Medical image focus labeling method and system based on deep learning

Abstract

The invention relates to the technical field of focus labeling, in particular to a medical image focus labeling method and system based on deep learning, and the method comprises the steps: collecting a multi-modal medical image, and carrying out the noise reduction and contrast enhancement preprocessing to obtain a medical enhanced image; segmenting a focus area through a deep convolutional network and generating a focus labeling mask image; and performing post-processing optimization on the mask image, inputting an optimization result into the sub-pixel level annotation generation model, and outputting a high-precision focus annotation coordinate sequence. According to the method, image quality and focus visibility are improved through multi-modal preprocessing, accurate segmentation is realized by using a deep convolutional network, traditional pixel-level precision limitation is broken through in combination with post-processing and a sub-pixel-level model, the problems of noise interference, contrast difference and boundary blur are effectively solved, and focus labeling fineness and spatial positioning precision are remarkably improved.

Description

A Deep Learning-Based Method and System for Labeling Lesions in Medical Images Technical Field

[0001] This invention relates to the field of lesion annotation technology, and in particular to a method and system for medical image lesion annotation based on deep learning. Background Technology

[0002] With the rapid development of medical imaging technology, multimodal imaging methods such as CT and MRI are playing an increasingly important role in clinical disease diagnosis, especially in the early detection and treatment assessment of major diseases such as tumors and stroke, where accurate lesion localization and annotation have become crucial. Traditional lesion annotation mainly relies on manual drawing by radiologists, which is not only time-consuming and labor-intensive but also easily influenced by subjective experience, leading to individual differences and poor repeatability in the annotation results. In recent years, deep learning technology has made significant progress in the field of medical image analysis, especially image segmentation models based on convolutional neural networks (such as U-Net and its variants), which have been widely used in automatic lesion segmentation tasks, greatly improving processing efficiency and consistency.

[0003] However, existing methods still face many challenges in practical applications: on the one hand, raw medical images are often affected by noise interference and low tissue contrast, especially in multimodal data fusion scenarios, where contrast differences caused by different imaging mechanisms further affect the accuracy of lesion identification; on the other hand, most segmentation models output lesion masks with pixel-level precision, and problems such as blurred boundaries and artifact interference limit their application in high-precision diagnosis and treatment (such as surgical navigation and radiotherapy target delineation). In addition, there is a lack of effective sub-pixel-level post-processing mechanisms between the segmentation results and the lesion coordinate annotations that can be used for quantitative analysis, making it difficult to meet the clinical demand for high precision in key geometric parameters such as lesion center and edge.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a medical image lesion annotation method and system based on deep learning, which aims to solve the technical problems of existing medical image lesion annotation methods, such as large noise interference, significant differences in multimodal contrast, blurred segmentation boundaries, and lack of sub-pixel level precise positioning mechanism, resulting in insufficient annotation accuracy and difficulty in meeting the needs of high-precision clinical diagnosis and treatment.

[0006] To achieve the above objectives, this invention provides a deep learning-based method for annotating medical image lesions, the method comprising:

[0007] Multimodal medical images are acquired using medical imaging equipment, and image processing techniques are used to sequentially reduce noise and enhance the contrast of lesions in the medical images to obtain enhanced medical images.

[0008] A deep convolutional network segmenter is constructed to segment the lesion region, and a lesion annotation mask image is generated based on the lesion segmentation results;

[0009] Post-processing techniques are used to optimize the lesion annotation mask image, and the optimized mask image is input into the sub-pixel level annotation generation model. The sub-pixel level annotation generation model outputs a high-precision lesion annotation coordinate sequence.

[0010] Optionally, the step of acquiring multimodal medical images using medical imaging equipment and sequentially performing noise reduction and lesion contrast enhancement processing on the medical images using image processing techniques to obtain enhanced medical images includes:

[0011] The system integrates CT and MRI dual-modal imaging equipment to acquire medical images. Based on the grayscale distribution feedback of the medical images, it uses an adaptive window width and window level adjustment algorithm to optimize the dynamic range of the images in real time, so as to eliminate imaging noise and balance the differences in multimodal contrast.

[0012] An improved nonlocal mean filtering algorithm is used to perform joint denoising on medical images, dynamically balancing the filtering weights for local similarity and global structural fidelity.

[0013] The reflectance component of the denoised medical image is decomposed based on the multi-scale Retinex-CLAHE fusion algorithm, and the reflectance component is fused with the local contrast adaptive histogram equalization result to enhance the contrast of the lesion area, thus obtaining the enhanced medical image.

[0014] Optionally, the reflectance component of the denoised medical image is decomposed based on the multi-scale Retinex-CLAHE fusion algorithm, and the reflectance component is fused with the local contrast adaptive histogram equalization result to enhance the contrast of the lesion area, resulting in a medically enhanced image, including:

[0015] The original medical image is converted to the HSV color space, the gradient magnitude map of the original medical image in the saturation channel is calculated, and the gradient magnitude map is binarized by the adaptive Otsu threshold segmentation algorithm to obtain the initial lesion candidate map.

[0016] The maximum inter-class variance thresholds of CT bone tissue and MRI soft tissue in the original medical images are integrated to generate a multimodal binary map. The initial lesion candidate map is then fused with the multimodal binary map and artifact regions are removed to obtain an optimized lesion label map.

[0017] The optimized lesion label map is used to perform morphological closing operations on the original medical images to capture the reflection components of the microscopic lesion edges, mesoscopic tissue structures and macroscopic background in the closing operation area.

[0018] Based on a weighted fusion strategy, the reflection components of the microscopic lesion edge, mesoscopic tissue structure, and macroscopic background are integrated to form a medical enhancement image with enhanced lesion contrast.

[0019] Optionally, the construction of a deep convolutional network segmenter to segment the lesion region and generate a lesion-annotated mask image based on the lesion segmentation result includes:

[0020] A deep convolutional network segmenter is constructed and trained end-to-end based on a predefined medical image training set. The trained deep convolutional network segmenter is then used to output a probability heatmap of the lesion region.

[0021] The encoder of the deep convolutional network segmenter is used as the encoder of the improved U-Net model, and the medical enhanced image is input into the encoder, and a high-dimensional feature map is output at the end of the encoder.

[0022] The probability heatmap is mapped to a channel weight vector through the channel attention module of the improved U-Net model. The channel weight vector is then multiplied with the high-dimensional feature map channel by channel to focus on the lesion region of the feature map.

[0023] In the decoder stage of the improved U-Net model, a multi-scale feature fusion module is introduced to gradually restore the resolution of the lesion region in the feature map, thereby obtaining the segmented lesion-annotated mask image.

[0024] Optionally, the multi-scale feature fusion module includes several dilated convolutional layers with different receptive fields, used to capture multi-scale lesion structural features from micro to macro on the same level feature map;

[0025] Among them, the small receptive field hollow convolutional layer is used to focus on the boundary details of individual lesion cells; the medium receptive field hollow convolutional layer is used to perceive the interface transition area between the lesion and the surrounding tissue; and the large receptive field hollow convolutional layer is used to analyze the overall morphology and spatial distribution pattern of the lesion.

[0026] Optionally, the step of optimizing the lesion annotation mask image using post-processing techniques and inputting the optimized mask image into a sub-pixel-level annotation generation model, and outputting a high-precision lesion annotation coordinate sequence through the sub-pixel-level annotation generation model, includes:

[0027] A distance transformation is performed on the lesion-annotated mask image to generate a distance field map, and the main edge components of the lesion are extracted based on the gradient distribution in the distance field map;

[0028] The contour refinement search range is set based on the main edge components of the lesion, and the lesion boundary is refined using an active contour model. At the same time, artifact interference is suppressed during the refinement process to obtain an optimized lesion boundary.

[0029] An adaptive edge filter is constructed in the distance field map based on the optimized lesion boundary, and a smooth lesion boundary image is generated using the adaptive edge filter;

[0030] Radial projection is performed on the lesion boundary image along the optimized lesion boundary to generate a boundary intensity curve. Subpixel-level peak detection is then performed on the boundary intensity curve using B-spline interpolation to obtain the lesion center coordinate sequence.

[0031] The lesion center coordinate sequence is input into the sub-pixel level annotation generation model to calculate the lesion annotation parameters to obtain a high-precision lesion annotation coordinate sequence. The lesion annotation parameters are then compared with the medical standard library, and the sub-pixel level annotation generation model is optimized based on the comparison results.

[0032] Optionally, the expression for the adaptive edge filter is:

[0033]

[0034] In the formula, The functional form representing the adaptive edge filter; Represents the original spatial domain coordinates; Indicates the principal direction angle of the filter; Indicates the center distance of the filter; This represents the standard deviation of a Gaussian function in the spatial domain. This indicates that the original spatial domain coordinates are rotated around the center of the lesion. The new coordinates.

[0035] Optionally, the process of suppressing artifact interference during refinement to obtain optimized lesion boundaries includes:

[0036] Morphological opening operations are performed on the distance field map to remove isolated artifacts, generating preliminary continuous lesion boundaries. The angles corresponding to the gradient peaks in the preliminary continuous lesion boundaries are extracted as a coarse estimation reference for the lesion boundaries.

[0037] Based on a coarse estimation reference and combined with the average size of lesions in current medical images, a contour energy threshold is set to filter out artifact regions below the contour energy threshold.

[0038] The initial continuous lesion boundaries are clustered according to curvature to generate the initial lesion boundary contours. The smoothness of the initial lesion boundary contours is verified, and the cluster centers are adjusted based on the verification results until they meet the medical anatomical constraints.

[0039] The broken boundaries are sequentially interpolated and repaired according to the initial lesion boundary contour that meets medical anatomical constraints. The accuracy of the lesion boundary is evaluated based on the repair results to obtain the final optimized lesion boundary.

[0040] Optionally, the step of radially projecting the lesion boundary image along the optimized lesion boundary to generate a boundary intensity curve, and then performing sub-pixel-level peak detection on the boundary intensity curve using B-spline interpolation to obtain the lesion center coordinate sequence includes:

[0041] Radial projection is performed on the lesion boundary image according to the normal direction of the optimized lesion boundary to obtain a one-dimensional boundary intensity curve. Dense interpolation points are generated at the discrete sampling points of the one-dimensional boundary intensity curve according to the sub-pixel interval.

[0042] The dense interpolation points on the smoothed boundary intensity curve are fitted using a cubic B-spline function, and the first derivative of the cubic B-spline function is obtained to locate the critical point as a candidate location for the lesion boundary.

[0043] Calculate the second derivative at the critical point and determine the type of extremum. If the second derivative is greater than or equal to zero, it indicates that the dense interpolation point is a local minimum point and corresponds to the lesion boundary. Otherwise, it is a local maximum point and corresponds to the background area.

[0044] Within the neighborhood of subpixel-level peaks, only the subpixel peaks corresponding to local minima are retained, and the subpixel peaks are back-inferred to the original image coordinate system based on the optimized lesion boundary to generate a high-precision lesion center coordinate sequence.

[0045] Furthermore, to achieve the above objectives, the present invention also provides a deep learning-based medical image lesion annotation system, the system comprising: a memory, a processor, and a deep learning-based medical image lesion annotation method program stored in the memory and executable on the processor, the deep learning-based medical image lesion annotation method program being configured to implement the steps of the deep learning-based medical image lesion annotation method as described above.

[0046] This invention provides a deep learning-based method for medical image lesion annotation. The method effectively improves image quality and lesion visibility by combining noise reduction and lesion contrast enhancement preprocessing of multimodal medical images, thereby enhancing the model's ability to identify complex lesions. A deep convolutional network segmenter is used to achieve accurate segmentation of the lesion region, generating a high-quality initial annotation mask. Furthermore, post-processing optimization and a sub-pixel-level annotation generation model are introduced to overcome the problems of blurred boundaries and low positioning accuracy in traditional segmentation methods. This achieves a leap from pixel-level segmentation to accurate sub-pixel-level coordinate output, significantly improving the precision and spatial positioning accuracy of lesion annotation. It meets the clinical application needs for high-precision lesion localization in quantitative analysis, surgical planning, and efficacy evaluation, and exhibits good robustness, automation, and clinical applicability. Attached Figure Description

[0047] Figure 1 is a flowchart illustrating an embodiment of the deep learning-based medical image lesion annotation method of the present invention;

[0048] Figure 2 is a schematic diagram of the specific steps in generating a lesion annotation mask image based on the lesion segmentation result in one embodiment of the deep learning-based medical image lesion annotation method of the present invention.

[0049] Figure 3 is a schematic diagram of the specific steps in an embodiment of the deep learning-based medical image lesion annotation method of the present invention, which outputs a high-precision lesion annotation coordinate sequence through a sub-pixel level annotation generation model.

[0050] 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

[0051] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] Referring to Figure 1, which is a flowchart of an embodiment of the deep learning-based medical image lesion annotation method of the present invention, an embodiment of the deep learning-based medical image lesion annotation method of the present invention is proposed.

[0053] In one embodiment, the deep learning-based medical image lesion annotation method includes:

[0054] Step S100: Multimodal medical images are acquired using medical imaging equipment, and image processing techniques are used to sequentially perform noise reduction and lesion contrast enhancement processing on the medical images to obtain enhanced medical images.

[0055] Medical enhanced images can be multimodal medical images that have undergone denoising and lesion contrast enhancement processing. These images can be used to improve lesion visibility and tissue boundary clarity, providing high-quality input for segmentation models. Furthermore, medical enhanced images can be obtained by sequentially applying denoising filtering and local contrast enhancement algorithms to original CT, MRI, or other images. Denoising and lesion contrast enhancement processing of medical images can be achieved by sequentially applying image denoising and local contrast enhancement algorithms to the original multimodal medical images. For example, denoising and lesion contrast enhancement processing of medical images can be achieved by using nonlocal mean denoising combined with CLAHE contrast enhancement, thereby suppressing imaging noise and enhancing the grayscale difference between lesions and surrounding tissues, improving lesion identifiability.

[0056] Step S200: Construct a deep convolutional network segmenter to segment the lesion region, and generate a lesion annotation mask image based on the lesion segmentation result.

[0057] The deep convolutional network segmenter can be a deep neural network with an encoder-decoder structure, used to segment lesion regions from medical images. It can be used to generate an initial lesion annotation mask image, achieving end-to-end lesion region recognition. In an exemplary embodiment, the deep convolutional network segmenter can be trained on an annotated dataset through supervised learning and utilize multi-scale feature fusion to improve segmentation accuracy. For example, the deep convolutional network segmenter can be one or more of the following, including but not limited to U-Net architecture segmenters, AttentionU-Net segmenters, and nnU-Net adaptive segmenters. The lesion annotation mask image can be a pixel-level image representing the lesion region in binary or probabilistic form, with foreground markers indicating the lesion location. It can be used as an intermediate result to be passed to the post-processing and sub-pixel annotation stages, carrying spatial distribution information of the lesions. For example, the lesion annotation mask image can include hard-threshold binary masks, soft-probabilistic masks, or multi-class tissue masks. Constructing a deep convolutional network segmenter to segment lesion regions can involve inputting medical enhanced images into a pre-trained deep convolutional network and outputting a probability map or binary mask of the lesion region. Furthermore, constructing a deep convolutional network segmenter to segment the lesion region can be achieved by using a U-Net backbone network in conjunction with the Dice loss function for training, thereby enabling automatic and high-fidelity segmentation of the lesion region and generating an initial labeled mask.

[0058] Step S300: Optimize the lesion annotation mask image using post-processing technology, and input the optimized mask image into the sub-pixel level annotation generation model. Output a high-precision lesion annotation coordinate sequence through the sub-pixel level annotation generation model.

[0059] The subpixel-level annotation generation model can be a deep learning or geometric modeling module that converts a pixel-level lesion mask into a continuous spatial coordinate sequence. It can output high-precision subpixel coordinates of lesion boundaries and key points, supporting quantitative analysis and surgical planning. In one specific embodiment, the subpixel-level annotation generation model can generate continuous coordinates based on the optimized mask image through mechanisms such as subpixel interpolation, edge gradient guidance, or parametric curve fitting. For example, the subpixel-level annotation generation model can include, but is not limited to, one or more of the following: a spline-fit-based subpixel annotation model, a gradient field inversion-based subpixel annotation model, and a super-resolution reconstruction-based subpixel annotation model. Post-processing techniques can be used to optimize the lesion annotation mask image by performing morphological operations, connected component analysis, or edge smoothing on the initial mask. Furthermore, post-processing techniques can be used to optimize the lesion annotation mask image by applying opening and closing operations to remove isolated noise points and connect broken regions, thereby eliminating segmentation artifacts, filling holes, smoothing boundaries, and improving the integrity of the mask structure. The optimized mask image is input into a subpixel-level annotation generation model, which outputs a high-precision sequence of lesion annotation coordinates. This can be achieved by using the optimized mask as input to drive the subpixel-level annotation generation model to calculate continuous spatial coordinates. In one exemplary embodiment, this operation can be achieved through subpixel interpolation localization based on the gradient direction of the mask edge, thus overcoming the limitations of the pixel grid and outputting subpixel-level lesion contours or center point coordinates that can be used for geometric measurements.

[0060] Taking the delineation of radiotherapy target areas for brain tumors as an example, the deep learning-based medical image lesion annotation method in this embodiment can be as follows: acquire multimodal MRI (T1, T2, FLAIR) images of the patient, input them into a trained nnU-Net segmenter after noise reduction and contrast enhancement to generate an initial tumor mask; optimize the mask through morphological closing operations and connected component screening to remove small-volume false positive regions; and feed the optimized mask into a gradient-guided subpixel annotation generation model to output the subpixel coordinate sequence of the tumor boundary, which is used by the radiotherapy planning system to accurately delineate the GTV (gross tumor volume) and avoid under-irradiation or damage to normal tissues due to pixel-level errors.

[0061] In one embodiment, multimodal medical images are acquired using medical imaging equipment, and image processing techniques are used to sequentially perform noise reduction and lesion contrast enhancement processing on the medical images to obtain enhanced medical images, including:

[0062] The system integrates CT and MRI dual-modal imaging equipment to acquire medical images. Based on the grayscale distribution feedback of the medical images, it uses an adaptive window width and window level adjustment algorithm to optimize the dynamic range of the images in real time, so as to eliminate imaging noise and balance the differences in multimodal contrast.

[0063] The adaptive window width and level adjustment algorithm can be an image preprocessing algorithm that dynamically adjusts the display window width and window level parameters based on the grayscale distribution of medical images. It can be used to optimize the dynamic range of image display, suppress noise interference, and balance contrast differences between CT and MRI modalities. In this embodiment, the adaptive window width and level adjustment algorithm can analyze the grayscale histogram or local statistical features of the input image in real time and automatically calculate the optimal window width and window level to maximize the visible dynamic range of the lesion region. Furthermore, the adaptive window width and level adjustment algorithm can include, but is not limited to, one or more of the following: percentile-based adaptive window width and level algorithms, window parameter optimization algorithms based on the entropy maximization criterion, and lesion-prior-guided window width and level adjustment algorithms.

[0064] Adaptive window width and level adjustment algorithms are used to optimize the dynamic range of images in real time to eliminate imaging noise and balance multimodal contrast differences. This can be achieved by dynamically calculating and applying optimal window width and level parameters based on the grayscale distribution feedback of CT / MRI images. Furthermore, this operation can be achieved by automatically setting the window width and level based on the 95% confidence interval of the grayscale histogram, or by dynamically locking the window center and adaptively expanding the window width based on the typical HU value range of the lesion. This unifies the visual representation of multimodal images, reduces noise interference in subsequent segmentation, and improves the discriminability of lesion areas.

[0065] An improved nonlocal mean filtering algorithm is used to perform joint denoising on medical images, dynamically balancing the filtering weights for local similarity and global structural fidelity.

[0066] The improved nonlocal mean filtering algorithm can be a denoising algorithm that introduces a dynamic weighting mechanism for local-global structural fidelity on the basis of traditional nonlocal mean filtering. It can be used to effectively suppress imaging noise while preserving lesion edges and tissue structure details, avoiding excessive smoothing. In an exemplary embodiment, the improved nonlocal mean filtering algorithm can dynamically adjust the filtering weights by calculating the similarity between pixel blocks and combining the local gradient change rate and the global structural consistency index. Exemplary examples include, but are not limited to, nonlocal mean filtering guided by structural tensors, nonlocal denoising based on depth feature similarity, and weighted nonlocal mean filtering based on multi-scale block matching.

[0067] An improved nonlocal mean filtering algorithm is used for joint denoising of medical images. The filtering weights, which dynamically balance local similarity and global structural fidelity, can be achieved by introducing structure-aware weights into the nonlocal mean framework to differentiate between local texture and global anatomical structure. Furthermore, this operation can be implemented by constructing a hybrid similarity measure combining local gradient magnitude and global curvature consistency, or by dynamically allocating filtering intensity weights using multi-scale block matching results. This allows for the removal of noise while maintaining lesion boundary sharpness and internal heterogeneity, thus avoiding structural distortion.

[0068] The reflectance component of the denoised medical image is decomposed based on the multi-scale Retinex-CLAHE fusion algorithm, and the reflectance component is fused with the local contrast adaptive histogram equalization result to enhance the contrast of the lesion area, thus obtaining the enhanced medical image.

[0069] The multi-scale Retinex-CLAHE fusion algorithm can be an enhancement algorithm that combines multi-scale Retinex reflectance component decomposition with local contrast adaptive histogram equalization. It can be used to enhance the local contrast of lesion areas while maintaining illumination invariance, thus improving the visibility of low-contrast lesions. In one specific embodiment, the multi-scale Retinex-CLAHE fusion algorithm can first separate the reflectance components of the image using a multi-scale Retinex model, and then perform a weighted fusion of these components with the CLAHE processing results. For example, the multi-scale Retinex-CLAHE fusion algorithm can include, but is not limited to, multi-scale Retinex-CLAHE fusion based on Gaussian pyramids, Retinex-CLAHE fusion based on wavelet domain reflectance component extraction, and Retinex reflectance component estimation and CLAHE fusion based on guided filtering.

[0070] This algorithm decomposes the reflectance component of denoised medical images using a multi-scale Retinex-CLAHE fusion method and then fuses it with the results of local contrast adaptive histogram equalization to enhance the contrast of lesion areas. This can be achieved by first extracting the reflectance component reflecting the tissue essence through multi-scale Retinex, and then fusing it with the CLAHE enhancement result. Furthermore, this operation can involve Gaussian scale spatial decomposition of the Retinex reflectance component followed by a weighted summation with the CLAHE output, or by using the Retinex reflectance map as input to CLAHE to guide the local contrast enhancement area, thereby strengthening the grayscale difference between the lesion and surrounding normal tissue, significantly improving lesion visibility, especially in low-contrast areas.

[0071] For example, in the scenario of multimodal early stroke lesion detection, the deep learning-based medical image lesion annotation method in this embodiment can acquire dual-modal images of the patient's CT and DWI-MRI. First, an adaptive window width and window level algorithm is used to optimize the display dynamic range of the two modalities, so that the ischemic area is in the optimal visual range in each modality. Then, an improved nonlocal mean filtering combined with denoising is used to suppress CT metal artifacts and MRI motion noise while preserving the edges of small infarcts. Finally, a multi-scale Retinex-CLAHE fusion algorithm is applied to enhance the lesion reflectivity, so that the early low-density infarct area on CT and the high signal area on MRI both achieve significant contrast improvement, generating high-quality medical enhanced images for subsequent use by the U-Net segmenter, thereby improving the detection rate of lesions in the hyperacute phase.

[0072] In one embodiment, the reflectance component of the denoised medical image is decomposed based on the multi-scale Retinex-CLAHE fusion algorithm, and the reflectance component is fused with the local contrast adaptive histogram equalization result to enhance the contrast of the lesion area, resulting in a medically enhanced image, including:

[0073] The original medical image is converted to the HSV color space, the gradient magnitude map of the original medical image in the saturation channel is calculated, and the gradient magnitude map is binarized by the adaptive Otsu threshold segmentation algorithm to obtain the initial lesion candidate map.

[0074] The adaptive Otsu thresholding algorithm can be an Otsu automatic thresholding method that dynamically adjusts the inter-class variance calculation window based on the local gray-level distribution of the image. It can be used to adaptively segment high-response regions on the gradient magnitude map as initial lesion candidates, improving sensitivity to low-contrast lesions. In an exemplary embodiment, the adaptive Otsu thresholding algorithm can calculate the local inter-class variance within a sliding window, select the threshold that maximizes the variance as the locally optimal segmentation point, and then fuse global statistical information for smoothing correction. Exemplarily, the adaptive Otsu thresholding algorithm can include, but is not limited to, one or more of the following: an adaptive Otsu algorithm based on local entropy constraints, an Otsu variant based on multi-scale histogram fusion, and an improved Otsu algorithm based on gradient direction consistency.

[0075] The maximum inter-class variance thresholds of CT bone tissue and MRI soft tissue in the original medical images are integrated to generate a multimodal binary map. The initial lesion candidate map is then fused with the multimodal binary map and artifact regions are removed to obtain an optimized lesion label map.

[0076] The multimodal binary map can be a joint binary mask generated by fusing the Otsu's maximum inter-class variance (Otsu) threshold segmentation results of CT and MRI respectively. It can be used to integrate the segmentation advantages of different modalities for bone and soft tissues, providing anatomical priors to assist in lesion region screening. In one specific embodiment, the multimodal binary map can apply Otsu's threshold segmentation to CT and MRI images separately, taking the union or weighted combination of high-density regions on CT and high-signal regions on MRI. Furthermore, the multimodal binary map can include, but is not limited to, CT-dominated multimodal binary maps, MRI-dominated multimodal binary maps, and balanced weighted multimodal binary maps.

[0077] The optimized lesion label map can be a binary mask generated by fusing an initial lesion candidate map with a multimodal binary map and removing artifacts, used to identify the real lesion region. In an exemplary embodiment, the optimized lesion label map can logically intersect or weightedly fuse the initial lesion candidate map obtained based on the saturation gradient magnitude map through an adaptive Otsu threshold, with a multimodal binary map generated by fusing CT bone tissue and MRI soft tissue maximum inter-class variance thresholds, and remove artifacts such as isolated small connected components. For example, the optimized lesion label map can be one or more of the following: lesion label map based on morphological constraints, lesion label map based on multimodal confidence weighting, and lesion label map based on edge consistency verification.

[0078] The initial lesion candidate image is fused with the multimodal binary image, and artifact regions are removed to obtain an optimized lesion label image. This can be achieved by performing a logical AND and / or weighted fusion on the initial lesion candidate image and the multimodal binary image, and removing connected regions that are too small or have abnormal shapes. Furthermore, this operation can be achieved by using a logical AND operation to retain regions supported by dual evidence, and then filtering out minor artifacts using an area threshold; or by weighted fusion of the two images based on multimodal confidence, and then using morphological opening operations to remove isolated points. This suppresses false positive responses caused by noise or non-lesion structures such as blood vessels, thereby improving the accuracy of lesion region localization.

[0079] The optimized lesion label map is used to perform morphological closing operations on the original medical images to capture the reflection components of the microscopic lesion edges, mesoscopic tissue structures and macroscopic background in the closing operation area.

[0080] The reflection components (microscopic lesion edges, mesoscopic tissue structures, and macroscopic background) can be illumination-invariant components reflecting the essential characteristics of tissues, decomposed from medical images at different spatial scales. These components correspond to lesion edge details, local anatomical structures, and overall background distribution, respectively, and can be used to provide structural awareness for multi-scale contrast enhancement, ensuring that the enhancement results balance detail preservation and overall consistency. In a specific embodiment, the reflection components may include, but are not limited to, microscopic lesion edge reflection components, mesoscopic tissue structure reflection components, and macroscopic background reflection components.

[0081] Morphological closing operations are performed on the original medical images using optimized lesion label maps. Within the closing operation region, the reflectance components of microscopic lesion edges, mesoscopic tissue structures, and macroscopic background are captured. This can be achieved by using the optimized label map as a mask, performing multi-scale structuring element closing operations on the original image within its coverage area, and extracting reflectance components at different scales. Furthermore, this operation can be achieved by sequentially performing closing operations using circular structuring elements of different sizes, corresponding to microscopic, mesoscopic, and macroscopic scales respectively, or by applying multi-scale guided filtering within the label map region to estimate reflectance components at each scale. This allows for structure-aware reflectance component decomposition under lesion semantic guidance, avoiding distortion of non-lesion areas caused by global processing.

[0082] Based on a weighted fusion strategy, the reflection components of the microscopic lesion edge, mesoscopic tissue structure, and macroscopic background are integrated to form a medical enhancement image with enhanced lesion contrast.

[0083] This approach integrates reflectance components from microscopic lesion edges, mesoscopic tissue structures, and macroscopic background using a weighted fusion strategy to create contrast-enhanced medical images. This can be achieved by assigning weights based on the local contrast or gradient energy of reflectance components at each scale, and then linearly or non-linearly fusing them to generate the final enhanced image. Furthermore, this operation can be implemented by dynamically assigning weights to microscopic components based on local standard deviations, fusing mesoscopic and macroscopic components at a fixed ratio, or using an attention mechanism to learn the importance of reflectance components at each scale and then weighting and synthesizing them. This allows for adaptive contrast enhancement of the lesion region while preserving edge sharpness and overall structural consistency.

[0084] Taking multimodal image enhancement for liver cancer as an example, the deep learning-based medical image lesion annotation method in this embodiment can be as follows: First, the images are converted to HSV space, the saturation channel gradient magnitude map is calculated, and an initial lesion candidate is generated through an adaptive Otsu algorithm. At the same time, Otsu thresholding is applied to CT (highlighting calcification / hemorrhagic) and MRI (highlighting edema / necrosis) to obtain multimodal binary maps. After the two are fused, connected components smaller than 50 pixels are removed to form an optimized lesion label map. Using this label map as a mask, multi-scale closing operations are performed on the original image within the liver parenchyma region to extract the reflection components of the tumor edge, internal heterogeneous structures, and surrounding liver background. Finally, the three-scale reflection components are fused according to the edge gradient intensity to generate a medically enhanced image with significantly enhanced lesion contrast and no background overexposure, which is used by the subsequent segmentation model to accurately identify microsatellite lesions.

[0085] In one embodiment, referring to Figure 2, a deep convolutional network segmenter is constructed to segment the lesion region, and a lesion annotation mask image is generated based on the lesion segmentation result, including:

[0086] Step S201: Construct a deep convolutional network segmenter and train the deep convolutional network segmenter end-to-end based on a predefined medical image training set. Use the trained deep convolutional network segmenter to output a probability heatmap of the lesion region.

[0087] The probability heatmap can be a spatial distribution map representing the probability of lesion presence, used as a guiding signal for the channel attention module. In this embodiment, the probability heatmap and the channel attention module work together; the heatmap provides the lesion confidence distribution, driving the attention mechanism to reweight the high-dimensional feature map.

[0088] Step S202: Use the encoder of the deep convolutional network segmenter as the encoder of the improved U-Net model, and input the medical enhanced image into the encoder, and output a high-dimensional feature map at the end of the encoder;

[0089] The improved U-Net model can be a lesion segmentation network that introduces a channel attention mechanism and a multi-scale feature fusion structure on the basis of the standard U-Net architecture, in order to improve the semantic focusing ability and boundary reconstruction accuracy of the lesion region. In an exemplary embodiment, the improved U-Net model reuses the encoder of a pre-trained deep convolutional network segmenter as the backbone, and embeds a channel attention module and a multi-scale feature fusion module in its decoding path. The high-dimensional feature map can be a multi-channel feature representation containing rich semantic information, used to carry the deep semantic features of the lesion. Furthermore, the high-dimensional feature map and the channel weight vector work together, and the feature focusing of the lesion region is achieved after channel-by-channel multiplication.

[0090] Step S203: Map the probability heatmap to a channel weight vector through the channel attention module of the improved U-Net model, and multiply the channel weight vector with the high-dimensional feature map channel by channel to focus on the lesion area of ​​the feature map;

[0091] The channel attention module can be a neural network substructure that dynamically generates channel weights based on the importance of input features, used to strengthen feature channels related to lesions and suppress irrelevant background channels. For example, the channel attention module receives a probability heatmap as a guiding signal and generates channel weight vectors through global pooling, fully connected layers, or convolutional layers. The channel attention module can be one or more of, including but not limited to, SE attention modules, CBAM channel branches, and ECA attention modules. Mapping the probability heatmap to a channel weight vector through the channel attention module of the improved U-Net model can be achieved by taking the probability heatmap as input, calculating the importance score of each channel through the channel attention module, and outputting a normalized weight vector. Furthermore, this operation can be performed by globally averaging the probability heatmap and then inputting it into a two-layer fully connected network to generate weights, thereby establishing an explicit correlation between lesion confidence and feature channel responses, achieving semantically guided feature selection.

[0092] Multiplying the channel weight vectors channel-by-channel with the high-dimensional feature map to focus on the lesion region of the feature map can be achieved by broadcasting the weight vectors to every spatial location of the high-dimensional feature map and performing a channel-by-channel scaling operation. In one specific embodiment, this operation can be implemented through Hadamard product to enhance the response intensity of lesion-related channels, suppress interfering features in non-lesion regions, and improve feature discriminability.

[0093] Step S204: In the decoder stage of the improved U-Net model, a multi-scale feature fusion module is introduced to gradually restore the lesion region resolution of the feature map and obtain the segmented lesion annotation mask image.

[0094] The multi-scale feature fusion module can be a network component that integrates feature maps from different levels of the encoder to recover high-resolution details, used to fuse high-level semantic information with low-level spatial details. In this embodiment, the multi-scale feature fusion module aggregates multi-level features through upsampling, concatenation, or weighted fusion, gradually reconstructing the fine structure in the decoder. The multi-scale feature fusion module can include, but is not limited to, FPN-style feature pyramid fusion, DeepLabASPP multi-scale context fusion, etc. Introducing a multi-scale feature fusion module into the decoder stage of the improved U-Net model to gradually restore the resolution of the lesion region of the feature map can be achieved by fusing low-level features from the corresponding encoder level with the current high-level features at each upsampling stage of the decoder. Furthermore, this operation can use skip connections to concatenate multi-scale features and then unify the number of channels through convolution, thereby preserving fine-grained boundary information and reducing structural distortion and blurring caused by multiple upsampling.

[0095] Taking precise preoperative segmentation of liver cancer as an example, the deep learning-based medical image lesion annotation method in this embodiment can be as follows: after denoising and contrast enhancement of the patient's CT and MRI multimodal images, the images are input into a pre-trained segmenter to generate a liver tumor probability heatmap; this heatmap guides the channel attention module of the improved U-Net to generate weights and focus on the lesions in the high-dimensional feature map output by the encoder; in the decoding stage, the multi-scale feature fusion module integrates shallow edge details and deep semantics to reconstruct a tumor mask with clear boundaries; after post-processing, this mask is sent to the sub-pixel annotation model to output high-precision tumor contour coordinates that can be used for laparoscopic surgical navigation.

[0096] In one embodiment, the multi-scale feature fusion module includes several dilated convolutional layers with different receptive fields, used to capture multi-scale lesion structural features from micro to macro on the same level feature map;

[0097] The multi-scale feature fusion module can be a feature extraction unit composed of multiple dilated convolutional layers with different receptive fields in parallel, used to capture multi-granularity lesion structure information at the same level. In this embodiment, the multi-scale feature fusion module can simultaneously model micro-boundaries, meso-transitions, and macro-morphology, improving the geometric consistency and detail fidelity of the segmentation results. Furthermore, the multi-scale feature fusion module can deploy dilated convolutions with different dilation rates in parallel on a single feature map, extracting local, regional, and global contextual features before fusion. The dilated convolutional layers with different receptive fields can employ dilated convolution operations with different dilation rates to expand the receptive field without increasing the number of parameters. For example, dilated convolutional layers with different receptive fields can achieve multi-scale context awareness of the same input feature map, avoiding information loss due to a fixed receptive field. In an exemplary embodiment, dilated convolutional layers with different receptive fields can include, but are not limited to, dilated convolutional layers with small receptive fields, dilated convolutional layers with medium receptive fields, and dilated convolutional layers with large receptive fields.

[0098] The dilated convolutional layer with a small receptive field can employ a low dilation rate (e.g., 1 or 2) to focus on the local pixel neighborhood. In one specific embodiment, the dilated convolutional layer with a small receptive field can preserve lesion cell-level boundary details, supporting high-precision edge reconstruction. Furthermore, the dilated convolutional layer with a small receptive field can collaborate with medium and large receptive field layers to provide fine structural information and compensate for the spatial ambiguity of the large receptive field layer. The dilated convolutional layer with a medium receptive field can employ a dilation rate (e.g., 4 or 6) to cover the boundary region between the lesion and adjacent tissue. In this embodiment, the dilated convolutional layer with a medium receptive field can enhance the lesion-background interface discrimination ability and suppress artifacts and missegmentation. The dilated convolutional layer with a medium receptive field can connect the small and large receptive field layers, establishing a transitional representation between local details and global semantics. The dilated convolutional layer with a large receptive field can employ a high dilation rate (e.g., 8 or 12) to cover the entire lesion and its distal context. Furthermore, dilated convolutional layers with large receptive fields can model the overall morphology and spatial distribution patterns of lesions, improving the structural rationality of the segmentation results. In an exemplary embodiment, dilated convolutional layers with large receptive fields can provide global constraints for small and medium receptive field layers, preventing local overfitting or fragmented segmentation.

[0099] Capturing multi-scale lesion structure features from micro to macro on the same level feature map can be achieved by feeding the same input feature map in parallel into dilated convolutional layers with different dilation rates, extracting multi-granularity contextual information separately. Furthermore, capturing multi-scale lesion structure features from micro to macro on the same level feature map can be achieved by using parallel dilated convolutional branches followed by concatenating channel dimensions or by weighted summation of the outputs of each branch to generate unified multi-scale features. This avoids relying on cross-level skip connections to pass details, achieving multi-scale context fusion within the same layer and improving the completeness of feature representation.

[0100] Small receptive field dilated convolutional layers focus on the boundary details of individual lesion cells, which can be achieved by using low-dilation dilated convolutions to extract high-frequency edge responses within the local neighborhood. Furthermore, this focus on the boundary details of individual lesion cells can be achieved through standard convolutions with a dilation rate of 1 or dilated convolutions with a dilation rate of 2 combined with edge enhancement loss, thus preserving sub-pixel-level resolvable boundary cues and providing a geometric basis for subsequent sub-pixel annotation. Medium receptive field dilated convolutional layers perceive the interface transition region between the lesion and surrounding tissue, which can be achieved by using convolutional kernels with a moderate dilation rate to cover a range extending several millimeters beyond the lesion edge, capturing tissue contrast changes. In a specific embodiment, the medium receptive field dilated convolutional layer can perceive the interface transition region between the lesion and surrounding tissue through 3×3 dilated convolutions with a dilation rate of 4 or by combining medium receptive field feature extraction with local contrast normalization, thereby enhancing lesion-normal tissue boundary discrimination and reducing missed or overflowed classifications due to low contrast. Large receptive field dilated convolutional layers can be used to analyze the overall morphology and spatial distribution of lesions. This can be achieved by using high-expansion-rate convolutional kernels to cover the entire lesion area and even adjacent anatomical structures, extracting global shape priors. Furthermore, large receptive field dilated convolutional layers can be used to analyze the overall morphology and spatial distribution of lesions by combining dilated convolutions with an expansion rate of 12 with global pooling or multi-branch large receptive field convolutions with integrated morphological attention. This can constrain the segmentation results to conform to anatomical rationality and avoid isolated fragments or non-physiological morphologies.

[0101] Taking the precise segmentation of glioblastoma boundaries as an example, the deep learning-based medical imaging lesion annotation method in this embodiment can be to input the feature map output by the improved U-Net encoder into a multi-scale feature fusion module after preprocessing of enhanced MRI images; wherein the small receptive field hollow convolution preserves the fine enhancement structure of the tumor infiltration edge, the medium receptive field layer distinguishes the transition zone between the tumor core area and the edema zone, and the large receptive field layer ensures the spatial coherence of the entire tumor mass; the features after the fusion of the three are reconstructed by the decoder to generate a mask with clear boundaries and reasonable shape, providing a reliable target area for neurosurgical navigation.

[0102] In one embodiment, referring to Figure 3, post-processing techniques are used to optimize the lesion annotation mask image, and the optimized mask image is input into a sub-pixel level annotation generation model. The sub-pixel level annotation generation model outputs a high-precision lesion annotation coordinate sequence, including:

[0103] Step 301: Perform distance transformation on the lesion-annotated mask image to generate a distance field map, and extract the main edge components of the lesion based on the gradient distribution in the distance field map;

[0104] The distance field map can be a scalar field image generated from the lesion-annotated mask image through distance transformation. Each pixel value represents its Euclidean distance to the nearest background point, which can be used to provide boundary gradient information in continuous space, supporting edge component extraction and subsequent sub-pixel localization. In this embodiment, the distance field map can be generated with a smooth gradient distribution by performing an Euclidean distance transformation on the binary mask. For example, the distance field map can be one or more of the following: a signed distance field map, a weighted distance field map, and an anisotropic distance field map. Generating the distance field map by performing a distance transformation on the lesion-annotated mask image can be achieved by performing an Euclidean distance transformation on the binary mask image, outputting the distance value from each foreground pixel to the nearest background boundary. Furthermore, this operation can be achieved by using the Fast Marching Model (FMM) to achieve efficient distance transformation or by using anisotropic weighted distance transformation to adapt to multimodal tissue characteristics, thereby transforming discrete boundaries into a continuous distance field, providing a mathematical basis for gradient analysis and sub-pixel localization.

[0105] Step 302: Set the contour refinement search range based on the main edge components of the lesion, and use the active contour model to refine the lesion boundary. At the same time, suppress artifact interference during the refinement process to obtain the optimized lesion boundary.

[0106] The active contour model can be a deformable curve model based on energy minimization, used to automatically fit target boundaries in an image. It can refine lesion boundaries within a limited search range while suppressing segmentation artifacts and topological noise. In an exemplary embodiment, the active contour model can combine external image forces (such as gradients) with internal smoothing constraints to iteratively evolve to the optimal boundary near the initial contour. For example, the active contour model can be one or more of the Snake model, GeodesicActiveContour model, and LevelSet active contour model. Refining lesion boundaries using the active contour model while suppressing artifact interference can be achieved by using the main edge components as the initial contour, driving the active contour model evolution within a limited search range, and combining regularization terms to suppress abnormal deformation. Furthermore, this operation can be achieved by using a level set-based active contour model to handle complex topological changes or by introducing an adversarial artifact loss term to constrain the contour evolution direction, thereby repairing segmentation defects such as breaks and jagged edges, and obtaining optimized lesion boundaries that are topologically correct and geometrically smooth.

[0107] Step 303: Construct an adaptive edge filter in the distance field map based on the optimized lesion boundary, and use the adaptive edge filter to generate a smooth lesion boundary image;

[0108] The adaptive edge filter can be an edge smoothing filter that dynamically adjusts its parameters based on the geometric features of the local lesion boundary. It can be used to suppress noise and artifacts while preserving the true boundary structure, generating a geometrically consistent smooth boundary image. In one specific embodiment, the adaptive edge filter can construct a local response function based on optimizing the curvature and gradient direction of the lesion boundary in the range field map. For example, the adaptive edge filter can be one or more of the following: a curvature adaptive Gaussian filter, a gradient-guided bilateral filter, and a directional median filter. Constructing the adaptive edge filter based on the optimized lesion boundary in the range field map and generating a smooth lesion boundary image using the adaptive edge filter can be achieved by designing a filter kernel based on the local geometric properties (such as curvature and gradient) of the optimized boundary in the range field and performing convolutional smoothing on the boundary image. Furthermore, this operation can be achieved by adaptively adjusting the standard deviation of the Gaussian kernel based on curvature or by constructing a guided filter using the gradient direction to maintain edge smoothness, thereby eliminating high-frequency noise and improving the geometric consistency of the boundary image while preserving the true anatomical structure.

[0109] Step 304: Project the lesion boundary image radially along the optimized lesion boundary to generate a boundary intensity curve, and perform sub-pixel level peak detection on the boundary intensity curve using B-spline interpolation to obtain the lesion center coordinate sequence;

[0110] The lesion center coordinate sequence can be a continuous set of center point coordinates along the radial projection direction of the lesion obtained through sub-pixel level peak detection. This set can be used as an intermediate geometric representation input to the sub-pixel level annotation generation model to calculate the final high-precision annotation parameters. For example, the lesion center coordinate sequence can be one or more of the following: centroid trajectory sequence, skeleton principal axis coordinate sequence, radial symmetry center sequence, etc. The lesion center coordinate sequence is obtained by performing sub-pixel level peak detection on the boundary intensity curve using B-spline interpolation. This can be achieved by fitting a B-spline function to the radially projected boundary intensity curve and locating the sub-pixel level peak at its derivative zero point. Furthermore, this operation can be improved in robustness by using cubic B-splines for global fitting and then calculating the derivative to locate extreme points, or by combining piecewise B-spline fitting with local maximum constraints. This allows overcoming pixel grid limitations and obtaining continuous, high-resolution lesion center trajectories.

[0111] Step 305: Input the lesion center coordinate sequence into the sub-pixel level annotation generation model to calculate the lesion annotation parameters to obtain a high-precision lesion annotation coordinate sequence, and compare the lesion annotation parameters with the medical standard library. Optimize the sub-pixel level annotation generation model based on the comparison results.

[0112] The lesion center coordinate sequence is input into a sub-pixel level annotation generation model to calculate lesion annotation parameters, resulting in a high-precision lesion annotation coordinate sequence. This can be achieved by using the lesion center coordinate sequence as input to drive the sub-pixel level annotation generation model to regress or optimize the final annotation parameters (such as boundary points, volume, major axis, etc.). Furthermore, this operation can be achieved by using a graph neural network to aggregate contextual information from the center sequence to predict boundary points, or by using the center sequence as initial conditions to drive a parameterized deformation model to output the final annotation. This integrates geometric priors with model inference, outputting a high-precision coordinate sequence that meets clinical standards.

[0113] Taking the quantitative assessment of stroke lesions as an example, the deep learning-based medical image lesion annotation method in this embodiment can be as follows: after obtaining the initial infarct mask through MRI-DWI image segmentation, a distance transformation is performed to generate a distance field map; the main edges are extracted based on gradient distribution, and the GeodesicActiveContour model is activated within the narrow search zone to refine the boundaries, effectively removing artifacts caused by partial volume effects; subsequently, an adaptive Gaussian filter is constructed based on the curvature of the refined boundary in the distance field to generate a smooth boundary image; the intensity curve is obtained by radial projection of the image, and the subpixel-level center trajectory is obtained through cubic B-spline interpolation; the trajectory is input into the subpixel-level annotation generation model to output the precise boundary coordinates of the infarct core, and compared with the stroke image standard library to dynamically calibrate the model output, ensuring that the volume measurement error is below the clinically acceptable threshold.

[0114] In one embodiment, the expression for the adaptive edge filter is:

[0115]

[0116] In the formula, The functional form representing the adaptive edge filter; Represents the original spatial domain coordinates; Indicates the principal direction angle of the filter; Indicates the center distance of the filter; This represents the standard deviation of a Gaussian function in the spatial domain. This indicates that the original spatial domain coordinates are rotated around the center of the lesion. The new coordinates.

[0117] The principal direction angle of the filter can be the dominant direction of the adaptive edge filter at the local lesion boundary, used to align the filter response axis with the true edge direction. In an exemplary embodiment, the principal direction angle of the filter can guide the filter to preserve structure in the direction parallel to the edge and suppress noise in the vertical direction, improving the directional smoothing effect. Further, the principal direction angle of the filter can be, but is not limited to, one or more of the following: direction angle based on gradient principal component analysis, direction angle based on local curvature extrema, and direction angle based on skeleton tangent estimation. The standard deviation of the spatial domain Gaussian function can be a scale parameter controlling the smoothing intensity of the adaptive edge filter in the spatial domain. For example, the standard deviation of the spatial domain Gaussian function can adjust the smoothing degree of the filter in different regions to achieve local adaptive denoising. In a specific embodiment, the standard deviation of the spatial domain Gaussian function can be, but is not limited to, one or more of the following: fixed global standard deviation, standard deviation based on distance field amplitude modulation, and standard deviation based on local signal-to-noise ratio estimation.

[0118] The new coordinates rotated around the lesion center can be obtained by rotating the original spatial coordinate system with the lesion's geometric center as the origin. In this embodiment, the new coordinates rotated around the lesion center can align the filter's principal axis with the local edge of the lesion, supporting direction-sensitive filtering operations. Furthermore, the new coordinates rotated around the lesion center can be reoriented using a two-dimensional rotation transformation matrix based on the principal direction angle. Rotating the original spatial domain coordinates around the lesion center can be achieved by using the lesion center as the rotation center and performing a two-dimensional rotation transformation on each pixel coordinate according to the principal direction angle. Further, rotating the original spatial domain coordinates around the lesion center can be achieved by using an affine transformation matrix to perform batch coordinate rotation, or by directly adjusting the angle offset in polar coordinates, thereby establishing a local coordinate system aligned with the local geometry of the lesion, providing a basis for directional filtering.

[0119] Constructing an adaptive edge filter with the principal direction angle and center distance as parameters can be achieved by using these parameters as control parameters and defining the filter's spatial response function in conjunction with the rotated coordinates. Furthermore, this adaptive edge filter can be constructed by defining an anisotropic Gaussian kernel in the rotated coordinate system, or by mapping the center distance to the filter's support domain radius to limit its effective range. This generates a filter kernel with direction selectivity and positional adaptability, accurately matching the geometric characteristics of the lesion boundary. Directional smoothing of the lesion boundary based on the standard deviation of the spatial domain Gaussian function can be achieved by applying a larger standard deviation along the filter's principal direction to preserve structure, and a smaller standard deviation in the vertical direction to suppress noise. Further, directional smoothing of the lesion boundary based on the standard deviation of the spatial domain Gaussian function can be achieved by using an elliptical Gaussian kernel to achieve directional differential smoothing, or by dynamically adjusting the standard deviation to match local boundary curvature changes. This effectively removes artifacts and jagged edges while maintaining edge sharpness, improving the geometric consistency of the boundary image.

[0120] For example, in the scenario of refining the boundary of liver tumors, the deep learning-based medical image lesion annotation method in this embodiment can be as follows: after the initial segmentation mask is refined by an active contour model, the lesion center is extracted and the principal direction angle of each boundary point is calculated (e.g., through local gradient PCA); the original coordinates are rotated around the center to align the x-axis with the principal direction; under this new coordinate system, an elliptic Gaussian filter is constructed with the center distance as the radial weight and the principal direction as the major axis, and its standard deviation is set to a larger value in the tangential direction to preserve the contour continuity and a smaller value in the normal direction to suppress spurs; after this directional smoothing, the boundary image significantly reduces the step effect, providing high-quality input for subsequent B-spline interpolation, and finally outputting sub-pixel-level tumor major axis coordinates that conform to the RECIST standard.

[0121] In one embodiment, suppressing artifact interference during the refinement process to obtain optimized lesion boundaries includes:

[0122] Morphological opening operations are performed on the distance field map to remove isolated artifacts, generating preliminary continuous lesion boundaries. The angles corresponding to the gradient peaks in the preliminary continuous lesion boundaries are extracted as a coarse estimation reference for the lesion boundaries.

[0123] The angle corresponding to the gradient peak can be the directional angle corresponding to the point of maximum radial gradient magnitude in the distance field map. This angle can be used to provide a rough directional estimate of the lesion boundary to guide subsequent energy threshold setting and contour initialization. In an exemplary embodiment, the angle corresponding to the gradient peak can be described in context as the principal gradient directional angle obtained by analyzing the local gradient distribution of the distance field map. For example, the angle corresponding to the gradient peak can be one or more of the following: local normal angle, principal gradient directional angle, and radial extremum angle. Morphological opening operations are performed on the distance field map to remove isolated artifacts and generate preliminary continuous lesion boundaries. This can be achieved by sequentially performing erosion and dilation operations to eliminate small-scale isolated high-value regions. Furthermore, this operation can be implemented by using an opening operation whose structuring element size is adapted to the lesion scale, thereby removing spurious local maxima caused by segmentation noise or partial volume effects and preserving the main lesion structure.

[0124] Based on a coarse estimation reference and combined with the average size of lesions in current medical images, a contour energy threshold is set to filter out artifact regions below the contour energy threshold.

[0125] The contour energy threshold can be an energy criterion set based on the average lesion size and image gradient information to distinguish between true boundaries and artifacts. It can be used to filter low-response artifact regions and retain anatomically significant high-energy boundary components. In a specific embodiment, the contour energy threshold can be described in context as dynamically calculated by combining the statistical size prior of similar lesions in the current medical image with the gradient magnitude of the distance field map. Furthermore, the contour energy threshold can include, but is not limited to, one or more of the following: an energy threshold based on area normalization, an energy threshold based on local contrast weighting, and an energy threshold based on multimodal fusion response.

[0126] Based on a coarse estimate and combined with the average lesion size in current medical images, a contour energy threshold is set to filter out artifact regions below the contour energy threshold. This can be achieved by determining the search direction based on the gradient peak angle and normalizing the gradient energy according to the average lesion size, thus setting a dynamic threshold. For example, this operation can be implemented by using the lesion volume prior to infer the equivalent diameter to set the energy threshold, thereby physically distinguishing true edges from low-energy interference and improving the semantic accuracy of boundary extraction.

[0127] The initial continuous lesion boundaries are clustered according to curvature to generate the initial lesion boundary contours. The smoothness of the initial lesion boundary contours is verified, and the cluster centers are adjusted based on the verification results until they meet the medical anatomical constraints.

[0128] The initial lesion boundary contour can be a candidate boundary representation that meets basic anatomical rationality, generated after curvature clustering and smoothness verification. It can serve as a benchmark structure for fracture repair and accuracy assessment, supporting the final optimized boundary generation. In an exemplary embodiment, the initial lesion boundary contour can be one or more of the following: closed spline contour, piecewise circular arc fitting contour, and skeleton-guided contour. Medical anatomical constraints can be prior knowledge rules reflecting the reasonable range of human organs or lesions in terms of morphology, curvature, and connectivity. These can be used to constrain the boundary optimization process, ensuring that the generated contour conforms to clinical anatomical rationality. Furthermore, medical anatomical constraints can be described in context as statistically summarizing typical lesion geometric features from standard medical atlases or large-scale labeled datasets. For example, medical anatomical constraints can be one or more of the following: curvature continuity constraints, topological simple connectivity constraints, and organ spatial location constraints.

[0129] The initial continuous lesion boundaries are clustered according to curvature to generate initial lesion boundary contours. The smoothness of these initial contours is then verified. Based on the verification results, the cluster centers are adjusted until they meet medical anatomical constraints. This can be achieved by clustering boundary points according to curvature features, fitting the initial contour, and verifying smoothness using indicators such as the rate of change of curvature and the number of inflection points, iteratively adjusting the clustering parameters. In one specific embodiment, this operation can be achieved by introducing an anatomical constraint loss function to guide the optimization of cluster centers, thereby ensuring that the contour maintains both mathematical smoothness and conforms to the morphological priors of the anatomical structure, avoiding overfitting noise.

[0130] The broken boundaries are sequentially interpolated and repaired according to the initial lesion boundary contour that meets medical anatomical constraints. The accuracy of the lesion boundary is evaluated based on the repair results to obtain the final optimized lesion boundary.

[0131] The broken boundaries are sequentially interpolated and repaired according to the initial lesion boundary contour that meets medical anatomical constraints. The accuracy of the lesion boundary is then evaluated based on the repair results to obtain the final optimized lesion boundary. This can be achieved by identifying the contour break points and performing spline or circular arc interpolation based on the nearby curvature and anatomical orientation, and calculating the geometric consistency index before and after repair. Furthermore, this operation can be achieved through directional interpolation based on the local Frenet coordinate system, thereby restoring boundary integrity. At the same time, accuracy evaluation ensures that the repair results are reliable and clinically credible.

[0132] Taking the refinement of liver tumor boundaries as an example, the deep learning-based medical imaging lesion annotation method in this embodiment can generate a distance field map based on the CT enhancement-phase segmentation mask, remove intrahepatic vascular artifacts through morphological opening operations, extract the gradient main peak angle as a reference for the tumor expansion direction, set the contour energy threshold in combination with the typical diameter of liver cancer (e.g., 3cm), and filter out small false positive areas; generate an initial contour by clustering the remaining boundary points according to curvature, and if the verification finds that the curvature mutation exceeds the allowable range of liver parenchyma, adjust the cluster center; then, for boundary breaks caused by low contrast, perform B-spline interpolation repair according to anatomical constraints, and calculate the Hausdorff distance between the repaired contour and the typical liver tumor shape as an accuracy evaluation index, and finally output an optimized boundary that meets the diagnostic standards of radiology.

[0133] In one embodiment, the lesion boundary image is radially projected along the optimized lesion boundary to generate a boundary intensity curve. Subpixel-level peak detection is then performed on the boundary intensity curve using B-spline interpolation to obtain the lesion center coordinate sequence, including:

[0134] Radial projection is performed on the lesion boundary image according to the normal direction of the optimized lesion boundary to obtain a one-dimensional boundary intensity curve. Dense interpolation points are generated at the discrete sampling points of the one-dimensional boundary intensity curve according to the sub-pixel interval.

[0135] The normal direction can be a unit vector direction perpendicular to the local tangent at each point on the optimized lesion boundary. It can be used to define the sampling path for radial projection, ensuring the projection direction is orthogonal to the lesion geometry and improving positioning accuracy. In an exemplary embodiment, the normal direction can include, but is not limited to, one or more of the following: inner normal direction, outer normal direction, and average curvature normal direction. The one-dimensional boundary intensity curve can be an intensity distribution function obtained by radially projecting the lesion boundary image along the normal direction of the optimized lesion boundary. It can be used to compress two-dimensional boundary information into a one-dimensional signal, facilitating sub-pixel level peak detection and analysis. For example, the one-dimensional boundary intensity curve can include, but is not limited to, grayscale projection curves, gradient magnitude projection curves, and multi-channel weighted projection curves. The one-dimensional boundary intensity curve is obtained by radially projecting the lesion boundary image along the normal direction of the optimized lesion boundary. This can be achieved by sampling the image intensity along its normal direction at each point on the optimized lesion boundary, forming a one-dimensional function. Furthermore, this operation can be achieved by sampling along the bidirectional normal (inside and outside) and taking the maximum response to generate an intensity curve, or by using a weighted window to integrate along the normal to enhance robustness, thereby transforming the two-dimensional boundary local structure into an analyzable one-dimensional signal while preserving the lesion-background transition features.

[0136] The dense interpolation points on the smoothed boundary intensity curve are fitted using a cubic B-spline function, and the first derivative of the cubic B-spline function is obtained to locate the critical point as a candidate location for the lesion boundary.

[0137] The cubic B-spline function can be a piecewise cubic polynomial function with C² continuity, used for smooth interpolation and fitting of discrete data points. It can provide a differentiable, smooth, sub-pixel-level continuous representation of the boundary intensity curve, supporting high-precision derivative analysis and extremum localization. In one specific embodiment, the cubic B-spline function can construct a combination of piecewise cubic polynomial basis functions based on dense interpolation points using least squares or interpolation constraints. Furthermore, the cubic B-spline function can include, but is not limited to, uniform cubic B-splines, non-uniform rational B-splines (NURBS), and periodic cubic B-splines. By fitting dense interpolation points on the smoothed boundary intensity curve using the cubic B-spline function and obtaining the first derivative of the cubic B-spline function, the critical point can be located as a candidate location for the lesion boundary. This can be achieved by generating dense points at sub-pixel intervals, fitting with the cubic B-spline, analytically differentiating, and setting the first derivative to zero to obtain the critical point. Furthermore, this operation can be achieved by using a node insertion strategy to improve the local conformity of the fit or by combining a regularization term to suppress overfitting noise regions, thereby enabling the localization of extreme value candidates in the continuous domain and breaking through the limitations of discrete sampling.

[0138] Calculate the second derivative at the critical point and determine the type of extremum. If the second derivative is greater than or equal to zero, it indicates that the dense interpolation point is a local minimum point and corresponds to the lesion boundary. Otherwise, it is a local maximum point and corresponds to the background area.

[0139] The second derivative at the critical point is calculated and the type of extremum is determined. Within the neighborhood of the sub-pixel-level peak, only the sub-pixel peak corresponding to the local minimum point is retained. This can be achieved by calculating the second derivative value for each critical point; if it is ≥0, it is determined to be a local minimum (corresponding to a low-intensity valley within the lesion) and retained. Further, this operation can be enhanced by introducing multi-scale second derivative verification to improve the robustness of the discrimination or by combining neighborhood consistency voting to filter isolated minimums, thereby effectively distinguishing the true boundary of the lesion (often manifested as an intensity valley) from background artifacts (often as peaks), improving the localization specificity. The sub-pixel peak can be the lesion boundary location point determined by continuous function extremum detection at sub-pixel resolution. It can be used as a high-precision geometric feature point for back-deriving the lesion center coordinate sequence. In a specific embodiment, the sub-pixel peak can be obtained by differentiating the cubic B-spline fitted curve and filtering the critical points that satisfy the local minimum condition. For example, sub-pixel peaks can include, but are not limited to, gradient minimum type sub-pixel peaks, curvature-dominated type sub-pixel peaks, and multimodal response fusion type sub-pixel peaks.

[0140] Within the neighborhood of subpixel-level peaks, only the subpixel peaks corresponding to local minima are retained, and the subpixel peaks are back-inferred to the original image coordinate system based on the optimized lesion boundary to generate a high-precision lesion center coordinate sequence.

[0141] The subpixel peak values ​​are back-calculated into the original image coordinate system based on the optimized lesion boundaries to generate a high-precision lesion center coordinate sequence. This can be achieved by using the offset of the subpixel peak values ​​in the normal direction and the coordinates of the corresponding boundary points, and then restoring their continuous coordinates in the original image through geometric mapping. Furthermore, this operation can be achieved by correcting coordinate distortion through bilinear inverse mapping or by optimizing the final coordinates using multi-view projection consistency, thereby outputting a lesion center trajectory with subpixel accuracy, supporting subsequent quantitative analysis and applications.

[0142] For example, in the scenario of defining the boundary of intervention operations in the target area, the deep learning-based medical image lesion annotation method in this embodiment can be as follows: after segmenting the enhanced MRI T1 image, obtain the optimized lesion boundary, and project the original image radially along its normal direction to generate a one-dimensional boundary intensity curve; interpolate at 0.1 pixel intervals on the curve to obtain dense points, fit with cubic B-splines and then differentiate to locate all critical points; filter out the local minimum points representing the transition valley between the target area and normal tissue through the second derivative, and remove background noise peaks; backcalculate these subpixel peaks back to the original MRI space according to their respective normal starting points and offsets to form a continuous high-precision center coordinate sequence, which is used by the surgical navigation system to accurately delineate the removal boundary and avoid missing the diffusion area or mistakenly cutting the key area.

[0143] Furthermore, to achieve the above objectives, the present invention also provides a deep learning-based medical image lesion annotation system, the system comprising: a memory, a processor, and a deep learning-based medical image lesion annotation method program stored in the memory and executable on the processor, the deep learning-based medical image lesion annotation method program being configured to implement the steps of the deep learning-based medical image lesion annotation method as described above.

[0144] 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 deep learning-based method for annotating lesions in medical images, characterized in that, The method includes: acquiring multimodal medical images using medical imaging equipment, and sequentially performing noise reduction and lesion contrast enhancement processing on the medical images using image processing techniques to obtain enhanced medical images; constructing a deep convolutional network segmenter to segment the lesion region, and generating a lesion annotation mask image based on the lesion segmentation result; optimizing the lesion annotation mask image using post-processing techniques, and inputting the optimized mask image into a sub-pixel level annotation generation model, and outputting a high-precision lesion annotation coordinate sequence through the sub-pixel level annotation generation model.

2. The deep learning-based medical image lesion annotation method as described in claim 1, characterized in that, The process involves acquiring multimodal medical images using medical imaging equipment and sequentially performing noise reduction and lesion contrast enhancement processing on the medical images using image processing techniques to obtain enhanced medical images. This includes: acquiring medical images using integrated CT and MRI dual-modal imaging equipment; optimizing the dynamic range of the images in real time using an adaptive window width and level adjustment algorithm based on the grayscale distribution feedback of the medical images to eliminate imaging noise and balance multimodal contrast differences; performing joint noise reduction processing on the medical images using an improved nonlocal mean filtering algorithm, dynamically balancing the filtering weights for local similarity and global structural fidelity; and decomposing the reflection component of the denoised medical images based on a multi-scale Retinex-CLAHE fusion algorithm, and fusing the reflection component with the local contrast adaptive histogram equalization result to enhance the contrast of the lesion area, thus obtaining enhanced medical images.

3. The deep learning-based medical image lesion annotation method as described in claim 2, characterized in that, The method involves decomposing the reflectance component of the denoised medical image using a multi-scale Retinex-CLAHE fusion algorithm and fusing the reflectance component with the local contrast adaptive histogram equalization result to enhance the contrast of the lesion area, resulting in a medically enhanced image. This includes: converting the original medical image to the HSV color space; calculating the gradient magnitude map of the original medical image in the saturation channel; binarizing the gradient magnitude map using an adaptive Otsu thresholding algorithm to obtain an initial lesion candidate image; fusing the maximum inter-class variance thresholds of CT bone tissue and MRI soft tissue in the original medical image to generate a multimodal binary image; fusing the initial lesion candidate image with the multimodal binary image and removing artifact regions to obtain an optimized lesion label map; performing morphological closing operations on the original medical image using the optimized lesion label map to capture the reflectance components of the microscopic lesion edge, mesoscopic tissue structure, and macroscopic background in the closing operation region; and integrating the reflectance components of the microscopic lesion edge, mesoscopic tissue structure, and macroscopic background based on a weighted fusion strategy to form a medically enhanced image with enhanced lesion contrast.

4. The deep learning-based medical image lesion annotation method as described in claim 1, characterized in that, The process of constructing a deep convolutional network segmenter to segment lesion regions and generating lesion-annotated mask images based on the lesion segmentation results includes: constructing a deep convolutional network segmenter and performing end-to-end training on the deep convolutional network segmenter based on a predefined medical image training set; using the trained deep convolutional network segmenter to output a probability heatmap of the lesion region; using the encoder of the deep convolutional network segmenter as the encoder of an improved U-Net model, and inputting the medical enhanced image into the encoder, outputting a high-dimensional feature map at the end of the encoder; mapping the probability heatmap to a channel weight vector through the channel attention module of the improved U-Net model, and multiplying the channel weight vector with the high-dimensional feature map channel by channel to focus on the lesion region of the feature map; introducing a multi-scale feature fusion module in the decoder stage of the improved U-Net model to gradually restore the lesion region resolution of the feature map, and obtaining the segmented lesion-annotated mask image.

5. The deep learning-based medical image lesion annotation method as described in claim 4, characterized in that, The multi-scale feature fusion module includes several dilated convolutional layers with different receptive fields, used to capture multi-scale lesion structural features from micro to macro on the same level feature map; among them, dilated convolutional layers with small receptive fields are used to focus on the boundary details of individual lesion cells; dilated convolutional layers with medium receptive fields are used to perceive the interface transition region between the lesion and the surrounding tissue; and dilated convolutional layers with large receptive fields are used to analyze the overall morphology and spatial distribution pattern of the lesion.

6. The deep learning-based medical image lesion annotation method as described in claim 1, characterized in that, The process of optimizing the lesion annotation mask image using post-processing techniques and inputting the optimized mask image into a sub-pixel-level annotation generation model to output a high-precision lesion annotation coordinate sequence includes: performing a distance transformation on the lesion annotation mask image to generate a distance field map; extracting the main edge components of the lesion based on the gradient distribution in the distance field map; setting a contour refinement search range based on the main edge components of the lesion and refining the lesion boundary using an active contour model while suppressing artifact interference during the refinement process to obtain an optimized lesion boundary; constructing an adaptive edge filter in the distance field map based on the optimized lesion boundary and generating a smooth lesion boundary image using the adaptive edge filter; radially projecting the lesion boundary image along the optimized lesion boundary to generate a boundary intensity curve, and performing sub-pixel-level peak detection on the boundary intensity curve using B-spline interpolation to obtain a lesion center coordinate sequence; inputting the lesion center coordinate sequence into the sub-pixel-level annotation generation model to calculate lesion annotation parameters to obtain a high-precision lesion annotation coordinate sequence, comparing the lesion annotation parameters with a medical standard library, and optimizing the sub-pixel-level annotation generation model based on the comparison results.

7. The deep learning-based medical image lesion annotation method as described in claim 6, characterized in that, The expression for the adaptive edge filter is: In the formula, The functional form representing the adaptive edge filter; Represents the original spatial domain coordinates; Indicates the principal direction angle of the filter; Indicates the center distance of the filter; This represents the standard deviation of a Gaussian function in the spatial domain. This indicates that the original spatial domain coordinates are rotated around the center of the lesion. The new coordinates.

8. The deep learning-based medical image lesion annotation method as described in claim 6, characterized in that, The process of suppressing artifact interference and obtaining optimized lesion boundaries during refinement includes: performing morphological opening operations on the distance field map to remove isolated artifacts, generating preliminary continuous lesion boundaries, and extracting the angle corresponding to the gradient peak in the preliminary continuous lesion boundaries as a coarse estimation reference for the lesion boundaries; setting a contour energy threshold based on the coarse estimation reference and combined with the average lesion size in the current medical images, and filtering out artifact regions below the contour energy threshold; clustering the preliminary continuous lesion boundaries according to curvature to generate initial lesion boundary contours, and verifying the smoothness of the initial lesion boundary contours, adjusting the cluster centers based on the verification results until they meet medical anatomical constraints; interpolating and repairing the broken boundaries sequentially according to the initial lesion boundary contours that meet medical anatomical constraints, and evaluating the accuracy of the lesion boundaries based on the repair results of the broken boundaries to obtain the final optimized lesion boundaries.

9. The deep learning-based medical image lesion annotation method as described in claim 8, characterized in that, The process of radially projecting the lesion boundary image along the optimized lesion boundary to generate a boundary intensity curve, and then performing sub-pixel-level peak detection on the boundary intensity curve using B-spline interpolation to obtain the lesion center coordinate sequence includes: radially projecting the lesion boundary image along the normal direction of the optimized lesion boundary to obtain a one-dimensional boundary intensity curve; generating dense interpolation points at sub-pixel intervals on the discrete sampling points of the one-dimensional boundary intensity curve; fitting the dense interpolation points on the smoothed boundary intensity curve using a cubic B-spline function, and calculating the first derivative of the cubic B-spline function to locate critical points as candidate positions of the lesion boundary; calculating the second derivative at the critical point and determining the type of extreme value; if the second derivative is greater than or equal to zero, it indicates that the dense interpolation point is a local minimum point and corresponds to the lesion boundary position; otherwise, it is a local maximum point and corresponds to the background area; retaining only the sub-pixel peak corresponding to the local minimum point in the sub-pixel-level peak neighborhood, and back-inferring the sub-pixel peak back to the original image coordinate system according to the optimized lesion boundary to generate a high-precision lesion center coordinate sequence.

10. A deep learning-based medical image lesion annotation system, characterized in that, The system includes: a memory, a processor, and a deep learning-based medical image lesion annotation method program stored in the memory and executable on the processor, the deep learning-based medical image lesion annotation method program being configured to implement the steps of the deep learning-based medical image lesion annotation method as described in any one of claims 1 to 9.

Images