Wavelet flow matching based OCT to OCTA generation method and device

Abstract

The application provides an OCT-to-OCTA generation method and device based on wavelet flow matching, and belongs to the technical field of medical image generation.The application implements three-dimensional discrete wavelet transform on a three-dimensional structure OCT image and a real three-dimensional OCTA image, and obtains corresponding wavelet domain representation; a wavelet domain conditional velocity field is constructed, a continuous flow matching model is combined, wavelet coefficients of the structure OCT are transformed into wavelet coefficients of a synthetic OCTA, and then a three-dimensional synthetic OCTA image is reconstructed through inverse wavelet transform.The application also introduces retinal depth contrast alignment constraints, spatial retinal layer projection constraints and structure focusing loss functions, and anatomical structure constraints are added to the generation process from three dimensions of blood vessel depth positioning, multi-view spatial consistency and blood vessel area key learning.The application can effectively reduce blood vessel fracture, adhesion and artifacts in the synthetic OCTA image, improve the retention capability of high-frequency details such as capillaries, and obtain a three-dimensional synthetic OCTA image which is more real and accurate in anatomy.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and cross-modal image generation, and in particular to an OCT to OCT generation method and apparatus based on wavelet flow matching. Background Technology

[0002] Optical coherence tomography (OCT), a core imaging technology for retinal disease diagnosis, has been widely used for morphological assessment of diseases such as diabetic retinopathy and age-related macular degeneration since its introduction in 1991. Based on this, optical coherence tomography angiography (OCTA) has been developed to achieve non-invasive depth-resolved visualization of retinal vessels by detecting blood flow contrast. In synergy with OCT, it provides a dual perspective of structure and vessels, significantly improving diagnostic accuracy. Specifically, this technology system covers the entire process from structural OCT image acquisition to vascular information reconstruction, including key steps such as three-dimensional volumetric translation, multi-scale feature extraction, and anatomical structural constraints.

[0003] However, existing OCT-to-OCTA generation methods directly use two-dimensional or three-dimensional generation models to treat OCT as a random conditional signal, without fully considering the sparsity and multi-scale tissue characteristics of retinal vascular signals. This may lead to anatomical inaccuracies such as difficulty in recovering fine capillary networks, vascular rupture, and artifact proliferation, thus affecting the reliability of clinical diagnosis. In addition, although traditional diffusion models generate high-quality signals, they require iterative reverse random paths, resulting in high computational costs and making it difficult to balance efficiency and accuracy. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose an OCT to OCT generation method based on wavelet flow matching.

[0006] Another objective of this invention is to provide an OCT-to-OCTA generation device based on wavelet flow matching.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes an OCT-to-OCTA generation method based on wavelet flow matching, comprising:

[0010] S1. Perform three-dimensional discrete wavelet transform on the acquired three-dimensional structure OCT image and the corresponding three-dimensional real OCT image to obtain their respective wavelet domain representations. S2, construct a velocity field conditioned on the source OCT wavelet coefficients in the wavelet domain, and transform the source OCT wavelet coefficients into synthesized OCT wavelet coefficients step by step along a straight path using a continuous flow matching model; S3, perform inverse wavelet transform on the synthesized OCTA wavelet coefficients to reconstruct the synthesized three-dimensional OCTA image; S4. The synthesized three-dimensional OCTA image is optimized through multi-dimensional anatomical structure constraints, including depth contrast alignment constraints, spatial projection constraints, and structural focusing constraints, to enhance the anatomical accuracy of the generated blood vessels.

[0011] In one embodiment of the present invention, S1 includes: The 3D discrete wavelet transform operator is used to transform the three-dimensional structure OCT image to obtain the source OCT wavelet coefficients; The same 3D discrete wavelet transform operator is used to transform the three-dimensional real OCTA image to obtain the real OCTA wavelet coefficients. These wavelet coefficients contain multiple scales s and directional subbands b, as follows: .

[0012] In one embodiment of the present invention, S2 includes: Define continuous flow in the wavelet domain Its variation is due to the source OCT wavelet coefficients Conditional learnable velocity field Control, satisfaction and from Points to The synthesized OCTA wavelet coefficients are obtained. .

[0013] In one embodiment of the present invention, it further includes: The continuous flow matching model is optimized using flow matching loss for paired samples. Define the linear interpolation path Target speed The flow matching loss is: .

[0014] In one embodiment of the present invention, S3 includes: Employing 3D inverse discrete wavelet transform operator The synthesized OCTA wavelet coefficients Transformed into a synthetic 3D OCTA image .

[0015] In one embodiment of the present invention, S4 includes: Extract slices along each axis from real OCTA images and synthetic OCTA images. and Using a shared depth encoder Map it to a normalized deep embedding vector and ; Contrastive learning is used to make the embedding vectors of the real and synthetic slices at the same depth location close to each other, while the embedding vectors at different depths are far apart. The loss function is... .

[0016] In one embodiment of the present invention, it further includes: The spatial projection constraint includes voxel-level L1 loss and layer projection loss. The voxel-level L1 loss is: The layer projection loss is within a predefined retinal layer range. A two-dimensional projection image is obtained by performing maximum or mean projection on a three-dimensional image. The L1 loss is calculated for the projection maps of the real and synthetic images: , in For the projection operator, the total projection constraint is: .

[0017] To achieve the above objectives, a second aspect of the present invention provides an OCT-to-OCTA generation device based on wavelet flow matching, comprising: The numerical wavelet transform module is used to perform three-dimensional discrete wavelet transform on the acquired three-dimensional structure OCT image and the corresponding three-dimensional real OCT image to obtain their respective wavelet domain representations. The wavelet domain velocity field construction and transformation module is used to construct a velocity field conditioned on the source OCT wavelet coefficients in the wavelet domain. The source OCT wavelet coefficients are gradually transformed into synthesized OCT wavelet coefficients along a straight path through a continuous flow matching model. The inverse wavelet image reconstruction module is used to perform inverse wavelet transform on the synthesized OCTA wavelet coefficients to reconstruct the synthesized three-dimensional OCTA image. A multi-dimensional constraint optimization module is used to optimize the synthesized three-dimensional OCTA image through multi-dimensional anatomical structure constraints, including depth contrast alignment constraints, spatial projection constraints, and structural focusing constraints, to enhance the anatomical accuracy of the generated blood vessels.

[0018] An OCT-to-OCTA generation method and apparatus based on wavelet flow matching according to an embodiment of the present invention can generate OCTA images with accurate anatomical structures and continuous and complete blood vessels from structural OCT images, effectively solving the problems of difficulty in restoring small blood vessels, blood vessel rupture and artifact proliferation in existing methods.

[0019] To achieve the above objectives, a third aspect of this application provides a computer device comprising: a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing the method described in the first aspect embodiment.

[0020] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the first aspect.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of an OCT to OCT generation method based on wavelet flow matching according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a deep learning network model based on wavelet flow matching according to an embodiment of the present invention; Figure 3 This is a visual comparison chart of the generation results of the present invention on the OCTA-500 dataset and existing methods; Figure 4 This is a comparison chart of ablation experiment results on the OCTA-500 dataset according to an embodiment of the present invention; Figure 5 This is a structural diagram of an OCT-to-OCTA generation device based on wavelet flow matching according to an embodiment of the present invention; Figure 6 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following describes, with reference to the accompanying drawings, an OCT-to-OCTA generation method and apparatus based on wavelet flow matching according to an embodiment of the present invention.

[0026] Figure 1 This is a flowchart of an OCT-to-OCTA generation method based on wavelet flow matching according to an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: S1. Perform three-dimensional discrete wavelet transform on the acquired three-dimensional structure OCT image and the corresponding three-dimensional real OCT image to obtain their respective wavelet domain representations. S2, construct a velocity field conditioned on the source OCT wavelet coefficients in the wavelet domain, and transform the source OCT wavelet coefficients into synthesized OCT wavelet coefficients step by step along a straight path using a continuous flow matching model; S3, perform inverse wavelet transform on the synthesized OCTA wavelet coefficients to reconstruct the synthesized three-dimensional OCTA image; S4. The synthesized three-dimensional OCTA image is optimized through multi-dimensional anatomical structure constraints, including depth contrast alignment constraints, spatial projection constraints, and structural focusing constraints, to enhance the anatomical accuracy of the generated blood vessels.

[0027] Furthermore, a deep learning network based on wavelet flow matching is designed. The core generative network of this method is a conditional flow matching model based on the wavelet domain. First, the input 3DOCT image is processed... and corresponding real 3DOCTA images Perform 3D Discrete Wavelet Transform (DWT) on each component to obtain its wavelet domain representation. and :

[0028] in, This is a 3D discrete wavelet transform operator, whose output includes multiple scales s and directional subbands. wavelet coefficients .

[0029] Furthermore, a continuous flow in a wavelet domain is defined. Its variation is determined by a source OCT wavelet coefficient. Conditional learnable velocity field control:

[0030] The flow from Start, accumulate points The synthesized OCTA wavelet coefficients are obtained. Finally, the synthesized OCTA image is reconstructed using inverse wavelet transform:

[0031] Furthermore, during the training phase, the network is optimized using a streaming matching loss. For paired samples... Define the linear interpolation path Target speed The flow matching loss is:

[0032] Furthermore, multi-dimensional anatomical constraints are designed. To ensure the anatomical accuracy of the generated blood vessels, this method introduces the following three constraint modules: Retinal Depth Contrast Alignment Module: This module introduces depth contrast alignment loss to address the characteristic of retinal vessels being distributed along specific narrow axial bands. It first extracts axial slices from real and synthetic OCTA images. and Using a shared depth encoder Map it to a normalized deep embedding vector and By using contrastive learning, the embedding vectors of the real and synthetic slices at the same depth location are made as close as possible, while the embedding vectors at different depths are kept far apart, thereby preventing diffusion and misalignment of blood vessels in the depth direction.

[0033] Spatial retinal layer projection constraint module: Based on clinical interpretation habits of OCTA images, this module introduces two complementary constraints. One is voxel-level L1 loss. The first is to preserve the fine vascular structure at the original spatial resolution; the second is layer projection loss, within a predefined range of the retinal layer. The 3D image is subjected to maximum or mean projection to obtain a 2D projection map, and the L1 loss is calculated for the projection maps of the real and synthetic images. .in This is the projection operator. The overall projection constraint is:

[0034] Structure-focusing loss module: Given the sparsity of vascular signals in OCTA images, a binary mask based on real OCTA signals is constructed. All the aforementioned losses (such as flow matching loss, voxel loss, etc.) are calculated only in regions where vascular signals are significant, allowing model training to focus more on the vascular structure itself and avoid being dominated by large background areas. The loss after applying the mask is denoted as... .

[0035] Furthermore, multi-dimensional anatomical constraints are designed. The overall training objective of this method is the weighted sum of the above losses, as shown in the following formula:

[0036] in, and Weighting coefficients to balance the various losses.

[0037] like Figure 2 As shown, the deep learning network model based on wavelet flow matching proposed in this invention mainly consists of four core parts. First, the model transforms the input structural OCT image and the paired real OCTA image into the wavelet frequency domain through three-dimensional wavelet transform, decomposing the image into high-frequency and low-frequency sub-bands of multiple scales and directions. Second, in the generation stage, the model adopts a noise-free conditional flow matching framework, starting with the wavelet coefficients of the OCT and transporting them along a deterministic straight-line path from the source distribution to the target distribution, learning a conditional velocity field to predict the direction of change at each step, thereby gradually mapping the wavelet coefficients of the OCT to the wavelet coefficient distribution of the OCTA. Third, the model introduces two anatomical constraints: one is a retinal depth contrast alignment module, which forces the generated vascular signals to converge at the correct depth level through contrastive learning, preventing axial diffusion or misalignment of blood vessels; the other is a spatial retinal layer projection constraint module, which applies constraints at both the three-dimensional voxel level and the two-dimensional clinically commonly used projection level to ensure the spatial consistency of the vascular structure. Finally, addressing the sparse nature of blood vessels, the model employs a structure-focused loss function and constructs a blood vessel mask using an adaptive threshold. This allows the network to focus on the blood vessel region rather than background voxels, significantly improving the quality of capillary generation. During inference, the model only needs to perform wavelet transform on the input OCT image, integrate the learned velocity field from this initial state in a very small number of steps, and then perform inverse wavelet transform to quickly reconstruct the final OCT image.

[0038] like Figure 3As shown, this invention presents a visual comparison of its generation results with several existing methods on the OCTA-500 dataset. The figure displays central slices from three orthogonal projection planes (axial, coronal, and sagittal) extracted from the 3D volume, while also providing magnified areas for a clearer comparison of angiogenesis details. PSNR and SSIM values ​​calculated compared to real OCTA images are labeled below each synthesized slice. The comparison results show that generative adversarial network-based methods (such as the 2D and 3D versions of Pix2Pix and AdjacentGAN) exhibit significant checkerboard artifacts or stripe noise, and overall low brightness. While the diffusion model produces relatively smooth results, it shows numerous erroneous connections and breaks in blood vessels. The TransPro method improves in structure reconstruction and brightness, but still suffers from excessive angiogenesis and adhesion. In comparison, the FlowOCT method proposed in this invention achieves the best visual effect on all projected views: the residual between the generated image and the real OCTA is the smallest, the vascular structure is clearer and more continuous, the capillary details are richer, the overall image quality is closest to the real OCTA, and the PSNR and SSIM quantization indicators also reach the highest values.

[0039] like Figure 4 As shown, ablation experiments were conducted on the OCTA-500 dataset to verify the effectiveness of each core module. The figure compares the generation results under five configurations: removal of the wavelet domain module, removal of the retinal depth contrast alignment module, removal of the spatial projection constraint module, removal of the structural focusing loss function, and the complete model. Visual comparisons show that removing the wavelet domain significantly degrades the overall image quality, resulting in severe loss of high-frequency details; removing the retinal depth contrast alignment module causes obvious depth misalignment and axial diffusion in the vascular structure; removing the spatial projection constraint disrupts the consistency between the 3D voxels and the 2D projection, leading to geometric inconsistencies across views; and removing the structural focusing loss significantly reduces the refinement of capillaries, resulting in blurred small vessels. The complete model of this invention performs best among all comparison configurations, generating continuous and complete vascular structures with accurate depth localization and projection accuracy. Figure 1 The capillary details are clearly visible, fully demonstrating the indispensability of the wavelet domain flow matching framework and various anatomical constraint modules in improving the quality of OCTA generation.

[0040] The embodiments of this invention also have the following technical effects: Improved generation quality: On the OCTA-500 dataset, compared with the best baseline method, this invention improves PSNR by 1.7%~2.8%, SSIM by 2.1%~4.8%, reduces LPIPS by 8.1%~16.6%, and reduces FID by 8.7%~29.0%. High-frequency detail preservation: The high-frequency energy ratio (HFER) of this invention reaches 94.2%, significantly higher than the diffusion model (78.5%) and Pix2Pix (71.3%), proving that wavelet domain flow matching effectively preserves high-frequency structures such as capillaries. Improved inference efficiency: This invention only requires 3 steps of ODE solving to complete the generation, compared with the traditional diffusion model (requiring 50~1000 steps), improving the inference speed by 1~2 orders of magnitude. Downstream diagnostic effectiveness: The disease classifier trained using the synthetic data of this invention achieves an accuracy of 46.67% on unseen datasets, significantly better than other synthetic data (23.33%~40.00%), proving that the generated images preserve the pathological features required for clinical diagnosis.

[0041] Improving Clinical Accessibility: This invention can synthesize high-quality OCTA images from conventionally acquired structural OCT images, enabling primary hospitals and medical institutions in remote areas without OCTA equipment to obtain vascular imaging information, reducing patient referral and examination costs. Assisting Clinical Diagnosis: The high-fidelity OCTA images generated by this invention can be used for downstream tasks such as vascular segmentation, FAZ measurement, and disease classification, providing doctors with objective vascular biomarkers to assist in the early screening and disease assessment of diseases such as diabetic retinopathy and age-related macular degeneration. Improving Diagnostic Efficiency: This invention requires only 3 steps of reasoning to generate a complete 3D OCTA volume, significantly shortening image acquisition and processing time compared to traditional methods, reducing motion artifacts caused by prolonged fixation in patients, and improving the overall efficiency of clinical workflows. Lowering Equipment Barriers: This invention enhances the vascular imaging capabilities of OCT equipment through software algorithms, expanding the functionality of existing OCT equipment without changing the hardware, and has good industrial promotion value.

[0042] To achieve the above embodiments, such as Figure 5 As shown, this embodiment also provides an OCT-to-OCTA generation device 10 based on wavelet flow matching, including: The numerical wavelet transform module 100 is used to perform three-dimensional discrete wavelet transform on the acquired three-dimensional structure OCT image and the corresponding three-dimensional real OCT image respectively to obtain their respective wavelet domain representations. The wavelet domain velocity field construction and transformation module 200 is used to construct a velocity field conditioned on the source OCT wavelet coefficients in the wavelet domain. The source OCT wavelet coefficients are gradually transformed into synthesized OCT wavelet coefficients along a straight path through a continuous flow matching model. The inverse wavelet image reconstruction module 300 is used to perform inverse wavelet transform on the synthesized OCTA wavelet coefficients to reconstruct the synthesized three-dimensional OCTA image. The multi-dimensional constraint optimization module 400 is used to optimize the synthesized three-dimensional OCTA image through multi-dimensional anatomical structure constraints, including depth contrast alignment constraints, spatial projection constraints, and structural focusing constraints, to enhance the anatomical accuracy of the generated blood vessels.

[0043] An OCT-to-OCTA generation device based on wavelet flow matching according to an embodiment of the present invention can generate OCTA images with accurate anatomical structures and continuous and complete blood vessels from structural OCT images, effectively solving the problems of difficulty in restoring small blood vessels, blood vessel rupture and artifact proliferation in existing methods.

[0044] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 6 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0045] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0046] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0047] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for generating OCT to OCT based on wavelet flow matching, characterized in that, include: S1. Perform three-dimensional discrete wavelet transform on the acquired three-dimensional structure OCT image and the corresponding three-dimensional real OCT image to obtain their respective wavelet domain representations. S2, construct a velocity field conditioned on the source OCT wavelet coefficients in the wavelet domain, and transform the source OCT wavelet coefficients into synthesized OCT wavelet coefficients step by step along a straight path using a continuous flow matching model; S3, perform inverse wavelet transform on the synthesized OCTA wavelet coefficients to reconstruct the synthesized three-dimensional OCTA image; S4. The synthesized three-dimensional OCTA image is optimized through multi-dimensional anatomical structure constraints, including depth contrast alignment constraints, spatial projection constraints, and structural focusing constraints, to enhance the anatomical accuracy of the generated blood vessels.

2. The method as described in claim 1, characterized in that, S1 includes: The 3D discrete wavelet transform operator is used to transform the three-dimensional structure OCT image to obtain the source OCT wavelet coefficients; The same 3D discrete wavelet transform operator is used to transform the three-dimensional real OCTA image to obtain the real OCTA wavelet coefficients. These wavelet coefficients contain multiple scales s and directional subbands b, as follows: 。 3. The method as described in claim 1, characterized in that, S2 includes: Define continuous flow in the wavelet domain Its variation is due to the source OCT wavelet coefficients Conditional learnable velocity field Control, satisfy and from Points to The synthesized OCTA wavelet coefficients are obtained. .

4. The method as described in claim 3, characterized in that, The method further includes: The continuous flow matching model is optimized using flow matching loss for paired samples. Define the linear interpolation path Target speed The flow matching loss is: 。 5. The method as described in claim 1, characterized in that, The S3 includes: Employing 3D inverse discrete wavelet transform operator The synthesized OCTA wavelet coefficients Transformed into a synthetic 3D OCTA image .

6. The method as described in claim 1, characterized in that, The S4 includes: Extract slices along each axis from real OCTA images and synthetic OCTA images. and Using a shared depth encoder Map it to a normalized deep embedding vector and ; Contrastive learning is used to make the embedding vectors of the real and synthetic slices at the same depth position close to each other, while the embedding vectors at different depths are far apart. The loss function is... .

7. The method as described in claim 1, characterized in that, The method further includes: The spatial projection constraint includes voxel-level L1 loss and layer projection loss. The voxel-level L1 loss is: The layer projection loss is within a predefined retinal layer range. A two-dimensional projection image is obtained by performing maximum or mean projection on a three-dimensional image. The L1 loss is calculated for the projection maps of the real and synthetic images: ， in For the projection operator, the total projection constraint is: .

8. An OCT-to-OCTA generation device based on wavelet flow matching, characterized in that, include: The numerical wavelet transform module is used to perform three-dimensional discrete wavelet transform on the acquired three-dimensional structure OCT image and the corresponding three-dimensional real OCT image to obtain their respective wavelet domain representations. The wavelet domain velocity field construction and transformation module is used to construct a velocity field conditioned on the source OCT wavelet coefficients in the wavelet domain. The source OCT wavelet coefficients are gradually transformed into synthesized OCT wavelet coefficients along a straight path through a continuous flow matching model. The inverse wavelet image reconstruction module is used to perform inverse wavelet transform on the synthesized OCTA wavelet coefficients to reconstruct the synthesized three-dimensional OCTA image. A multi-dimensional constraint optimization module is used to optimize the synthesized three-dimensional OCTA image through multi-dimensional anatomical structure constraints, including depth contrast alignment constraints, spatial projection constraints, and structural focusing constraints, to enhance the anatomical accuracy of the generated blood vessels.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images