A liver tumor automatic recognition and measurement method based on CT image
By introducing a liver component knowledge graph and cascade network pattern into liver tumor identification and segmentation, the components are decomposed before tumor identification and segmentation, which solves the applicability and accuracy problems of existing liver tumor identification and segmentation technologies, and achieves more efficient and interpretable tumor identification and measurement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUZHOU BAINA MEDICAL TECHNOLOGY CO LTD
- Filing Date
- 2025-05-12
- Publication Date
- 2026-05-05
AI Technical Summary
Existing end-to-end deep learning-based methods have poor applicability and robustness in liver tumor identification and segmentation, making it difficult to guarantee the accuracy of the results and lacking model interpretability. In particular, they are difficult to identify and segment liver tumors in complex medical images.
By introducing a liver component knowledge graph and cascaded network model, the patient's abdominal CT images are first decomposed into liver components, separating multiple liver component images. Then, a liver tumor segmentation network is used for tumor identification and segmentation, and finally, tumor measurement and analysis are performed.
The model's robustness and accuracy have been improved, its applicability to different image qualities has been enhanced, and the liver component images are interpretable, assisting doctors in reviewing and correcting the results.
Smart Images
Figure CN120471884B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, specifically to a method for automatic identification and measurement of liver tumors based on CT images. Background Technology
[0002] In liver tumor surgery, tumor identification, segmentation, and measurement based on preoperative CT images are crucial and directly determine the surgical plan. However, manual tumor marking, including tumor identification, segmentation, and measurement, requires highly specialized medical skills and is a tedious and repetitive task, which is detrimental to the high-quality and efficient conduct of liver tumor surgery. Therefore, developing automated methods for liver tumor identification, segmentation, and measurement is of great value.
[0003] Currently, commonly used methods for liver tumor identification and segmentation include threshold segmentation, region growing, edge detection, machine learning methods such as support vector machines, and deep learning methods represented by convolutional neural networks. Among these, deep learning methods are currently the most mainstream and efficient. However, conventional end-to-end deep learning methods have some shortcomings and limitations in automatically identifying liver tumors, for the following reasons:
[0004] 1. Medical imaging features are complex, tumors have diverse morphologies, present different shapes, and have blurred boundaries and poor contrast; at the same time, liver tumors have a large density range (including low density, isodense or high density) and great variation (different density areas may exist within a tumor, usually caused by factors such as internal hemorrhage, necrosis, calcification, etc.).
[0005] 2. Noise and artifacts in CT images can interfere with tumor identification, such as quantum noise, motion artifacts, and metal artifacts.
[0006] 3. Individual differences: There are certain differences in the anatomical structure of the liver among different patients, including the size, shape, position of the liver and the distribution of blood vessels. This makes it difficult to accurately identify tumors due to some liver variations or malformations, bifurcation of blood vessels, enlargement, etc.
[0007] 4. Liver tumor annotation data is scarce because manual annotation of liver tumors is extremely difficult and labor-intensive, resulting in a severe shortage of high-quality annotation data.
[0008] Therefore, traditional deep learning methods that are entirely driven by labeled data have poor applicability and robustness in liver tumor identification and segmentation tasks, making it difficult to guarantee the accuracy of the results and lacking model interpretability.
[0009] To address the above issues, this invention presents an automatic liver tumor identification and measurement method based on CT images. By introducing a liver component knowledge graph and a cascaded network pattern, the robustness, accuracy, and interpretability of the model are improved. Summary of the Invention
[0010] To address the shortcomings of traditional methods, such as poor applicability and robustness, difficulty in ensuring accuracy, and lack of model interpretability, this invention provides an automatic liver tumor identification and measurement method based on CT images. This invention first decomposes the patient's abdominal CT images into multiple liver component images, including: a complete liver image, a normal liver tissue image, a high-density tissue image, and a low-density tissue image. All component images are then input into a liver tumor segmentation network for tumor identification and segmentation, ultimately achieving tumor measurement and analysis.
[0011] To achieve the above objectives, the present invention provides the following technical solution:
[0012] An automatic liver tumor identification and measurement method based on CT images includes the following steps: First, the patient's abdominal CT image is preprocessed and then input into a liver component decomposition network. The liver component decomposition network effectively decomposes the liver image, separating normal and abnormal liver tissues to obtain decomposed liver component images I1-I4. Next, the liver component images I1-I4 are input into a liver tumor segmentation network. The liver tumor segmentation network extracts and analyzes features from the input liver component images I1-I4, segments the tumor region in the liver, and outputs the tumor segmentation result. Finally, the segmentation result is input into a tumor measurement and analysis module for tumor analysis to calculate detailed tumor information.
[0013] Furthermore, this includes the model training phase and the model inference phase;
[0014] The model training phase is used to train the liver component decomposition network and the liver tumor segmentation network.
[0015] The model inference stage involves the following steps: First, the patient's abdominal CT images are preprocessed and then input into the liver component decomposition network to obtain the decomposed liver component images I1-I4. Next, the liver component images I1-I4 are input into the liver tumor segmentation network to segment the liver tumor and output the tumor segmentation results. Finally, the segmentation results are input into the tumor measurement and analysis module for tumor analysis and calculation of detailed tumor information.
[0016] Furthermore, the detailed information of the tumors includes the major diameter, minor diameter, average radius, average density, maximum density, minimum density, maximum cross-sectional area, total volume, and total mass of each tumor.
[0017] Furthermore, the model inference stage comprises four steps:
[0018] 1.1 Data preprocessing;
[0019] 1.2 Liver component decomposition: Abdominal CT images are input into the liver component decomposition network to obtain liver component images I1-I4;
[0020] 1.3 Tumor segmentation: Input the liver component images I1-I4 into the liver tumor segmentation network to obtain the tumor segmentation results;
[0021] 1.4 Tumor Measurement and Analysis: The tumor measurement and analysis module measures and analyzes the tumor segmentation results and outputs detailed tumor information.
[0022] Furthermore, in step 1.1, data preprocessing, the specific steps are as follows: the patient's abdominal CT images will be pre-parsed from the DICOM format data and the images will be normalized. Specifically, the images will be resampled to a uniform resolution, which is the median resolution of all image resolution distributions, and the image value range will be normalized to [0, 1].
[0023] Furthermore, the tumor measurement and analysis module uses methods including: three-dimensional connected domain detection, internal cavity filling, major and minor axis calculation, area and volume calculation, and density and mass calculation.
[0024] Furthermore, the model training phase includes four steps:
[0025] 2.1 Data Preprocessing: Preprocess the patient's abdominal CT images;
[0026] 2.2 Data augmentation;
[0027] 2.3 Model Training:
[0028] The preprocessed abdominal CT image is input and then sequentially fed into the liver component decomposition network and the liver tumor segmentation network. Then, the forward derivation of the liver component decomposition network and the liver tumor segmentation network are performed and the network weight gradient is calculated to obtain the output of the liver component decomposition network, namely the liver component images I1-I4, and the output of the liver tumor segmentation network, namely the segmented tumor mask P.
[0029] During the gradient backpropagation and weight update phases, the liver component decomposition network and the liver tumor segmentation network are used alternately:
[0030] First, freeze the weights of the liver component decomposition network, calculate the segmentation loss value L2 between the segmented tumor mask P and the labeled tumor gold standard T, and then perform gradient backpropagation and network weight update on the liver tumor segmentation network.
[0031] Secondly, the weights of the liver component decomposition network are unfrozen, and the weights of the liver tumor segmentation network are frozen. Then, the mutual information loss is calculated using the liver component images I1-I4 and the input abdominal CT images, and the sum is accumulated to obtain the loss value L1. Then, the similarity SSIM loss value L0 is calculated between the liver component images I1-I4 and the manually decomposed liver component images J1-J4. Finally, the L1 and L0 loss values are added to the previously calculated segmentation loss value L2 to obtain the final loss value L3. Then, gradient backpropagation and network weight update are performed on the liver component decomposition network.
[0032] Once the liver component decomposition network and the liver tumor segmentation network have completed gradient backpropagation and network weight updates, they enter the next training loop. In each training loop, different data is sampled for training.
[0033] 2.4 Model Validation: After training for a specified number of rounds, use validation data to test the effectiveness of the current model weights, and select the model weights with the best performance for saving.
[0034] Furthermore, step 2.2, data augmentation, specifically involves performing random image flipping, random image translation, random image rotation, random image elastic deformation, and random image contrast adjustment on the preprocessed abdominal CT images.
[0035] Furthermore, the segmentation loss value L2 = Dice loss value + CE cross-entropy loss value.
[0036] Furthermore, the main body of the liver component decomposition network is a 3D UNet network. The input abdominal CT image is used to extract depth features and decode them into component weight maps through the 3D UNet network. The component weight maps have the same size as the input abdominal CT image and are converted into component weight probability maps through a softmax activation function. Finally, the input abdominal CT image is multiplied by each weight probability map to obtain the decomposed liver component images I1-I4.
[0037] The backbone of the liver tumor segmentation network is a UNet network structure. The input liver component images I1-I4 are downsampled and encoded three times, and then upsampled and decoded three times. In between, cross-layer connections are used to connect features of the same scale. The output of the final decoding layer is then passed through a sigmoid activation function to obtain the final output tumor segmentation result, i.e., the segmented tumor mask P.
[0038] Compared with the prior art, the beneficial effects of this invention are as follows: 1) This invention is highly robust. Compared with the traditional single liver tumor segmentation network, this invention introduces a liver component decomposition network before the segmentation network, which effectively filters out other useless information and has higher learning efficiency and accuracy.
[0039] 2) This invention has strong applicability. Compared with traditional standalone liver tumor segmentation networks, the liver component decomposition network of this invention outputs standardized image components, making the input domain of the liver tumor segmentation network simple and singular, and thus applicable to CT images of various image qualities.
[0040] 3) This invention has high accuracy. This invention fully incorporates prior information about liver tumor components, including the extent of liver tissue, normal liver tissue components, high-density components, and low-density components. Therefore, it can improve the effectiveness of model feature engineering and make the output results more accurate.
[0041] 4) The model of this invention is interpretable. The liver component images I1-I4 of the liver component decomposition network in this invention are intuitively interpretable. The liver component images I1-I4 sequentially reflect the tissue extent of the liver, normal liver tissue, high-density liver tissue, and low-density liver tissue. Attached Figure Description
[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0043] Figure 1 This is a flowchart of the automatic liver tumor identification and measurement method based on CT images according to the present invention;
[0044] Figure 2 This is a schematic diagram of the model training phase in this invention;
[0045] Figure 3 This is a schematic diagram of the liver component decomposition network in this invention;
[0046] Figure 4 This is a schematic diagram of the liver tumor segmentation network in this invention. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0048] Example 1: In some embodiments, please refer to the accompanying drawings. Figure 1-2 An automatic identification and measurement method for liver tumors based on CT images, including a model training stage and a model inference stage;
[0049] The model inference stage involves the following steps: First, the patient's abdominal CT images are preprocessed and then input into the liver component decomposition network to obtain the decomposed liver component images I1-I4. Next, the liver component images I1-I4 are input into the liver tumor segmentation network to segment the liver tumor and output the tumor segmentation results. Finally, the segmentation results are input into the tumor measurement and analysis module for tumor analysis and calculation of detailed tumor information, including the major diameter, minor diameter, average radius, average density, maximum density, minimum density, maximum cross-sectional area, total volume, and total mass of all tumors.
[0050] The model inference phase consists of four steps:
[0051] 1.1 Data Preprocessing: The patient's abdominal CT images are pre-parsed from DICOM format data and normalized. Specifically, the images are resampled to a uniform resolution, which is the median resolution of all image resolution distributions, and the image value range is normalized to [0, 1].
[0052] 1.2 Liver component decomposition: Abdominal CT images are input into the liver component decomposition network to obtain liver component images I1-I4;
[0053] 1.3 Tumor segmentation: Input the liver component images I1-I4 into the liver tumor segmentation network to obtain the tumor segmentation results;
[0054] 1.4 Tumor Measurement and Analysis: The tumor measurement and analysis module measures and analyzes the tumor segmentation results, and outputs the major diameter, minor diameter, average radius, average density, maximum density, minimum density, maximum cross-sectional area, total volume, and total mass of all tumors.
[0055] The tumor measurement and analysis module uses methods including: three-dimensional connected domain detection, internal cavity filling, major and minor axis calculation, area and volume calculation, density and mass calculation, etc.
[0056] This invention decomposes the abdominal CT images of patients into components and identifies and segments tumors based on the decomposed components. This network design of decomposing components first and then segmenting tumors can effectively improve the identification ability and segmentation accuracy of liver tumors. At the same time, the decomposed component map is also a feature analysis of the model for identifying and segmenting tumors, which can help doctors to review and correct the data.
[0057] The model training phase incorporates a liver component knowledge graph to guide the training of a cascaded network, which includes a liver component decomposition network and a liver tumor segmentation network. The liver component decomposition network effectively decomposes liver images, separating normal and abnormal liver tissues. The liver tumor segmentation network extracts and analyzes features from the input liver components to segment tumor regions within the liver.
[0058] The model training phase is used to train the liver component decomposition network and the liver tumor segmentation network. This training process is end-to-end and consists of four steps:
[0059] 2.1 Data Preprocessing: Preprocess the patient's abdominal CT images. Refer to step 1.1 for the preprocessing method.
[0060] 2.2 Data Augmentation: The preprocessed abdominal CT images are subjected to random image flipping, random image translation, random image rotation, random image elastic deformation, and random image contrast adjustment.
[0061] 2.3 Model Training:
[0062] The preprocessed abdominal CT image is input and then sequentially fed into the liver component decomposition network and the liver tumor segmentation network. Then, the forward derivation of the liver component decomposition network and the liver tumor segmentation network are performed and the network weight gradient is calculated to obtain the output of the liver component decomposition network, namely the liver component images I1-I4, and the output of the liver tumor segmentation network, namely the segmented tumor mask P.
[0063] During the gradient backpropagation and weight update phases, the liver component decomposition network and the liver tumor segmentation network are used alternately:
[0064] First, freeze the weights of the liver component decomposition network, calculate the segmentation loss value L2 between the segmented tumor mask P and the labeled tumor gold standard T (segmentation loss value L2 = Dice loss value + CE cross-entropy loss value), and then perform gradient backpropagation and network weight update on the liver tumor segmentation network.
[0065] Secondly, the weights of the liver component decomposition network are unfrozen, while the weights of the liver tumor segmentation network are frozen. Then, the mutual information loss (MI Loss) is calculated using the liver component images I1-I4 and the input abdominal CT images, and the sum is accumulated to obtain the loss value L1. Next, the similarity SSIM loss value L0 is calculated between the liver component images I1-I4 and the manually decomposed liver component images J1-J4. Finally, the L1 and L0 loss values are added to the previously calculated segmentation loss value L2 to obtain the final loss value L3. Then, gradient backpropagation and network weight updates are performed on the liver component decomposition network.
[0066] Once both the liver component decomposition network and the liver tumor segmentation network have completed gradient backpropagation and network weight updates, they enter the next training cycle. In each training cycle, different data is sampled for training.
[0067] Among them, the liver component images I1-I4 are interpretable, representing: the extent of liver tissue, the normal liver tissue area, the high-density liver area, and the high-visceral-low-density liver area, respectively. The manually decomposed liver component images J1-J4 are manually labeled liver tissue extent, normal liver tissue area, high-density liver area, and high-visceral-low-density liver area.
[0068] 2.4 Model Validation: After training for a specified number of rounds, use validation data to test the effectiveness of the current model weights, and select the model weights with the best performance for saving.
[0069] Example 2: In some embodiments, such as Figure 3 As shown, the main body of the liver component decomposition network is a classic 3D UNet network. The input is an abdominal CT image, which is used to extract depth features and decode them into component weight maps. The component weight maps have the same size as the input abdominal CT image and are converted into component weight probability maps through a softmax activation function. Finally, the input abdominal CT image is multiplied by each weight probability map to obtain the decomposed liver component images I1-I4.
[0070] Example 3: In some embodiments, such as Figure 4 As shown, the backbone of the liver tumor segmentation network is a UNet network structure. The input liver component images I1-I4 are downsampled and encoded three times, and then upsampled and decoded three times. In between, cross-layer connections are used to connect features of the same scale. The output of the final decoding layer is then passed through a sigmoid activation function to obtain the final output tumor segmentation result, i.e., the segmented tumor mask P.
[0071] The present invention has the following technical effects:
[0072] 1) This invention is highly robust. Compared to traditional standalone liver tumor segmentation networks, this invention introduces a liver component decomposition network before the segmentation network, effectively filtering out other useless information, resulting in higher learning efficiency and accuracy.
[0073] 2) This invention has strong applicability. Compared with traditional standalone liver tumor segmentation networks, the liver component decomposition network of this invention outputs standardized image components, making the input domain of the liver tumor segmentation network simple and singular, and thus applicable to CT images of various image qualities.
[0074] 3) This invention has high accuracy. This invention fully incorporates prior information about liver tumor components, including the extent of liver tissue, normal liver tissue components, high-density components, and low-density components. Therefore, it can improve the effectiveness of model feature engineering and make the output results more accurate.
[0075] 4) The model of this invention is interpretable. The liver component images I1-I4 of the liver component decomposition network in this invention are intuitively interpretable. The liver component images I1-I4 sequentially reflect the tissue extent of the liver, normal liver tissue, high-density liver tissue, and low-density liver tissue.
[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for automatic identification and measurement of liver tumors based on CT images, characterized in that, The process includes the following steps: First, the patient's abdominal CT images are preprocessed and then input into a liver component decomposition network. This network effectively decomposes the liver images, separating normal and abnormal liver tissue to obtain decomposed liver component images I1-I4. These images sequentially represent the tissue extent, normal liver tissue, high-density liver tissue, and low-density liver tissue. Next, these images are input into a liver tumor segmentation network. This network extracts and analyzes features from the images, segmenting the tumor region within the liver and outputting the segmentation results. Finally, the segmentation results are input into a tumor measurement and analysis module for tumor analysis, calculating detailed tumor information. This includes the model training phase and the model inference phase; The model training phase is used to train the liver component decomposition network and the liver tumor segmentation network. The model inference stage involves the following steps: First, the patient's abdominal CT image is preprocessed and then input into the liver component decomposition network to obtain the decomposed liver component images I1-I4. Next, the liver component images I1-I4 are input into the liver tumor segmentation network to segment the liver tumor and output the tumor segmentation results. Finally, the segmentation results are input into the tumor measurement and analysis module for tumor analysis and to calculate detailed tumor information. The model training phase includes four steps: 2.1 Data Preprocessing: Preprocess the patient's abdominal CT images; 2.2 Data augmentation; 2.3 Model Training: Input the preprocessed abdominal CT image and input it sequentially into the liver component decomposition network and the liver tumor segmentation network; then perform forward derivation on the liver component decomposition network and the liver tumor segmentation network and calculate the network weight gradient to obtain the liver component images I1-I4 and the segmented tumor mask P. During the gradient backpropagation and weight update phases, the liver component decomposition network and the liver tumor segmentation network are used alternately: First, freeze the weights of the liver component decomposition network, calculate the segmentation loss value L2 between the segmented tumor mask P and the labeled tumor gold standard T, and then perform gradient backpropagation and network weight update on the liver tumor segmentation network. Secondly, the weights of the liver component decomposition network are unfrozen, and the weights of the liver tumor segmentation network are frozen. Then, the mutual information loss is calculated using the liver component images I1-I4 and the input abdominal CT images, and the sum is accumulated to obtain the loss value L1. Then, the similarity SSIM loss value L0 is calculated between the liver component images I1-I4 and the manually decomposed liver component images J1-J4. Finally, the L1 and L0 loss values are added to the previously calculated segmentation loss value L2 to obtain the final loss value L3. Then, gradient backpropagation and network weight update are performed on the liver component decomposition network. Once the liver component decomposition network and the liver tumor segmentation network have completed gradient backpropagation and network weight updates, they enter the next training loop. In each training loop, different data is sampled for training. 2.4 Model Validation: After training for a specified number of rounds, use validation data to test the effectiveness of the current model weights, and select the model weights with the best performance for saving.
2. The method for automatic identification and measurement of liver tumors based on CT images according to claim 1, characterized in that, The detailed information of the tumors includes the major diameter, minor diameter, average radius, average density, maximum density, minimum density, maximum cross-sectional area, total volume, and total mass of each tumor.
3. The method for automatic identification and measurement of liver tumors based on CT images according to claim 1, characterized in that, The model inference phase consists of four steps: 1.1 Data preprocessing; 1.2 Liver component decomposition: Abdominal CT images are input into the liver component decomposition network to obtain liver component images I1-I4; 1.3 Tumor segmentation: Input the liver component images I1-I4 into the liver tumor segmentation network to obtain the tumor segmentation results; 1.4 Tumor Measurement and Analysis: The tumor measurement and analysis module measures and analyzes the tumor segmentation results and outputs detailed tumor information.
4. The method for automatic identification and measurement of liver tumors based on CT images according to claim 3, characterized in that, Step 1.1: Data preprocessing. The specific steps are as follows: The patient's abdominal CT images will be pre-parsed from the DICOM format data and the images will be normalized. Specifically, the images will be resampled to a uniform resolution, which is the median resolution of all image resolution distributions. At the same time, the image value range will be normalized to [0, 1].
5. The method for automatic identification and measurement of liver tumors based on CT images according to claim 3, characterized in that, The methods used in the tumor measurement and analysis module include: three-dimensional connected domain detection, internal cavity filling, major and minor axis calculation, area and volume calculation, and density and mass calculation.
6. The method for automatic identification and measurement of liver tumors based on CT images according to claim 1, characterized in that, Step 2.2, Data Augmentation, specifically involves performing random image flipping, random image translation, random image rotation, random image elastic deformation, and random image contrast adjustment on the preprocessed abdominal CT images.
7. The method for automatic identification and measurement of liver tumors based on CT images according to claim 6, characterized in that, The segmentation loss value L2 = Dice loss value + CE cross-entropy loss value.
8. The method for automatic identification and measurement of liver tumors based on CT images according to claim 1, characterized in that, The main body of the liver component decomposition network is a 3D UNet network. The input is an abdominal CT image. The 3D UNet network extracts depth features and decodes them into component weight maps. The component weight maps have the same size as the input abdominal CT image and are converted into component weight probability maps through a softmax activation function. Finally, the input abdominal CT image is multiplied by each weight probability map to obtain the decomposed liver component images I1-I4. The backbone of the liver tumor segmentation network is a UNet network structure. The input liver component images I1-I4 are downsampled and encoded multiple times, and then upsampled and decoded multiple times. In between, cross-layer connections are used to connect features of the same scale. The output of the final decoding layer is then activated by the sigmoid activation function to obtain the segmented tumor mask P.