A liver tumor segmentation method and system based on plain scan CT and a storage medium

Abstract

This invention discloses a liver tumor segmentation method, system, and storage medium based on plain CT scans, belonging to the field of medical image processing technology. It addresses the problem of the lack of existing liver tumor segmentation methods that can both improve the clarity of liver tumor boundaries and preserve detailed information. The key technical points of this invention include: Step 1, acquiring plain CT scan data of the liver from liver tumor patients; Step 2, segmenting the plain CT images using a pre-trained segmentation model to obtain a liver region mask; the segmentation model is based on the UNet model, using PVT-V2 as the backbone network of the encoder, and the decoder adopts a cascaded upsampling path, fusing a residual denoising module, a gated attention module, a multi-scale feature fusion module, and a depth supervision mechanism in each stage; Step 3, further segmenting the liver region mask using the segmentation model to obtain a liver tumor mask; Step 4, extracting features and classifying the liver region mask using a classification model to obtain the benign or malignant tumor result.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, to a method, system, and storage medium for segmenting liver tumors based on plain CT scans. Background Technology

[0002] CT imaging, as a core tool for liver tumor detection, provides detailed tissue structure information due to its high resolution. However, in low-dose plain CT images, liver tumors often appear as low-contrast areas with blurred boundaries, posing a significant challenge to automated segmentation techniques. Furthermore, existing segmentation techniques struggle to preserve tumor region details while smoothing segmentation boundaries, directly impacting the accuracy of tumor detection and the efficiency of clinical diagnosis.

[0003] In recent years, the rapid development of artificial intelligence technology has brought new opportunities to medical image analysis. Deep learning methods based on convolutional neural networks (CNNs) can automatically extract multi-level image features. X-Net, developed by Chi et al., combines parallel 2D and 3D networks; the former segments tumors while the latter extracts liver features, achieving accurate segmentation of the liver and tumors, and is widely used in tasks such as lesion classification and segmentation. For example, Du et al. used Res-Net combined with spatial attention mechanisms to distinguish between pneumonia-type lung cancer and pneumonia; Zhao et al. used 3D U-Net and ResNet to segment renal masses; and Yang et al. combined multi-channel features with random forests to classify pancreatic cystic tumors. However, current technologies still lack liver tumor segmentation methods that can both improve the clarity of liver tumor boundaries and preserve detailed information. Summary of the Invention

[0004] The technical problem to be solved by this invention is:

[0005] There is a lack of liver tumor segmentation methods in the current technology that can both improve the clarity of liver tumor boundaries and preserve detailed information.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a method for identifying liver tumors based on plain CT scans, comprising the following steps:

[0008] Step 1: Collect plain CT scan data of the liver from patients with liver tumors;

[0009] Step 2: The pre-trained segmentation model is used to segment the plain CT image to obtain a liver region mask. The segmentation model is based on the UNet model, using PVT-V2 as the backbone network of the encoder and the decoder adopts a cascaded upsampling path. In each stage, a residual denoising module, a gated attention module, a multi-scale feature fusion module, and a deep supervision mechanism are integrated to enhance the expressive power of semantic features.

[0010] Step 3: Use a segmentation model to segment the liver region mask again to obtain the liver tumor mask;

[0011] Step 4: Use a pre-trained classification model to extract features and classify the liver region mask to obtain the results of benign or malignant tumors.

[0012] Furthermore, the residual denoising module includes two sets of 3×3 convolution and batch normalization modules, and at the end, the residual connection between the input and output is realized through identity mapping; its forward propagation process is as follows: the input features are first processed by a set of 3×3 convolution → batch normalization → ReLU activation; after being processed by 3×3 convolution and batch normalization, they are added element-wise with the original input; finally, the output is activated by ReLU to form a purified feature map.

[0013] Furthermore, the functional implementation process of the gated attention module is as follows: the feature map x of the encoder and the gated signal g of the self-decoder are convolved respectively to unify the channel dimension, the two are added together and then activated by ReLU and BatchNorm; then convolved and activated by Sigmoid to generate attention weights; finally, the original input features are multiplied pixel by pixel and weighted by the weights before output.

[0014] Furthermore, the decoder employs a 4-level upsampling path, introducing a depth supervision branch at the output of the first to third level decoders. Each branch contains a 1×1 convolutional output layer, which is upsampled to the same spatial resolution as the main output through bilinear interpolation and participates in the total loss calculation together with the main output.

[0015] Furthermore, step two also includes: enhancing the obtained liver mask. First, the main liver region is preserved through binarization and maximum connected component filtering. Then, the obtained mask undergoes morphological dilation processing, calculated as follows:

[0016] (1)

[0017] First, the multiple channels of the original image are combined into a multi-channel representation for subsequent processing, where Y, U, and V correspond to the luminance and color components of the image, respectively. Then, the merged channels are linearly normalized, and the luminance channel Y is subjected to histogram equalization operation H(Y). M' represents the liver mask after morphological dilation, and ⊙ represents element-wise multiplication.

[0018] Furthermore, the loss function of the segmentation model is constructed as follows:

[0019]

[0020]

[0021] w i Represents the weights of each loss function, i = 1, 2, loss BCE Represents cross-entropy loss, loss Tversky Indicates Tversky loss, y (i) p represents the true label of the i-th sample. (i) This represents the probability that the model predicts the i-th sample to be positive, N represents the total number of samples, and α and β are parameters that weigh false positives (FP) and false negatives (FN).

[0022]

[0023] Let be the combined loss function of BCE loss and Tversky loss, where a is the weighting factor;

[0024] With deep supervision, the total loss of the network is given by the formula:

[0025]

[0026] Among them, auxi loss Output the loss for each level of depth supervision, i = 1, 2, 3.

[0027] Furthermore, the classification model is based on ResNet34-base and uses a hierarchical residual structure to achieve deep feature reuse and stable gradient propagation.

[0028] This invention provides a liver tumor segmentation system based on plain CT scan. The system has a program module corresponding to the steps of any of the above-described technical solutions, and executes the steps in the above-described liver tumor segmentation method based on plain CT scan when running.

[0029] The present invention provides a computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps of the liver tumor segmentation method based on plain CT scan as described in any of the above technical solutions.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] In this invention, the encoder stage uses PVT-V2 to extract global information and enhance spatial details. At each stage, the decoder guides the network to automatically focus on the target region (such as the tumor boundary or internal region) through a fusion-gated attention mechanism, suppressing background interference. A residual denoising module is introduced for feature purification to suppress high-frequency artifacts and redundant information introduced by the upsampling process, significantly enhancing features in regions with blurred boundaries or small tumor volumes. Multi-scale fusion improves the semantic consistency of feature expression, enhancing feature reconstruction capabilities and improving the recognition performance of small tumor regions. A deep supervision mechanism is introduced to accelerate model convergence, improve the expressive power of intermediate feature layers, and enhance the hierarchical expression of features. Attached Figure Description

[0032] Figure 1 This is a flowchart of a liver tumor segmentation method based on plain CT scan in an embodiment of the present invention;

[0033] Figure 2 This is a structural diagram of the segmentation model in an embodiment of the present invention;

[0034] Figure 3 This is a structural diagram of the residual denoising module in an embodiment of the present invention;

[0035] Figure 4 This is a structural diagram of the gated attention module in an embodiment of the present invention;

[0036] Figure 5 This is a flowchart of the LE operation in an embodiment of the present invention;

[0037] Figure 6 The images shown are three-dimensional tomographic scan images in this embodiment of the invention. Figure a is the original plain CT image, Figure b is the image processed using LE, Figure c is a real label drawn by the doctor by hand, and Figure d is the effect image segmented by the model.

[0038] Figure 7 The figures show the comparative test results of the various models in the embodiments of the present invention. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0041] Specific Implementation Plan 1: Combining Figure 1 and Figure 2 As shown, this invention provides a method for identifying liver tumors based on plain CT scans, comprising the following steps:

[0042] Step 1: Collect plain CT scan data of the liver from patients with liver tumors;

[0043] Step 2: The pre-trained segmentation model is used to segment the plain CT image to obtain a liver region mask. The segmentation model is based on the UNet model, using PVT-V2 as the backbone network of the encoder and the decoder adopts a cascaded upsampling path. In each stage, a residual denoising module, a gated attention module, a multi-scale feature fusion module, and a deep supervision mechanism are integrated to enhance the expressive power of semantic features.

[0044] Step 3: Use a segmentation model to segment the liver region mask again to obtain the liver tumor mask;

[0045] Step 4: Use a pre-trained classification model to extract features and classify the liver region mask to obtain the results of benign or malignant tumors.

[0046] PVT-V2 is used to extract multi-scale long-range dependent features from input CT images. A gated attention mechanism improves the model's response to key regions (such as tumors), suppresses background interference, and enhances the quality of feature fusion. A residual denoising module cleans up features during decoding, suppressing high-frequency noise introduced by upsampling. A deep supervision mechanism accelerates model convergence, enhances the expressive power of intermediate feature layers, and strengthens the hierarchical representation of features. A single 256×256 slice of CT liver image is cropped as input to the network model, and after processing by the encoder and decoder structure, the segmentation prediction result is output.

[0047] This implementation uses PVT-v2-b3 as the backbone network of the encoder, which has stronger local continuity modeling capabilities and greater flexibility in processing images of arbitrary resolution. The structure of PVT-v2 is divided into four stages (Stage 1 to Stage 4), with each stage progressively extracting more abstract semantic features while achieving progressive downsampling of spatial resolution, forming a multi-scale feature representation with a pyramid structure.

[0048] The specific structure is as follows:

[0049] Stage 1: The input image is first subjected to preliminary feature encoding through the Overlapping Patch Embedding module, which outputs a low-level feature map E1 with high spatial resolution and low channel dimension (64 channels).

[0050] Stage 2 to Stage 4: Each stage consists of multiple Transformer Blocks, and the feature map size is gradually reduced and the number of channels is increased through the PatchMerging module, outputting E2 (128 channels), E3 (320 channels) and E4 (512 channels) in sequence to obtain deeper global semantic information.

[0051] To better adapt to the characteristics of liver tumor segmentation tasks, this implementation scheme adjusts PVT-v2 by removing the classification head from the original structure and outputting four feature maps (E1–E4) at different scales in each downsampling stage for the decoder to perform feature fusion and layer-by-layer reconstruction. This pyramid-like multi-scale representation method provides effective support for the model to capture global and local information of the tumor region, which is especially beneficial for the fine identification of small lesions.

[0052] The decoder stage employs a four-level upsampling path, and in each level, it integrates a gated attention mechanism, a residual denoising block, and a multi-scale feature fusion structure to enhance feature reconstruction capabilities and improve the recognition performance of small-volume tumor regions.

[0053] The global information is first extracted by PVT, and then the local features are refined by CNN to enhance spatial details.

[0054] In each decoding stage, the input features are first upsampled through transposed convolution, and then cleaned by a residual denoising module to suppress high-frequency artifacts and redundant information introduced by the upsampling process. Subsequently, combined with features from the same layer of the encoder, a gated attention mechanism guides the network to automatically focus on the target region (such as the tumor boundary or internal region) and suppress interference from the background region. The attention-weighted encoder features are concatenated with the output of the current layer of the decoder and fused, and semantic information is integrated through a set of convolutional operations (ConvBlock). This approach better preserves global information while refining local features and restoring spatial details. A single CT liver image slice cropped to 256×256 is fed into the network model, and after processing by the encoder and decoder structures, the segmentation prediction result is output.

[0055] Specific Implementation Scheme Two: The residual denoising module adopts as follows: Figure 3The lightweight residual denoising block shown contains two sets of 3×3 convolutions and batch normalization modules, and implements a residual connection between the input and output through an identity shortcut at the end. Its forward propagation process is as follows: the input features are first processed through a set of 3×3 convolutions → batch normalization → ReLU activation; after further processing with 3×3 convolutions and batch normalization, they are added element-wise with the original input; finally, the output is activated by ReLU to form the purified feature map. The calculation is expressed by the formula:

[0056]

[0057] Where x represents the input feature and y represents the output feature. The residual connection mechanism effectively alleviates the gradient vanishing problem in deep networks while enhancing the preservation of low-level structural information and improving the integrity of feature representation. In this model, the residual denoising module is embedded in each upsampling stage of the decoder, reshaping and suppressing noise in the reconstruction process, especially showing significant feature enhancement for regions with blurred boundaries or small tumor volumes. Other aspects of this implementation scheme are the same as in Implementation Scheme 1.

[0058] Specific Implementation Scheme 3: The gating attention module is as follows: Figure 4 As shown, the functional implementation process is as follows: The encoder feature map x and the gated signal g of the self-decoder are used as inputs. The encoder features are weighted by channel-wise attention coefficients. 1×1 convolutions are performed on x and g respectively to unify the channel dimensions. The two are then summed and activated by ReLU and BatchNorm. After another 1×1 convolution and Sigmoid activation function, attention weights are generated. Finally, these weights are used to perform pixel-wise multiplication on the original input features x before outputting. Other aspects of this implementation scheme are the same as in specific implementation scheme two.

[0059] Implementation Scheme 4: The concatenated feature maps are further compressed and fused using 1×1 and 3×3 convolutions to improve the semantic consistency of feature representation. Furthermore, to enhance the stability of model training and the discriminative power of intermediate layer features, deep supervised branches are introduced at the outputs of the first to third level decoders. Each branch contains a 1×1 convolutional output layer, which is upsampled to the same spatial resolution as the main output through bilinear interpolation. This upsampled layer participates in the total loss calculation along with the main output, effectively mitigating the vanishing gradient problem and promoting multi-layer feature collaborative optimization. The entire decoding process presents a structural design combining layer-by-layer upsampling, feature fusion, denoising, and supervision, possessing both local detail restoration capabilities and global semantic awareness, making it particularly suitable for fine segmentation tasks of small-volume liver tumors in complex backgrounds. Other aspects of this implementation scheme are the same as in Implementation Scheme 3.

[0060] Specific implementation plan five: Step two also includes: such as Figure 5 and Figure 6 As shown, after obtaining the liver mask, the ROI of the liver region is enhanced based on the LE (Liver Enhancement) module. First, a trained segmentation model is used to predict the liver mask from the abdominal CT image, and the main liver region is retained by binarization and maximum connected component filtering. Subsequently, the obtained mask is subjected to appropriate morphological dilation to ensure coverage of small tumors or boundary structures that may exist near the liver edge, thereby improving the coverage of subtle lesions.

[0061] This invention's enhancement method, while preserving background information, significantly highlights the internal structure of the liver, providing more discriminative features for subsequent segmentation and classification networks. The calculation expression is:

[0062] (1)

[0063] In the formula, the MergeChannels operation first combines multiple channels of the original image into a multi-channel representation for subsequent processing, where Y, U, and V correspond to the luminance and color components (chroma channels) of the image, respectively, thus comprehensively describing the luminance and color information of the image. Then, the merged channels are linearly normalized to ensure that each pixel value is uniformly mapped to a standard numerical range, which is beneficial for stabilizing the subsequent image enhancement and network learning process. To enhance the contrast of the region of interest in the liver, a histogram equalization operation H(Y) is performed on the luminance channel Y. By redistributing the probability distribution of pixel gray levels, local contrast is improved, which helps to reveal detailed information and improve the visibility of low-contrast structures. Furthermore, the symbol M' represents a liver mask after morphological dilation, used to expand the spatial extent of the original liver region to cover the liver edges and any small lesions that may exist, ensuring that these important areas are not missed during the enhancement process. Finally, element-wise multiplication, represented by the symbol ⊙, is used to multiply the liver mask and the enhanced image features at the pixel level. This highlights the features of the region of interest and suppresses background-irrelevant information, thereby generating more discriminative and stable input features for subsequent network segmentation and classification modules. Other aspects of this implementation scheme are the same as in specific implementation scheme four.

[0064] Specific implementation plan six: The loss function of the segmentation model is constructed as follows:

[0065]

[0066]

[0067] wi Represents the weights of each loss function, i = 1, 2, loss BCE Represents cross-entropy loss, loss Tversky Indicates Tversky loss, y (i) p represents the true label of the i-th sample. (i) This represents the probability that the model predicts the i-th sample to be positive, N represents the total number of samples, and α and β are parameters that weigh false positives (FP) and false negatives (FN).

[0068]

[0069] Let be the combined loss function of BCE loss and Tversky loss, where a is the weighting factor;

[0070] With deep supervision, the total loss of the network is given by the formula:

[0071]

[0072] Among them, auxi loss Output the loss for each level of depth supervision, i = 1, 2, 3.

[0073] This implementation plan is otherwise the same as Implementation Plan Five.

[0074] Specific Implementation Scheme Seven: The classification model is based on ResNet34-base, and uses a hierarchical residual structure to achieve deep feature reuse and stable gradient propagation. This implementation scheme is otherwise the same as Specific Implementation Scheme Six.

[0075] The liver tumor segmentation method (algorithm) based on plain CT proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0076] Based on the method proposed in this invention, a liver tumor segmentation system based on plain CT scan is developed using a programming language. This system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned liver tumor segmentation method based on plain CT scan when running.

[0077] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-described liver tumor segmentation method based on plain CT scan when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0078] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0079] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0080] The beneficial effects of the present invention will be described below with reference to specific embodiments.

[0081] Example 1

[0082] The experiments in this embodiment were conducted in the following hardware and software environment to ensure the efficiency of model training and the reliability of the results: CPU: 16 vCPU Intel(R) Xeon(R) Platinum 8352VCPU@2.10GHz; GPU: RTX4090 (24GB); RAM: 120GB; Operating System: Ubuntu 22.04; Development Environment: Python 3.12, CUDA 12.4, PyTorch 2.5.1; Model Batch Size: 8; Initial Learning Rate: 1e-4; Optimizer: AdamW; 500 epochs per model. Data augmentation techniques such as random cropping and random rotation were used to expand the dataset. The segmentation results of each model for liver tumors in plain CT scans are shown in Table 1 and... Figure 7 As shown.

[0083] Table 1

[0084]

[0085] As shown in Table 4, TransUnet combines the Transformer and U-Net architectures, enabling it to capture global contextual information while preserving local details, thus demonstrating certain advantages in liver tumor segmentation. MISSFormer, a recently proposed Transformer-type segmentation model, enhances its ability to detect small targets through multi-scale information fusion, resulting in improvements in Recall and Precision, but its overall performance still falls short of the method proposed in this invention. Overall, the method proposed in this invention significantly outperforms the baseline model Attention-Unet++L4 across all evaluation metrics, with DSC increasing from 0.6486 to 0.8379, Jaccard from 0.5719 to 0.7860, and Recall and Precision increasing to 0.8835 and 0.8512, respectively. This indicates that even under low-contrast plain CT conditions, the method proposed in this invention can still effectively identify tumor regions while maintaining a low false detection rate.

[0086] Among the comparative methods, TransUnet showed some improvement in DSC and Jaccard on plain scan data, but the increase in recall and precision was limited, indicating that its global feature modeling ability has limited adaptability to low-contrast data. MISSFormer had a lower recall (0.6266) but a higher precision (0.8782) on plain scan data, indicating that this method tends to reduce false positives, but its sensitivity in detecting tumor regions is insufficient. In contrast, the method of this invention achieves a better balance between recall and precision, ensuring both a high detection rate of tumor regions and a low false positive rate, demonstrating superior robustness and cross-domain adaptability.

[0087] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for identifying liver tumors based on plain CT scan, characterized in that, Includes the following steps: Step 1: Collect plain CT scan data of the liver from patients with liver tumors; Step 2: The pre-trained segmentation model is used to segment the plain CT image to obtain a liver region mask. The segmentation model is based on the UNet model, using PVT-V2 as the backbone network of the encoder and the decoder adopts a cascaded upsampling path. In each stage, a residual denoising module, a gated attention module, a multi-scale feature fusion module, and a deep supervision mechanism are integrated to enhance the expressive power of semantic features. Step 3: Use a segmentation model to segment the liver region mask again to obtain the liver tumor mask; Step 4: Use a pre-trained classification model to extract features and classify the liver region mask to obtain the results of benign or malignant tumors.

2. The method according to claim 1, characterized in that, The residual denoising module includes two sets of 3×3 convolution and batch normalization modules, and at the end, the residual connection between the input and output is realized through identity mapping. Its forward propagation process is as follows: the input features are first processed by a set of 3×3 convolution → batch normalization → ReLU activation; after being processed by 3×3 convolution and batch normalization, they are added element-wise with the original input; finally, the output is activated by ReLU to form a purified feature map.

3. The method according to claim 2, characterized in that, The function implementation process of the gated attention module is as follows: the feature map x of the encoder and the gated signal g of the self-decoder are convolved respectively to unify the channel dimension. The two are added together and then activated by ReLU and Batch Norm. After convolution and Sigmoid activation function, attention weights are generated. Finally, the original input features are multiplied pixel by pixel and weighted before output.

4. The method according to claim 3, characterized in that, The decoder employs a 4-level upsampling path, introducing a depth supervision branch at the output of the first to third level decoders. Each branch contains a 1×1 convolutional output layer, which is upsampled to the same spatial resolution as the main output through bilinear interpolation and participates in the total loss calculation together with the main output.

5. The method according to claim 4, characterized in that, Step two also includes: enhancing the obtained liver mask. First, the main liver region is preserved through binarization and maximum connected component filtering. Then, the obtained mask undergoes morphological dilation processing, calculated as follows: （1） First, the multiple channels of the original image are combined into a multi-channel representation for subsequent processing, where Y, U, and V correspond to the luminance and color components of the image, respectively. Then, the merged channels are linearly normalized, and the luminance channel Y is subjected to histogram equalization operation H(Y). M' represents the liver mask after morphological dilation, and ⊙ represents element-wise multiplication.

6. The method according to claim 5, characterized in that, The loss function of the segmentation model is constructed as follows: w i Represents the weights of each loss function, i = 1, 2, loss BCE Represents cross-entropy loss, loss Tversky Indicates Tversky loss, y (i) p represents the true label of the i-th sample. (i) This represents the probability that the model predicts the i-th sample to be positive, N represents the total number of samples, and α and β are parameters that weigh false positives (FP) and false negatives (FN). Let be the combined loss function of BCE loss and Tversky loss, where a is the weighting factor; With deep supervision, the total loss of the network is given by the formula: Among them, auxi loss Output the loss for each level of depth supervision, i = 1, 2, 3.

7. The method according to claim 6, characterized in that, The classification model is based on ResNet34-base and uses a hierarchical residual structure to achieve deep feature reuse and stable gradient propagation.

8. A liver tumor segmentation system based on plain CT scan, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 7, and executes the steps in the liver tumor segmentation method based on plain CT scan described above when running.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the liver tumor segmentation method based on plain CT scan as described in any one of claims 1 to 7.

Images