A method for segmenting lung cancer target volume in radiotherapy planning

Abstract

The present invention discloses a method for segmenting a lung cancer target area in a radiotherapy plan, comprising the following steps: A. importing a lung CT image file and identifying the coordinates of a white area in the image; B. creating a rectangular frame that completely surrounds the white area in the image; C. selecting a minimum rectangular frame that can completely surround all white areas, and using this rectangular frame to crop the image; D. using a SEA-Unet model to segment the cropped image; the SEA-Unet model includes an encoder and a decoder, and two parallel jump paths are set between the encoder and the decoder. The present invention can improve the deficiencies of the prior art and improve segmentation accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, in particular to a method for segmenting a lung cancer target area in a radiotherapy plan. Background Art

[0002] Lung cancer is one of the most common malignant tumors in the world, with more than 2.2 million new cases and about 1.7 million deaths each year. It is the type of cancer with the most new cases and the second highest mortality rate. According to statistics from the World Health Organization's International Agency for Research on Cancer, the number of deaths caused by lung cancer worldwide in 2021 was about 1.8 million, making it the most deadly type of cancer. In addition, data released by the National Cancer Center of China in February 2022 showed that lung cancer is not only the most common malignant tumor in China, but also the leading cause of cancer-related deaths.

[0003] Despite significant progress in multimodal treatment over the past decade, lung cancer remains one of the leading causes of death, accounting for 25% of all cancer-related deaths. Current treatment modalities include surgery, radiotherapy, chemotherapy, and immunotherapy, of which radiotherapy is one of the main methods, and approximately 60%-70% of lung cancer patients require this treatment. In particular, stereotactic body radiation therapy (SBRT) is suitable for patients with early-stage non-small cell lung cancer, which uses multiple conformal coplanar and non-coplanar beams to focus high doses of radiation on small targets in several treatments. Accurate delineation of gross tumor volumes (GTVs) is crucial for the precise creation of clinical target volumes (CTVs) and planning target volumes (PTVs).

[0004] Traditional manual contouring methods require radiation oncologists to delineate the tumor region based on computed tomography (CT) images. However, manual delineation may lead to observer variation in GTV contours, affecting the accuracy of treatment planning and may cause random and systematic changes in treatment plans, thereby affecting the distribution of treatment dose. Therefore, automatic segmentation methods are becoming increasingly important in clinical SBRT.

[0005] With the rapid development of medical imaging and computer technology, tumor radiotherapy has entered the era of high-precision image guidance. In radiotherapy planning, accurate delineation of target areas and organs at risk (OARs) is particularly critical. Despite the availability of unified principles and consensus for reference, manual delineation of radiotherapy targets still relies mainly on the experience of practitioners, which is not only highly variable and time-consuming, but may also affect the effectiveness of radiotherapy. The application of artificial intelligence technology, especially convolutional neural networks (CNNs), a deep learning technology, has significantly improved the efficiency and consistency of target delineation in radiotherapy due to its insensitivity to image noise, blur and contrast. In the field of oncology radiology, trained CNN models can be accelerated by graphics processing units (GPUs) to quickly complete the segmentation tasks of gross tumor volume (GTV) and normal tissues and organs.

[0006] Although deep learning-based automatic segmentation techniques have been rapidly applied to the delineation of GTV in lung cancer radiotherapy, there is still a lack of research on automatic segmentation in lung cancer radiotherapy planning. Current challenges include the low contrast between tumors and surrounding tissues in medical images, which greatly increases the difficulty of automatic identification and segmentation; the relatively small proportion of tumor areas in lung CT images, which leads to an unbalanced dataset and directly affects the performance of automatic segmentation algorithms; and the high heterogeneity within lung cancer tumors, such as cystic areas, calcifications, and necrosis, which makes the automatic segmentation task more complicated. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a method for segmenting a lung cancer target area in a radiotherapy plan, which can solve the deficiencies of the prior art and improve segmentation accuracy.

[0008] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.

[0009] A method for segmenting a lung cancer target area in a radiotherapy plan comprises the following steps:

[0010] A. Import the lung CT image file and identify the coordinates of the white area in the image;

[0011] B. Create a rectangular box that completely surrounds the white area in the image;

[0012] C. Select the smallest rectangular frame that can completely surround all white areas and use this rectangular frame to crop the image;

[0013] D. Use the SEA-Unet model to segment the cropped image; the SEA-Unet model includes an encoder and a decoder, and two parallel jump paths are set between the encoder and the decoder.

[0014] Preferably, the encoder comprises:

[0015] 8 convolutional units, each of which includes a 4x4 convolutional layer, a batch normalization layer, and a ReLU activation function to expand the receptive field;

[0016] The convolution unit uses convolution downsampling with a stride of 2 to reduce the loss of feature information.

[0017] Preferably, the decoder uses eight upsampling steps to restore the image resolution. The upsampling step first scales the deep feature map by methods such as nearest neighbor interpolation, and then fuses the scaled feature map with the feature map of the corresponding scale in the encoding path as input for the next upsampling step.

[0018] Preferably, the jumping paths are respectively an attention module path and a compression excitation module path, and the attention module and the compression excitation module share the same input at the same level.

[0019] Preferably, before entering the attention module, the deep feature map is expanded by deconvolution with a kernel size of 4x4 and a stride of 2, and the deep feature map uses its deep semantic information to guide the redistribution of feature weights in the attention module path at the same level as the shallow feature map; at the same time, the shallow feature map is also used as the input of the compression excitation module path, where the channel information weight is adjusted by the compression excitation module; finally, the output from the attention module path is combined with the output from the compression excitation module path to form the input of the next attention module path.

[0020] Preferably, the attention module path adjusts the weight ratio between the target area and the non-target area through deep semantic analysis, emphasizes the target area and provides key feature information for the model; the deep feature map and the shallow feature map first pass through a convolution layer with a 1x1 convolution kernel, and then adjust the number of channels and feature mixing through a batch normalization layer; then, the deep feature map and the shallow feature map are superimposed to highlight the focus of model learning, and the non-target area is suppressed through the ReLU activation function; the attention coefficient is adjusted to between 0 and 1 through the convolution layer, batch normalization layer and Sigmoid activation layer; finally, the generated attention coefficient matrix is ​​multiplied with the original shallow feature map to generate a new feature map whose weights are optimally adjusted to enhance the model's ability to identify important features.

[0021] Preferably, the compression excitation module path embeds the SE module into the residual block, including a convolution path and a shortcut connection. The convolution path consists of three convolution blocks, each of which includes a convolution layer, a batch normalization layer and a ReLU activation function layer. The SE module is located between the batch normalization layer and the shortcut connection of the third convolution block, and is used to analyze the relationship between feature channels. The SE module automatically learns the importance of different channel features, wherein the squeezing operation averages the two-dimensional feature map of each channel through a global average pooling layer to eliminate spatial information interference; then the extraction operation is completed through two fully connected layers and a Sigmoid activation layer; the channel weight coefficient is trained in the fully connected layer to reflect the correlation between channels; the output of the Sigmoid layer is multiplied by the output of the last batch normalization layer in the residual block; the shortcut connection adds the shallow feature map to the SE module output, which is then processed through the ReLU activation function layer.

[0022] Preferably, the SEA-Unet model uses a hybrid loss function obtained by a weighted combination of binary cross entropy and Dice loss function.

[0023] The binary cross entropy (BCE) loss function is,

[0024]

[0025] Among them, y i To predict the mask, is the label, N is the total amount of training;

[0026] The Dice loss function is,

[0027]

[0028] Where X is the label, Y is the predicted mask, |X∩Y| is the intersection between the label and the predicted mask, |X| is the number of label samples, and |Y| is the number of predicted mask samples;

[0029] The binary cross entropy loss accounts for α, the Dice loss accounts for 1-α, and the mixed loss function is,

[0030] L=αL bce +(1-α)L dice .

[0031] Preferably, the α is 0.9.

[0032] The beneficial effects of adopting the above technical solution are: the present invention uses RAU-Net to accurately locate the GTV area by using object detection in ROI extraction, thereby improving segmentation accuracy. The multi-scale feature processing and segmentation accuracy are improved by optimizing the U-Net model equipped with parallel skip connections and integrated attention and SE-Res modules. By combining binary cross entropy and Dice loss functions, the model effectively handles the imbalance problem and enhances pixel-level classification for small area recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is the SEA-Unet model architecture diagram of the present invention.

[0034] Figure 2 This is the attention module path structure diagram of the present invention.

[0035] Figure 3 This is a path structure diagram of the compression excitation module of the present invention.

[0036] Figure 4 The figure is a comparison diagram of the segmentation results of the present invention and the existing image segmentation method. DETAILED DESCRIPTION

[0037] In the field of medical image analysis, the U-net architecture and its derivatives represent a major leap forward in image segmentation technology. Since the introduction of U-net, its basic structure has been widely used and continuously improved to meet various medical image processing needs. UNet++ is a significant enhancement to the U-net framework, using a deeply supervised encoder-decoder network and a redesigned dense skip path to effectively reduce the semantic differences in feature maps between the encoder and decoder subnetworks, and outperform the original U-Net and wide U-Net in segmentation performance. In addition, Swin-Unet represents an innovative attempt to integrate the Transformer architecture into U-net. This pure Transformer-based U-net model, using the U-shaped encoder-decoder structure and skip connections, learns local-global semantic features, and outperforms traditional pure convolutional methods in multi-organ and heart segmentation tasks. Specifically targeting the challenge of medical image segmentation, nn-UNet demonstrated its superiority at the 2021 International MICCAI Brainlesion Workshop. The research expands on the winning work of the previous year's competition, including an increase in network size, replacement of batch normalization with group normalization, and application of axial attention in the decoder. These improvements achieve remarkable results, reaching Dice scores of 88.35%, 88.78%, and 93.19% on brain tumor segmentation for the enhancing tumor, tumor core, and entire tumor, respectively.

[0038] U-net and its variants are widely used in various medical image segmentation tasks. and KemalPolat introduced a novel polyp segmentation method combining U-Net with visual saliency features, which surpassed the existing state-of-the-art methods and significantly improved the accuracy of colonoscopy polyp detection. and colleagues developed a novel model combining image processing and deep learning in the field of automatic diatom image segmentation. Their model outperformed previous methods and included a publicly available diatom image dataset. In the field of retinal blood vessel segmentation, M SHABANI and his team proposed an improved active contour model combined with discrete wavelet transform, and its performance was comparable to the state-of-the-art methods on multiple databases.

[0039] In this context, the integration of attention mechanisms and residual networks has brought deeper optimization to the U-net architecture. For example, attention mechanisms such as SE, CBAM, and GCNet improve the model's ability to identify target areas and segmentation accuracy by enhancing the feature responses of specific regions or channels. The residual network (Resnet) solves the problems of network degradation and gradient vanishing in deep networks by introducing jump connections, allowing the model to ensure effective training and accurate information transfer while maintaining depth. The fusion of these technologies not only improves the performance of U-net in medical image segmentation tasks, but also improves its ability to handle complex structures, demonstrating the power of deep learning in the field of medical image processing.

[0040] We introduce a two-stage RAU-Net network, where the first stage preprocesses the image to be segmented using the idea of ​​object detection, the second stage applies our modified U-Net variant model for fine segmentation, and finally restores the result to the original image size output.

[0041] In the first stage, we accurately locate the target frame through a customized region of interest (ROI) extraction network. The network first imports the label files of all training images. Subsequently, we accurately identify the coordinates of the white areas in these labeled images. Based on these coordinates, we create and update a specially designed optimal rectangular box to completely surround the white areas in the image. In addition, we conduct a comprehensive analysis of the rectangular boxes of all training labels to determine a minimum rectangular box that can completely surround all the training image areas to be segmented. By accurately cropping the image with this rectangular box, we can highlight the key areas for segmentation while significantly reducing background noise, effectively improving the accuracy and clarity of segmentation. In addition, based on two observations, the size of the rectangular box was finally set to 256×256: first, the size of the chest part is large enough to be included in the size of 256×256; second, this size perfectly meets the requirements of the second-stage deep network. The rectangular box obtained in this study is (194,161,450,417).

[0042] In the second stage, based on the target frame, we crop the image area to be segmented and apply the proposed SEA-Unet model for fine segmentation. After the segmentation is completed, we use the cropped coordinates of the first stage to restore the segmentation result to a size of 512×512. This model adopts a codec architecture. The specific network structure is as follows Figure 1 As shown below. We will describe the encoder, decoder, and two skip paths of the SEA-Unet model in detail.

[0043] exist Figure 1In the encoder shown in the upper part of , we integrate eight convolution units, each of which includes a 4x4 convolution layer, batch normalization (BN), and ReLU activation function to expand the receptive field. In order to reduce the loss of feature information, convolution downsampling with a stride of 2 is used instead of the traditional pooling operation. The number of convolution layers in this encoder is twice that of the standard U-Net to meet the high demand for multi-scale features in medical image segmentation. Through layered convolution, the encoder captures features at different scales, and the shallow feature maps focus on capturing detailed information such as small object attributes and boundary characteristics; the deep feature maps are more effective in coarse positioning, but less in detailed processing. The design of the decoder emphasizes the effective integration of these multi-scale feature maps to achieve more accurate image segmentation results.

[0044] The decoder displays Figure 1 The lower half of the encoder forms a symmetrical structure with the encoder in the network. After the encoder has been downsampled eight times, the decoder must upsample eight times to restore the image resolution. In the standard U-Net model, the upsampling process first scales the deep feature maps through methods such as nearest neighbor interpolation, and then these expanded feature maps are fused with the feature maps of the corresponding scale in the encoding path as the input of the next upsampling step. Due to the significant semantic differences between deep and shallow feature maps, and the loss of image details caused by the increase in network depth, traditional simple skip connections are not enough to solve these problems. Our SEA-Unet model innovatively designs two parallel skip connection paths in the decoder to effectively address these challenges.

[0045] Figure 2 The structure of the attention module is shown in detail, and Figure 3 Figure 1 shows the complexity of the SE-Res module. The two modules share the same input at the same level: G represents the deep feature map and x represents the shallow feature map. Before entering the attention module, the deep feature map G is expanded by using deconvolution with a kernel size of 4x4 and a stride of 2. G uses its deep semantic information to guide the redistribution of feature weights in the attention path at the same level as x. At the same time, x is also used as the input of the SE-Res path, where the channel information weights are adjusted by the SE module. Finally, the output X' from the attention path is combined with the output X" from the SE-Res path to form the input G of the next attention path. This structural design not only avoids the direct connection between the shallow and deep feature maps, thereby reducing the feature fusion problem caused by the semantic difference in U-Net, but also ensures that the small object features that may be lost in the attention module are compensated by adding shallow feature maps. This strategy not only prevents the dilution of the semantic information of small objects, but also effectively utilizes multi-scale features to preserve richer shallow semantic information.

[0046] SA module (attention module): The added attention path not only promotes the fusion of deep semantic information in the decoding path, but also effectively reduces the noise in medical images. Through deep semantic analysis, this path adjusts the weight ratio between the target area and the non-target area, emphasizes the target area and provides key feature information for the model. Figure 2-3 In the above example, x and G first pass through a convolutional layer with a 1x1 convolution kernel, and then adjust the number of channels and feature mixing through a BN layer. Subsequently, the two feature maps are superimposed to highlight the focus of model learning, and non-target areas are suppressed through the ReLU activation function. The attention coefficient is adjusted between 0 and 1 through the convolutional layer, BN layer, and Sigmoid activation layer. Finally, the generated attention coefficient matrix is ​​multiplied with the original x to generate a new feature map whose weights are optimally adjusted to enhance the model's ability to identify important features.

[0047] SE-Res module (combination of residual and SE modules): Figure 3 As shown in Figure 1, the internal structure of the SE-Res module embeds the SE module into a residual block, which includes a convolution path and a shortcut connection. The convolution path consists of three convolution blocks, each of which includes a convolution layer, a batch normalization layer, and a ReLU activation function layer. The SE module is located between the BN layer and the shortcut connection of the third convolution block, and is specifically used to analyze the relationship between feature channels. Through the SE block, the module automatically learns the importance of different channel features, where the squeezing operation averages the two-dimensional feature map of each channel through the global average pooling layer to eliminate spatial information interference. The extraction operation is then completed through two fully connected layers and a Sigmoid activation layer. The first fully connected layer reduces the number of channels from n to n / 8 to reduce the computational burden, while the second restores them. The channel weight coefficients are trained in the fully connected layer and reflect the correlation between channels. The output of the Sigmoid layer is multiplied by the output of the last BN layer in the Resblock. Due to the semantic difference between the shallow feature maps in the encoder and the deep feature maps in the decoder, the SE block is not independently set on the skip connection. Instead, it adds the input feature map x to the SE output via a shortcut connection and then processes it through a ReLU activation layer.

[0048] We used a hybrid loss function that combines the binary cross entropy (BCE) and Dice loss functions to optimize the SEA-UNet model. This method was chosen for several reasons: First, the binary cross entropy (BCE) loss function performs well when dealing with pixel-level binary classification problems, and effectively evaluates the difference between the model output probability and the true label. Second, the Dice loss function is particularly suitable for dealing with problems with unbalanced datasets, such as small regions or rare events in medical images. It helps improve the accuracy of the model in identifying small target areas by calculating the overlap between the predicted segmentation and the true segmentation. The integration of these two loss functions not only utilizes their respective advantages, but also complements each other, thereby improving the performance of the SEA-UNet model in complex medical image processing tasks. Formula of the binary cross entropy (BCE) loss function

[0049]

[0050] Among them, y i To predict the mask, is the label and N is the total amount of training.

[0051] The Dice loss function is a complement to the Dice coefficient and is used for optimization during training. It is often used to evaluate the degree of difference between two samples, especially in image segmentation. The calculation formula of the Dice coefficient is as follows,

[0052]

[0053] Where X is the label, Y is the predicted mask, |X∩Y| is the intersection between the label and the predicted mask, |X| is the number of label samples, and |Y| is the number of predicted mask samples.

[0054] In order to improve the accuracy and comprehensiveness of model evaluation, this study introduces a hybrid loss function, called combined loss (CLoss), which is achieved by fusing two existing loss functions. Specifically, this new loss function consists of a weighted combination of the two, where the binary cross entropy (BCE) loss accounts for α and the Dice loss accounts for 1-α (α=0.9 in this model). This combination aims to synergistically utilize the advantages of the two loss functions, thereby promoting a more comprehensive and effective model training process. The specific formula is

[0055] L=αL bce +(1-α)L dice .

[0056] The model training process is as follows:

[0057] The lung cancer segmentation model in this study was developed on the Pytorch platform and deployed on the NVIDIA RTX3080ti GPU. The environment used included Cuda 11.7, Pytorch 2.0.1, Windows 10, and VsCode1.75.1. All cross-sectional 2D CT image data were divided into training and validation sets in a 4:1 ratio. The training set contained 1139 images and the validation set contained 302 images. The resolution of each image was 512×512. We used Adam as the optimizer, and the initial learning rate was set to 1×10 -3 , the minimum learning rate is 1×10 -8 , the weight decay is 1×10 -5 , with a momentum of 0.9. If the validation set loss does not decrease for two consecutive training epochs, we halve the learning rate. The entire training process consists of 200 epochs with a batch size of 16. On our dataset, the RAU-Net model takes about three hours to complete training.

[0058] The model was developed using radiological data from 114 lung cancer patients with a median age of 64 years (range, 40-89 years) who received radiotherapy between October 2008 and September 2023.

[0059] The images used for analysis were obtained from CT scans of patients for radiotherapy positioning. The enhanced CT scans were performed using a SOMATOM Definition AS20-slice CT simulator system produced by Siemens (Germany), with the following scanning parameters: tube voltage 120 kVp; tube current 165 mAs; the upper limit of the scanning range was the mid-neck, and the lower limit was the upper abdomen (including the esophagus and the entire lung). The acquired images had a pixel size of 512 × 512, the scanning slice thickness was set to 5 mm, and the field of view (FOV) was 400 mm. After scanning, the reconstructed CT images were transferred to a dedicated radiotherapy three-dimensional treatment planning system, Pinnacle3 version 9.16. The features of each tissue were extracted after wavelet transform filtering preprocessing.

[0060] In our study, the performance of comparing various methods was evaluated by Dice coefficient as well as specificity and sensitivity metrics. These metrics provide a comprehensive framework for quantitatively evaluating the accuracy and reliability of medical image segmentation. The Dice coefficient plays an important role in evaluating the consistency between the segmentation results and the actual images. In contrast, specificity and sensitivity provide unique insights into the model's ability to correctly identify negative and positive samples.

[0061] Dice coefficient:

[0062] Sensitivity: Sensitivity measures the proportion of pixels that are actually brain tissue that are correctly identified. It indicates how effective the segmentation algorithm is in accurately detecting brain tissue.

[0063] Specificity:

[0064] Table 1 shows the experimental evaluation results of various image segmentation methods on the lung cancer rough tumor volume (GTV) target region dataset. The evaluation includes the baseline model U-Net and its variants UNet++ and UNet++Deep Supervision, as well as advanced models such as nnU-Net, SwinU-Net and our proposed RAU-Net. The analysis shows that RAU-Net leads with a Dice coefficient of 77.13%, a significant improvement of 28.34% compared to the baseline U-Net. In addition, RAU-Net is 16.50% higher than the highly adaptive nnU-Net and 4.1% higher than SwinU-Net, which is famous for its excellent global feature extraction capabilities. These results fully demonstrate the significant advantages of RAU-Net in accurately capturing details and integrating global information.

[0065] In terms of the Specificity metric, all evaluated networks consistently maintained above 99%, indicating their excellent performance in correctly distinguishing between background and non-target areas, effectively avoiding mislabeling non-target areas as targets. Crucially, from the perspective of sensitivity, RAU-Net significantly outperformed the baseline model U-Net, scoring 80.38%, more than double the improvement of U-Net's 57.42%. Even compared to the second-best SwinU-Net, there was a 6.25% improvement. This significant difference highlights RAU-Net's superior ability to accurately identify and label target areas, especially in reducing false negatives and improving the accuracy of medical image analysis.

[0066] Table 1

[0067]

[0068]

[0069] exist Figure 4In the experiment, we observed that the segmentation results of RAU-Net are very consistent with the actual situation and accurately delineate the boundaries of the gross tumor volume (GTV). In contrast, other segmentation methods such as U-Net, U-Net++, SwinUN-Net, UNet++Deep Supervision and nnU-Nnet are not satisfactory in dealing with over-segmentation and under-segmentation problems. The dataset used in this experiment includes 2D CT image slices of the target area of ​​lung cancer GTV, which are usually located in very small lung regions and need to be accurately segmented. The small proportion of the target area in the entire image increases the challenge of maintaining the integrity of the target information.

[0070] Specifically, in row A, only RAU-Net succeeds in achieving accurate segmentation, while U-Net exhibits significant over-segmentation, U-Net++ and SwinUN-Net fail to capture the target region, and UNet++Deep Supervision and nnU-Nnet exhibit false positives. In row B, U-Net++ severely under-segmentates, while U-Net exhibits false positives. Although nnU-Nnet and SwinUN-Net perform better, they are still inferior to RAU-Net in segmentation details and accuracy. In rows C and D, although all networks demonstrate some degree of segmentation, they are still inferior to RAU-Net in edge and detail processing. Finally, in row E, UNet and UNet++ significantly miss the target region, and UNet++Deep Supervision, nnU-Nnet, and SwinUN-Net also have inferior segmentation details and accuracy compared to RAU-Net. In row F, no method achieves correct segmentation, indicating that our method has considerable room for improvement when dealing with complex segmented regions. These observations highlight the superior performance of RAU-Net in handling details and edge localization, especially its ability to accurately segment small regions.

[0071] To verify the effectiveness of the first stage (region of interest extraction), attention mechanism (Att), SE-Res module and combined loss function (CLoss) in medical image processing, this study designed a series of ablation experiments. Different network configurations were evaluated on our custom dataset, and the detailed results are listed in Table 2. It is worth noting that all network models achieved excellent levels in specificity indicators. However, significant differences were observed between different network structures in two key indicators, Dice coefficient and sensitivity.

[0072] Table 2

[0073]

[0074]

[0075] Initially, by comparing experiments with and without ROI (N-ROI), we observed that in most cases, the model optimized for ROI significantly outperformed the N-ROI model in terms of Dice coefficient and sensitivity. Specifically, in the baseline model, the Dice coefficient and sensitivity of ROI were 64.22% and 63.78%, respectively, while those of N-ROI were 58.63% and 58.21%, respectively, an improvement of 5.59% and 5.57%, respectively. In addition, the Dice coefficient and sensitivity of the RAU-Net model in terms of ROI reached 77.13% and 80.38%, respectively, compared to 72.03% and 78.26% of N-ROI, an improvement of 5.1% and 2.12%, respectively. This result emphasizes the importance of specifically treating the region of interest in improving the segmentation accuracy and sensitivity of the model.

[0076] Next, we analyze the impact of the attention mechanism on model performance and find that it has a positive effect. Although adding the attention mechanism did not significantly improve performance in the N-ROI scenario, in the ROI scenario, it increased the Dice coefficient of the baseline model from 64.22% to 67.46% and the sensitivity from 63.78% to 65.90%, an increase of 3.24% and 2.12% respectively. This finding verifies the effectiveness of the attention mechanism in focusing model resources on more critical areas.

[0077] Finally, when the SE module, attention mechanism, and combined loss function (CLoss) are applied simultaneously, the Dice coefficient and sensitivity are significantly improved for both N-ROI and ROI. For example, for N-ROI, the Dice coefficient and sensitivity of Baseline+SE+Att+CLoss are improved to 72.03% and 78.26%, which are significantly higher than using SE or Att alone; in ROI, the combination of Baseline+SE+Att+Closs reaches up to 77.13% and 80.38%, significantly exceeding any other single combination. This result reveals the key role of the loss function in further optimizing model performance.

[0078] We introduced the RAU-Net model, which focuses on high-precision segmentation of lung cancer GTV regions in CT images. RAU-Net achieved a remarkable 77.13% in Dice coefficient, which is 28.34% higher than the 48.79% of the traditional U-Net model and even 4.1% higher than the advanced SwinU-Net model. This breakthrough highlights the superiority of RAU-Net in fine processing of details and accurate positioning of edges, especially in the precise segmentation of small tumor regions. In terms of sensitivity evaluation, RAU-Net outperformed other models with a score of 80.38%, once again confirming its effectiveness in identifying and marking target regions. Our ablation study deeply analyzes the contribution of each component of RAU-Net and reveals the importance of integrating the ROI extraction stage, attention mechanism, SE-Res module and combined loss function (CLoss) to improve the model performance. The addition of the ROI extraction stage alone improves the Dice coefficient by 5.1%, highlighting the model's precise focus on the region of interest and effective reduction of background interference, thereby improving the model's segmentation accuracy and efficiency. In addition, the combination of the attention mechanism, SE-Res module and CLoss further improves the accuracy of the model in segmenting complex image features, thanks to their synergistic effects in feature emphasis and key area labeling.

[0079] The RAU-Net model disclosed in the present invention successfully achieves the goal of accurately segmenting the lung cancer GTV area in CT images. By integrating ROI extraction, attention mechanism, SE-Res module and combined loss function, RAU-Net shows excellent performance in improving segmentation accuracy and processing details, especially in the segmentation of small and complex areas.

[0080] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for segmenting a lung cancer target area in a radiotherapy plan, characterized in that The following steps are involved: A. Import the lung CT image file and identify the coordinates of the white area in the image; B. Create a rectangular box that completely surrounds the white area in the image; C. Select the smallest rectangular frame that can completely surround all white areas and use this rectangular frame to crop the image; D. Use the SEA-Unet model to segment the cropped image; the SEA-Unet model includes an encoder and a decoder, and two parallel jump paths are set between the encoder and the decoder; The encoder includes 8 convolution units, each of which includes a 4x4 convolution layer, a batch normalization layer and a ReLU activation function for expanding the receptive field, and the convolution unit uses convolution downsampling with a step size of 2 to reduce the loss of feature information; The decoder uses eight upsampling steps to restore the image resolution. The upsampling first scales the deep feature map by the nearest neighbor interpolation method, and then fuses the scaled feature map with the feature map of the corresponding scale in the encoding path as the input of the next upsampling step. The jump paths are respectively an attention module path and a compression excitation module path, and the attention module and the compression excitation module share the same input at the same level; Before entering the attention module, the deep feature map is expanded by using deconvolution with a kernel size of 4x4 and a stride of 2. The deep feature map uses its deep semantic information to guide the redistribution of feature weights in the attention module path at the same level as the shallow feature map; at the same time, the shallow feature map is also used as the input of the compression-excitation module path, where the channel information weight is adjusted by the compression-excitation module; finally, the output from the attention module path is combined with the output from the compression-excitation module path to form the input of the next attention module path; The attention module path adjusts the weight ratio between the target area and the non-target area through deep semantic analysis, emphasizes the target area and provides key feature information for the model; the deep feature map and the shallow feature map first pass through the convolution layer with a 1x1 convolution kernel, and then adjust the number of channels and feature mixing through the batch normalization layer; Subsequently, the deep feature map and the shallow feature map are superimposed to highlight the focus of model learning, and the non-target area is suppressed through the ReLU activation function; The attention coefficient is adjusted to between 0 and 1 through the convolution layer, batch normalization layer and sigmoid activation layer; finally, the generated attention coefficient matrix is ​​multiplied with the original shallow feature map to generate a new feature map whose weights are optimally adjusted to enhance the model's ability to identify important features; The compression-excitation module path embeds the SE module into the residual block, including a convolution path and a shortcut connection. The convolution path consists of three convolution blocks, each of which includes a convolution layer, a batch normalization layer and a ReLU activation function layer. The SE module is located between the batch normalization layer and the shortcut connection of the third convolution block, and is used to analyze the relationship between feature channels. The SE module automatically learns the importance of different channel features, where the squeezing operation averages the two-dimensional feature map of each channel through the global average pooling layer to eliminate spatial information interference; then the extraction operation is completed through two fully connected layers and a Sigmoid activation layer; the channel weight coefficient is trained in the fully connected layer to reflect the correlation between channels; the output of the Sigmoid layer is multiplied by the output of the last batch normalization layer in the residual block; the shortcut connection adds the shallow feature map to the SE module output, which is then processed through the ReLU activation function layer.

2. The method for segmenting lung cancer target area in radiotherapy planning according to claim 1, characterized in that: The SEA-Unet model uses a hybrid loss function obtained by a weighted combination of binary cross entropy and Dice loss functions. The binary cross entropy (BCE) loss function is, Among them, y i To predict the mask, is the label, N is the total amount of training; The Dice loss function is, Where X is the label, Y is the predicted mask, |X∩Y| is the intersection between the label and the predicted mask, |X| is the number of label samples, and |Y| is the number of predicted mask samples; The binary cross entropy loss accounts for α, the Dice loss accounts for 1-α, and the mixed loss function is, L=αL bce +(1-α)L dice 。 3. The method for segmenting lung cancer target area in radiotherapy planning according to claim 2, characterized in that: The α is 0.9.

Images