full use of deep semantics and shallow details and prevent key interaction information from being diluted in the deep network.

[0165] The network training employs a simulated click sampling strategy, combining random sampling and iterative sampling to approximate real user interaction behavior; the loss function uses SoftIoU loss; finally, after multiple iterations, users can obtain accurate segmentation results by adding a small number of clicks.

[0166] Step S34: Receive user feedback information, decide whether to iterate the segmentation prediction result based on the user feedback information, and output the final segmentation result.

[0167] It should be understood that after receiving feedback from the user on the segmentation prediction results, the system can decide whether to terminate the iteration based on the feedback. If the feedback is satisfied, the segmentation prediction result of the current round is output as the high-precision final segmentation result; otherwise, the next round of interactive optimization is started.

[0168] Furthermore, step S34 specifically includes the following steps: Receive the segmentation prediction results from the user through an interactive interface. Receive user feedback information from users through an interactive interface regarding the segmentation prediction results; When the user feedback is satisfactory, no iteration is performed, and the segmentation prediction result is used as the final segmentation result and output. When the user feedback is unsatisfactory, the process is repeated according to a preset number of iterations until the user feedback is satisfactory.

[0169] It should be noted that the system receives user feedback on the segmentation prediction results in real time through an interactive interface: when the user feedback is "satisfied", the current segmentation prediction result is directly output as the final segmentation result; when the user feedback is "unsatisfied", the system automatically triggers the next round of iterative optimization process (based on a preset upper limit of the number of iterations), continuously generates new prediction results and waits for user feedback until the user confirms satisfaction, ensuring that the output results fully match the clinical diagnostic needs.

[0170] In the specific implementation, users can provide real-time feedback on the segmentation prediction results of the current round through an interactive interface (such as clicking the confirmation button, marking the correction area, or adjusting the segmentation boundary). They can also make dynamic decisions based on preset satisfaction thresholds (e.g., the user confirms "the result is accurate" or the correction error is lower than the set threshold): if the feedback meets the conditions (e.g., the user marks it as "satisfactory" or the boundary error is ≤5%), the iteration process is terminated, and the final segmentation result (i.e., a high-precision fiber cap region segmentation probability map, containing clear boundaries and pixel-level confidence) is output; otherwise, the system uses the user's newly added interactive information (such as corrected click points) as input, re-triggers the Information Preprocessing Module (HICAM), KAN backbone feature extraction, and dual-branch fusion strategy to generate an optimized segmentation prediction, realizing multiple rounds of interactive iteration until the user confirms that the result meets the clinical diagnostic accuracy requirements, ensuring that the final output result is highly consistent with the doctor's actual needs.

[0171] It is understood that, compared with the prior art, this embodiment has the following significant advantages: High precision and low interaction cost: Through innovative KAN backbone network, HICAM weighting, probabilistic graph modulation and dual-branch fusion strategy, this invention can make full use of the information provided by each user interaction, and achieve segmentation accuracy far exceeding that of existing interactive methods with very few clicks (e.g., only 3.56 clicks on average to achieve 80% IoU), greatly reducing the operational burden on doctors.

[0172] Highly targeted and robust: This invention is the first interactive algorithm specifically designed to address the challenge of segmenting the fibrous cap of the carotid artery in OCT images. Its network structure and modulation mechanism fully consider the characteristics of high noise and blurred boundaries in carotid artery OCT images, and can maintain stable performance even under complex image conditions, making it highly clinically applicable.

[0173] Innovative use of the KAN module: The Kolmogorov-Arnold Network (KAN) is introduced into the field of medical image segmentation as a core component of the network. KAN's learnable activation function provides stronger function fitting capabilities, and compared to traditional networks using fixed activation functions, it can more accurately model the complex boundary features of fibrous caps, which is a key technological foundation for improving segmentation accuracy. Ablation experiments demonstrate that removing the KAN module leads to a significant performance degradation.

[0174] Excellent generalization ability: It not only performs well on the carotid artery OCT dataset, but also achieves leading performance on public datasets such as BUSI (breast ultrasound) and AISD (stroke CT), proving that the framework has good versatility and transferability.

[0175] This embodiment employs the above scheme, using a dual-branch fusion strategy and a channel-level weighting mechanism to calculate the importance weights of each channel in the deep semantic feature map corresponding to the deep semantic features and the interaction feature map corresponding to the original interaction information. Based on these importance weights, the deep semantic feature map and the interaction feature map are scaled channel-by-channel to enhance the key texture and boundary features of the fibrous cap region in both maps, while suppressing vascular background noise. The weighted deep semantic feature map and the interaction feature map are then fused element-wise to obtain a fused feature map. The fused feature map is then processed... The system uses the Sigmoid activation function to obtain the segmentation prediction result for the current round. It receives user feedback and determines whether to iterate the segmentation prediction result based on this feedback, outputting the final segmentation result. This significantly improves the boundary clarity and pixel-level accuracy of carotid OCT image fibrous cap segmentation. It also supports real-time interactive iterative correction, ensuring the segmentation result highly matches clinical diagnostic needs. This effectively solves the problems of low segmentation accuracy, lack of interactivity, and poor efficiency of manual correction in traditional methods. The system offers high image segmentation accuracy, low interaction cost, strong targeting, good robustness, and excellent generalization ability, improving the speed and efficiency of carotid OCT fibrous cap segmentation.

[0176] Accordingly, the present invention further provides a carotid artery OCT fiber cap segmentation device.

[0177] Reference Figure 6 , Figure 6 This is a functional block diagram of the first embodiment of the carotid artery OCT fiber cap segmentation device of the present invention.

[0178] In a first embodiment of the carotid artery OCT fiber cap segmentation device of the present invention, the carotid artery OCT fiber cap segmentation device includes: Weighting module 10 is used to acquire the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round, and to weight the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain the weighted input features; Feature extraction module 20 is used to extract features from the input features using the KAN backbone feature extraction module to obtain deep semantic features, dynamically adjust the segmentation probability map according to the user's new click to obtain the modulated target probability map, and generate the original interaction information according to the target probability map; The fusion prediction module 30 is used to fuse the deep semantic features and the original interaction information with attention features through a dual-branch fusion strategy to obtain the segmentation prediction result of the current round, and decide whether to iterate the segmentation prediction result based on user feedback information, and output the final segmentation result.

[0179] The steps for implementing each functional module of the carotid artery OCT fiber cap segmentation device can be referred to in the various embodiments of the carotid artery OCT fiber cap segmentation method of the present invention, and will not be repeated here.

[0180] Furthermore, this embodiment of the invention also proposes a storage medium storing a carotid artery OCT fibrous cap segmentation program, which, when executed by a processor, performs the operations described in the above-described carotid artery OCT fibrous cap segmentation method embodiment.

[0181] Those skilled in the art will understand that all or part of the steps in the methods described above can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium is a computer-readable storage medium, including: USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media that can store program code.

[0182] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0183] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0184] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for segmenting the carotid artery using an OCT fibrous cap, characterized in that, The carotid artery OCT fibrous cap segmentation method includes: The original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round are obtained. The original carotid artery OCT image, the user click feature map, and the segmentation probability map are weighted to obtain the weighted input features. The input features are extracted using the KAN backbone feature extraction module to obtain deep semantic features. The segmentation probability map is dynamically adjusted according to the user's new clicks to obtain the modulated target probability map. The original interaction information is generated based on the target probability map. The deep semantic features and the original interaction information are fused using a dual-branch fusion strategy to obtain the segmentation prediction result for the current round. Based on user feedback, it is determined whether to iterate the segmentation prediction result and output the final segmentation result.

2. The method for segmenting the carotid artery using an OCT fibrous cap as described in claim 1, characterized in that, The process of acquiring the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round, and then weighting the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain weighted input features, includes: Raw carotid artery optical coherence tomography (OCT) images are acquired via a medical imaging system interface. A user click feature map is obtained by using an interactive user interface to generate user click marks on the original carotid artery OCT image; Load the segmentation probability map predicted by the previous round of carotid artery OCT fibrous cap segmentation; The HICAM information preprocessing module adaptively assigns weights to the original carotid artery OCT image, the user-clicked feature map, and the segmentation probability map to obtain weighted input features.

3. The carotid artery OCT fibrous cap segmentation method as described in claim 2, characterized in that, The step involves adaptively assigning weights to the original carotid artery OCT image, the user-clicked feature map, and the segmentation probability map using the HICAM information preprocessing module to obtain weighted input features, including: The first-level channel attention modulation module of the information preprocessing module HICAM obtains the global descriptor of each channel through global average pooling, and dynamically learns the activation weight of each channel using gate control. The original carotid artery OCT image, the user click feature map, and the segmentation probability map are processed in parallel using convolutional kernels of different sizes through HICAM's multi-scale feature extraction module to obtain contextual information at different scales. The HICAM's two-level channel attention modulation module fuses contextual information at different scales to generate an adaptive weight map. Based on the global descriptor and the activation weight, the confidence scores of the original carotid artery OCT image, the user click feature map, and the segmentation probability map are weighted and fused to obtain the weighted input features.

4. The carotid artery OCT fibrous cap segmentation method as described in claim 1, characterized in that, The KAN backbone feature extraction module is used to extract features from the input features to obtain deep semantic features. The segmentation probability map is dynamically adjusted based on user clicks to obtain a modulated target probability map. Original interaction information is generated based on the target probability map, including: The KAN backbone feature extraction module is used to achieve flexible nonlinear mapping through multiple KAN linear transformation layers. Wavelet convolution spatial enhancement is introduced to decompose the input features at multiple scales to obtain deep semantic features. The modulation radius is adaptively calculated based on the minimum Euclidean distance between the current new click point in the new user clicks and the opposite category click point in the previous segmentation. A linear distance mask is generated based on the modulation radius, and the segmentation probability map is nonlinearly modulated by combining gamma correction and additive masking to obtain the modulated target probability map. Original interaction information is generated based on the target probability map, the original carotid artery OCT image, and the user click feature map.

5. The carotid artery OCT fibrous cap segmentation method as described in claim 4, characterized in that, The step of generating original interaction information based on the target probability map, the original carotid artery OCT image, and the user click feature map includes: The target probability map, the original carotid artery OCT image, and the user click feature map are stitched together at the channel level to obtain a stitched image; The stitched image is input into a lightweight residual network, and the local details and global contextual semantics of the newly clicked area by the user are captured through convolutional layers to generate original interaction information that integrates user interaction intent and area confidence.

6. The method for segmenting the carotid artery using an OCT fibrous cap as described in claim 1, characterized in that, The deep semantic features and the original interaction information are fused using a dual-branch post-fusion strategy to obtain the segmentation prediction result for the current round. Based on user feedback, a decision is made on whether to iterate the segmentation prediction result, and the final segmentation result is output, including: The importance weights of each channel in the deep semantic feature map corresponding to the deep semantic features and the interaction feature map corresponding to the original interaction information are calculated by using a channel-level weighting mechanism after the dual-branch fusion strategy. The deep semantic feature map and the interaction feature map are scaled channel by channel according to the importance weights to enhance the key texture and boundary features of the fibrous cap region in the deep semantic feature map and the interaction feature map, while suppressing the vascular background noise in the deep semantic feature map and the interaction feature map. The weighted deep semantic feature map and the interaction feature map are then fused element by element to obtain a fused feature map. The fused feature map is processed by the Sigmoid activation function to obtain the segmentation prediction result for the current round; Receive user feedback information, decide whether to iterate the segmentation prediction result based on the user feedback information, and output the final segmentation result.

7. The carotid artery OCT fibrous cap segmentation method as described in claim 6, characterized in that, The step of receiving user feedback, determining whether to iterate the segmentation prediction result based on the user feedback information, and outputting the final segmentation result includes: Receive the segmentation prediction results from the user through an interactive interface. Receive user feedback information from users through an interactive interface regarding the segmentation prediction results; When the user feedback is satisfactory, no iteration is performed, and the segmentation prediction result is used as the final segmentation result and output. When the user feedback is unsatisfactory, the process is repeated according to a preset number of iterations until the user feedback is satisfactory.

8. A carotid artery OCT fiber cap segmentation device, characterized in that, The carotid artery OCT fiber cap segmentation device includes: The weighting module is used to acquire the original carotid artery OCT image, the user click feature map, and the segmentation probability map predicted in the previous round, and to weight the original carotid artery OCT image, the user click feature map, and the segmentation probability map to obtain the weighted input features; The feature extraction module is used to extract features from the input features using the KAN backbone feature extraction module to obtain deep semantic features, dynamically adjust the segmentation probability map according to the user's new clicks to obtain the modulated target probability map, and generate the original interaction information according to the target probability map. The fusion prediction module is used to fuse the deep semantic features and the original interaction information with attention features through a dual-branch fusion strategy to obtain the segmentation prediction result of the current round, and decide whether to iterate the segmentation prediction result based on user feedback information, and output the final segmentation result.

9. A carotid artery OCT fiber cap segmentation device, characterized in that, The carotid artery OCT fibrous cap segmentation device includes: a memory, a processor, and a carotid artery OCT fibrous cap segmentation program stored in the memory and executable on the processor, the carotid artery OCT fibrous cap segmentation program being configured to implement the steps of the carotid artery OCT fibrous cap segmentation method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a carotid artery OCT fibrous cap segmentation program, which, when executed by a processor, implements the steps of the carotid artery OCT fibrous cap segmentation method as described in any one of claims 1 to 7.

Images