Method for organ segmentation and focus identification based on CT (Computed Tomography) plain-scan image

Abstract

The invention provides an organ segmentation and focus identification method based on a CT (Computed Tomography) plain-scan image. The organ segmentation and focus identification method based on the CT plain-scan image comprises the steps that a, data preprocessing is conducted, specifically, CT image data are cut, resampled and normalized, the CT image data are cut to a non-zero value area so as to reduce the calculation burden of invalid data, the voxel pitch is unified through a third-order spline interpolation method in the resampling process, and the intensity value is standardized through a z-score formula in the normalization process; the invention provides an organ segmentation and focus identification method based on a CT (Computed Tomography) plain-scan image, which is characterized in that voxel spacing and z-score normalized standardized intensity values are unified through a data preprocessing module by adopting a third-order spline interpolation method, so that the problems of non-uniform intensity distribution and inconsistent voxel spacing of the CT image are solved from the source, and the accuracy of the CT image intensity distribution is improved. And the consistency of image input and the adaptability of the segmentation model are obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of organ segmentation and lesion recognition, and specifically provides a method for organ segmentation and lesion recognition based on CT plain scan images. Background Art

[0002] With the continuous development of Internet technology and the gradual rise of interdisciplinary fields, image segmentation technology is becoming an important part of the field of computer vision. Image segmentation is to extract regions with the same type of features (such as color, texture, gray level, etc.) in an image, so that each segmented region has the same characteristic attributes. The automatic segmentation method relying on algorithms is the current main research direction. With the development of deep learning, the image segmentation method based on deep neural network has gradually become the mainstream of research, greatly improving the segmentation accuracy and efficiency.

[0003] Medical image segmentation is an important step in medical image processing. Images are obtained by CT plain scanning of human organs and blood vessels, and key parts such as organs, blood vessels, bones, and lesions are marked on these images. Deep learning is performed using the marked images to obtain a large model of human tissues. The large model data obtained by training can be used to predict and segment human organs and lesions.

[0004] However, the existing medical image three-dimensional reconstruction technology often relies on specific CT and MRI devices, resulting in the limitation that its model can only be used under specific devices. In the case of large image sizes, the existing medical image segmentation methods usually cannot effectively identify organs and lesions due to large memory overhead and limited receptive fields. The current technical solutions may be difficult to effectively segment lesions in CT images, affecting the early detection and diagnosis of diseases. Traditional image segmentation methods may not be able to effectively process and optimize the segmentation results, resulting in inaccurate or unreasonable predictions. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a method for organ segmentation and lesion recognition based on CT plain scan images, which solves the problems that the medical image three-dimensional reconstruction technology often relies on specific devices; in the case of large image sizes, the medical image segmentation methods usually cannot effectively identify organs and lesions due to large memory overhead and limited receptive fields; it is difficult to effectively segment lesions in CT images, affecting the early detection and diagnosis of diseases; and it is unable to effectively process and optimize the segmentation results, resulting in inaccurate or unreasonable predictions.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for organ segmentation and lesion recognition based on CT plain scan images, including:

[0007] a. Data preprocessing: Crop, resample, and normalize the CT image data. The cropping is performed to the non-zero value region to reduce the computational burden of invalid data. The resampling unifies the voxel spacing through the third-order spline interpolation method. The normalization standardizes the intensity values using the z-score formula as follows:

[0008] f(x) = a i + b i (x - x i ) + c i (x - x i ) 2 + d i (x - x i ) 3

[0009]

[0010] Where:

[0011] a i = y i

[0012]

[0013] μ is the intensity mean, σ is the intensity standard deviation, b i , c i , d i are interpolation parameters. The data normalization is performed by collecting all the intensity values I that appear within the segmentation mask of the training dataset and initially normalizing the entire dataset by clipping the [0.5, 99.5] percentiles of these intensity values I 0.5,0.95 . Then, the mean μ and standard deviation of the intensity values are calculated based on I 0.5,0.95 to obtain the normalized result through z-score normalization;

[0014] b. Cascade segmentation network training: Construct a segmentation network based on the cascaded 3D U-Net, where the cascaded 3D U-Net is responsible for extracting multi-resolution features;

[0015] c. Adaptive optimization training: Improve the generalization ability of the model by dynamically adjusting the learning rate, optimizing the loss function, and expanding the data augmentation method. The loss function uses a combination of weighted Dice loss and cross-entropy loss, and the formula is as follows:

[0016] L = α · L Dice + β · L CE

[0017] The data augmentation includes random rotation, scaling, mirror flipping, and random noise addition to expand the diversity of data samples;

[0018] d. Segmentation result fusion and optimization: Optimize the segmentation results by combining the patch overlap method and the multi-model fusion strategy, and adopt the following weighted voting mechanism for fusion:

[0019]

[0020] And optimize the features of key regions through the attention mechanism;

[0021] e. Post-processing analysis: Optimize the segmentation results through connectivity analysis and morphological operations, retain the largest connected component and correct the segmentation boundary. The morphological operations include dilation, erosion, and region growing;

[0022] f. 3D reconstruction and lesion annotation: Reconstruct the 3D model of the CT image based on the segmentation results, implement voxel reconstruction using bilinear interpolation, and achieve accurate annotation of the lesion area through self-supervised learning methods combined with expert annotation;

[0023] g. Performance evaluation and model update: Evaluate the segmentation performance through indicators such as the Dice coefficient, Hausdorff distance, and average surface distance, and iteratively optimize the model by combining transfer learning methods;

[0024] Preferably, the cascaded 3DU-Net extracts low-resolution global features on the downsampled image and optimizes the fine-grained segmentation effect on the full-resolution image.

[0025] Preferably, the adaptive optimization training adopts a dynamic learning rate adjustment strategy. When the improvement of the training error in the recent 30 epochs is less than the set threshold, the learning rate is automatically reduced.

[0026] Preferably, the weighted Dice loss formula of the loss function is:

[0027]

[0028] where w i is the weight, p i is the predicted value, and g i is the ground truth.

[0029] Preferably, the segmentation result fusion optimizes the features of key regions through the channel attention module and the spatial attention module introducing the attention mechanism.

[0030] Preferably, the post-processing analysis is based on connectivity analysis, removes all regions smaller than the largest connected component, and further optimizes the boundary through the region growing algorithm.

[0031] Preferably, the three-dimensional reconstruction realizes high-precision three-dimensional model reconstruction through voxel interpolation combined with a surface fitting algorithm. The lesion annotation combines self-supervised learning and a multi-task learning framework to simultaneously predict the lesion location, volume, and type labels. The transfer learning method combines a public dataset and specific scenario data to improve the adaptability of the model under different image acquisition conditions through multiple rounds of fine-tuning.

[0032] The present invention provides a method for organ segmentation and lesion recognition based on non-contrast CT images, having the following beneficial effects:

[0033] The present invention proposes a method for organ segmentation and lesion recognition based on non-contrast CT images. Through a data preprocessing module, the voxel spacing is unified and the intensity values are standardized by z-score normalization using the third-order spline interpolation method, fundamentally solving the problems of uneven intensity distribution and inconsistent voxel spacing in CT images, and significantly improving the consistency of image input and the adaptability of the segmentation model. Through dynamic learning rate adjustment, adaptive optimization training, and various enhancement techniques, the generalization ability and robustness of the model are further improved, making it applicable to CT image data of different sizes and acquisition conditions.

[0034] In the optimization and post-processing of the segmentation results, the present invention introduces a patch overlap prediction and multi-model fusion strategy. By combining a weight voting mechanism and an attention module, the accuracy and reliability of the segmentation results are greatly improved. The post-processing module combines connectivity analysis, morphological operations, and region growing algorithms to optimize and correct the segmentation boundaries, ensuring the integrity and rationality of the segmentation results. In terms of three-dimensional reconstruction and lesion annotation, high-precision three-dimensional model reconstruction is achieved through bilinear interpolation and a surface fitting algorithm. At the same time, by combining self-supervised learning and a multi-task learning framework, the volume, position, and type of lesions are accurately annotated, meeting the clinical needs for early lesion recognition and diagnosis. Through performance evaluation and transfer learning optimization, the performance of the model is more stable under various data conditions, with broad adaptability and versatility, solving multiple problems in existing medical image segmentation and three-dimensional reconstruction technologies, and providing efficient and accurate technical support for medical image analysis and clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a schematic flow chart of the present invention;

[0036] Figure 2 is a schematic diagram of the advantages of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] As Figure 1-2 shown, the embodiment of the present invention provides a method for organ segmentation and lesion recognition based on CT plain scan images, including: a. Data preprocessing: performing cropping, resampling, and normalization operations on CT image data, where cropping is performed to the non-zero value region to reduce the computational burden of invalid data, resampling unifies the voxel spacing through the third-order spline interpolation method, and normalization standardizes the intensity values through the z-score formula. The formula is as follows:

[0039] f(x) = a i +b i (x - x i ) + c i (x - x i ) 2 +d i (x - x i ) 3

[0040]

[0041] Where:

[0042] a i = y i

[0043]

[0044]

[0045] μ is the intensity mean, σ is the intensity standard deviation, b i , c i , d i are interpolation parameters. Among them, data normalization is achieved by collecting all intensity values I that appear within the segmentation mask of the training data set and initially normalizing the entire data set by cropping the [0.5, 99.5] percentiles I 0.5,0.95 of these intensity values. Then, the average value μ and standard deviation of the intensity values are calculated based on I 0.5,0.95 , and z-score normalization is performed to obtain the normalized result.

[0046] b. Cascade segmentation network training: Construct a segmentation network based on the cascaded 3D U-Net. The cascaded 3D U-Net is responsible for extracting multi-resolution features. The cascaded 3D U-Net extracts low-resolution global features from the downsampled images.

[0047] c. Adaptive optimization training: Improve the generalization ability of the model by dynamically adjusting the learning rate, optimizing the loss function, and expanding data augmentation methods. The loss function adopts a combination of weighted Dice loss and cross-entropy loss. The formula is as follows:

[0048] L = α·L Dice +β·L CE

[0049] Data augmentation includes random rotation, scaling, mirror flipping, and random noise addition to expand the diversity of data samples. Adaptive optimization training adopts a dynamic learning rate adjustment strategy. When the improvement of the training error in the last 30 epochs is less than the set threshold, the learning rate is automatically reduced.

[0050] d. Segmentation result fusion and optimization: Optimize the segmentation results by combining the patch overlap method and the multi-model fusion strategy. The following weight voting mechanism is used for fusion:

[0051]

[0052] And optimize the features of the key regions through the attention mechanism. The segmentation result fusion optimizes the features of the key regions through the channel attention module and the spatial attention module introducing the attention mechanism.

[0053] e. Post-processing analysis: Optimize the segmentation results through connectivity analysis and morphological operations. Retain the largest connected component and correct the segmentation boundary. Morphological operations include dilation, erosion, and region growing. The weighted Dice loss formula of the loss function is:

[0054]

[0055] where w i is the weight, p i is the predicted value, g i is the ground truth value. The post-processing analysis is based on connectivity analysis, removes all regions smaller than the largest connected component, and further optimizes the boundary through the region growing algorithm.

[0056] f. 3D Reconstruction and Lesion Annotation: Based on the segmentation results, a 3D model of the CT image is reconstructed. Bilinear interpolation is used to achieve voxel reconstruction, and precise annotation of the lesion area is realized through self-supervised learning combined with expert annotation. 3D reconstruction is achieved by combining voxel interpolation with a surface fitting algorithm to reconstruct a high-precision 3D model. Lesion annotation combines self-supervised learning and a multi-task learning framework to simultaneously predict the location, volume, and type labels of the lesions. The transfer learning method combines public datasets with specific scenario data, and the model's adaptability under different image acquisition conditions is improved through multiple rounds of fine-tuning.

[0057] g. Performance Evaluation and Model Update: The segmentation performance is evaluated through metrics such as the Dice coefficient, Hausdorff distance, and average surface distance, and the model is iteratively optimized using the transfer learning method.

[0058] Experimental Example

[0059] Experimental Purpose:

[0060] To verify the effectiveness and superiority of the organ segmentation and lesion recognition method based on non-contrast CT images proposed in the present invention, and mainly evaluate the accuracy, efficiency of the segmentation model, and its adaptability in real clinical data.

[0061] Experimental Dataset:

[0062] The publicly available Liver Tumor Segmentation dataset is selected. This dataset contains liver CT image data from different patients and corresponding lesion annotation files, specifically including:

[0063] Training Data: 100 cases of liver CT images and lesion annotation files;

[0064] Test Data: 50 cases of liver CT images.

[0065] Experimental Steps:

[0066] Data Preprocessing:

[0067] Use the third-order spline interpolation method to unify the CT image spacing to 1mm×1mm×1mm;

[0068] Crop the image intensity values, and the retention range is the [0.5, 99.5] percentile;

[0069] Perform z-score normalization on the cropped images to ensure consistent intensity distributions of different samples.

[0070] Model Training:

[0071] Cascaded 3D U-Net: Train the first-level network on the downsampled CT images to extract global features, and train the second-level network on the full-resolution images to optimize the boundaries.

[0072] - The loss function adopts a combination of weighted Dice loss and cross-entropy loss:

[0073] L = α·L Dice + β·L CE

[0074] where α = 0.7 and β = 0.3.

[0075] - Data augmentation includes random rotation (angle range ±15°), scaling (range 0.8 - 1.2 times), mirror flipping, and random noise addition.

[0076] Segmentation result optimization:

[0077] The patch repetition method (patch size: 64×64×64, repetition rate 25%) is used to segment large-sized CT images;

[0078] The multi-model fusion strategy is used to fuse the segmentation results through the following voting mechanism:

[0079]

[0080] where w i is the confidence weight of each model.

[0081] Post-processing:

[0082] Use connectivity analysis to remove small regions with a volume less than 5% of the largest connected component;

[0083] Use morphological operations (dilation and erosion) to optimize the segmentation boundary.

[0084] Three-dimensional reconstruction and evaluation:

[0085] Reconstruct the three-dimensional liver model and the lesion area based on the segmentation results;

[0086] Use the Dice coefficient, Hausdorff distance (HD), and average surface distance (ASD) to evaluate the model performance.

[0087] Experimental results:

[0088] Evaluation index The method of the present invention Using 3DU-Net alone Dice coefficient (%) 94.2 88.5 Hausdorff distance (mm) 6.4 9.8 Average surface distance (mm) 1.8 2.5 Inference time (seconds per image) 8.3 7.6

[0089] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for organ segmentation and lesion recognition based on CT plain scan images, characterized in that: include: a. Data preprocessing: CT image data is cropped, resampled and normalized. Cropping to non-zero value areas can reduce the computational burden of invalid data. Resampling uses third-order spline interpolation to unify voxel spacing. Normalization uses the z-score formula to standardize the intensity value. The formula is as follows: f(x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 in: to i =and i μ is the mean intensity, σ is the standard deviation of intensity, b i ,c i ,d i is the interpolation parameter, where data normalization is performed by collecting all intensity values ​​I appearing in the segmentation mask of the training dataset and clipping the [0.5, 99.5] percentile I of these intensity values 0.5,0.95 , perform preliminary normalization on the entire data set. Then, according to I 0.5,0.95 Calculate the mean μ and standard deviation of the intensity value, perform z-score normalization to obtain the normalized result; b. Cascade segmentation network training: Construct a segmentation network based on cascade 3DU-Net. Cascade 3D, U-Net is responsible for extracting multi-resolution features; c. Adaptive optimization training: The generalization ability of the model is improved by dynamically adjusting the learning rate, optimizing the loss function and extending the data enhancement method. The loss function uses a combination of weighted Dice loss and cross entropy loss. The formula is as follows: L=α·L Dice +β·L CE Data augmentation includes random rotation, scaling, mirror flipping, and random noise addition to expand data sample diversity; d. Segmentation result fusion and optimization: The segmentation results are optimized by combining the patch overlap method and the multi-model fusion strategy, and the following weighted voting mechanism is used for fusion: And optimize key area features through attention mechanism; e. Post-processing analysis: Optimize the segmentation results through connectivity analysis and morphological operations, retain the largest connected components and correct the segmentation boundaries. Morphological operations include dilation, erosion and region growing; f. 3D reconstruction and lesion labeling: Reconstruct the 3D model of the CT image based on the segmentation results, use bilinear interpolation to achieve voxel reconstruction, and use self-supervised learning methods combined with expert labeling to achieve accurate labeling of the lesion area; g. Performance evaluation and model update: The segmentation performance is evaluated by indicators such as Dice coefficient, Hausdorff distance and average surface distance, and the model is iteratively optimized in combination with transfer learning methods.

2. The method for organ segmentation and lesion identification based on CT plain scan images according to claim 1, characterized in that: The cascaded 3DU-Net extracts low-resolution global features on the downsampled images and optimizes the fine-grained segmentation effect on the full-resolution images.

3. The method for organ segmentation and lesion recognition based on CT plain scan images according to claim 1, characterized in that: The adaptive optimization training adopts a dynamic learning rate adjustment strategy, and automatically reduces the learning rate when the increase of the training error in the last 30 cycles is less than a set threshold.

4. The method for organ segmentation and lesion identification based on CT plain scan images according to claim 1, characterized in that: The weighted Dice loss formula of the loss function is: Among them, w i is the weight, p i is the predicted value, g i is the true value.

5. The method for organ segmentation and lesion recognition based on CT plain scan images according to claim 1, characterized in that: The segmentation result fusion optimizes key area features by introducing a channel attention module and a spatial attention module of the attention mechanism.

6. The method for organ segmentation and lesion identification based on CT plain scan images according to claim 1, characterized in that: The post-processing analysis is based on connectivity analysis, removing all regions smaller than the largest connected component, and further optimizing the boundaries through a region growing algorithm.

7. The method for organ segmentation and lesion identification based on CT plain scan images according to claim 1, characterized in that: The three-dimensional reconstruction achieves high-precision three-dimensional model reconstruction through voxel interpolation combined with a surface fitting algorithm. The lesion annotation combines self-supervised learning with a multi-task learning framework to simultaneously predict the lesion location, volume and type label. The transfer learning method combines public data sets with specific scene data to improve the adaptability of the model under different image acquisition conditions through multiple rounds of fine-tuning.

Images