A bladder tumor segmentation method and system based on parallel branch focused linear attention

Abstract

This invention discloses a bladder tumor segmentation method and system based on parallel branch-focused linear attention, belonging to the field of bladder tumor segmentation technology. The method includes: a main branch extracting specific morphological and structural features from cystoscopy images, combined with a novel focused linear attention module to achieve high efficiency and expressiveness; a fast convolutional branch capturing spatial details to compensate for the limitations of the main branch in full-size prediction; a spatial channel reconstruction convolutional residual module reducing redundant computation and promoting representative feature learning, thus lowering complexity and computational cost; and a prediction head connecting the main branch and the fast convolutional branch, fusing the coarse-grained features extracted by the main branch with the fine-grained features extracted by the fast convolutional branch to predict the segmentation map. This invention improves the morphological specificity of bladder tumors, the distinction between bladder tumors and normal tissues, and the problem of blurred bladder tumor boundaries, significantly improving computational efficiency.

Description

Technical Field

[0001] This invention relates to the field of tumor segmentation technology, and more specifically to a method and system for bladder tumor segmentation based on parallel branch-focused linear attention. Background Technology

[0002] Bladder cancer is a common malignant tumor of the urinary system. Diagnosis and treatment of bladder cancer primarily rely on the layering of the bladder wall and the staging of the tumor. High-grade T1 bladder cancer is a non-muscle-invasive bladder cancer with a high risk of recurrence and progression, with a five-year recurrence rate of 42% and a progression rate of 20-40%. Despite these worrying statistics, bladder cancer is one of the cancer types with the highest survival rates. In cases of early detection, the cancer remains confined to the bladder, and 94% of patients have been successfully treated and survive. However, for later detection, the survival rate drops significantly to 6%. Therefore, accurately identifying and segmenting the various layers of the bladder wall (such as the epithelial layer, lamina propria, muscular layer, and muscular mucosa) as well as the location and size of the tumor is crucial for early detection, auxiliary diagnosis, and treatment selection of bladder cancer, effectively reducing morbidity, mortality, and costs.

[0003] Currently, commonly used medical image types for bladder tumor detection and analysis include magnetic resonance imaging (MRI), computed tomography (CT), and cystoscopy. However, due to problems such as noise, blurring, low contrast, morphological diversity, and changing perspectives, bladder tumor image segmentation is a challenging task. Furthermore, different methods have different applicability and performance on different types of images. Cystoscopy is a medical examination method that directly observes the inside of the bladder. Compared to imaging examinations such as CT and MRI, it is more intuitive, allowing direct observation of the bladder's internal condition, including the size, shape, color, and location of tumors, which is crucial for diagnosis and treatment planning; it is more real-time, allowing real-time observation of the bladder, while CT and MRI only provide static images; it is more operable, allowing doctors to perform procedures such as biopsies and tumor resections; and it is more economical and convenient, with a lower cost compared to equipment like CT and MRI.

[0004] Cystoscopy is an endoscopic technique used to observe and treat the bladder and urethra; it is a safe and effective method. It can help diagnose and treat conditions such as hematuria, urinary tract infections, bladder tumors, bladder stones, and benign prostatic hyperplasia. Transurethral resection of bladder tumors (TURBT) is a surgical procedure used to diagnose and treat bladder tumors, and is an alternative to open surgery. TURBT involves inserting a cystoscope through the urethra into the bladder and removing the tumor from the bladder wall. TURBT is also a safe and effective method for diagnosing and treating non-muscle-invasive bladder cancer. This type of cancer grows only on the surface of the bladder wall and does not invade deeper tissues. Both cystoscopy and TURBT produce white light images during the procedure. Studies have shown that widely used white light imaging cystoscopy has a bladder cancer identification rate of 68%–79.2%. Therefore, accurate identification and segmentation of bladder tumors under cystoscopy has significant clinical importance.

[0005] In recent years, deep learning, as a powerful machine learning technique, has made significant progress in the analysis of natural and medical images. Deep learning can automatically learn high-level feature representations from large amounts of data, thereby improving the accuracy and robustness of tumor image segmentation. However, due to some special characteristics of cystoscopy-acquired images, deep learning-based bladder tumor segmentation research has not yet been well resolved. For example, bladder cancer images have problems such as unclear tumor boundaries, small tumor regions, and discontinuous tumor distribution. Existing segmentation algorithms have a large number of parameters, are computationally complex, and their segmentation accuracy needs improvement. In addition, medical images generally contain complex noise and blur, making it difficult for traditional image segmentation algorithms, which require manually designing features based on prior knowledge, to obtain accurate segmentation results.

[0006] Meanwhile, due to the pathological characteristics of bladder tumors and various imaging methods, the detected bladder tumor images possess features that distinguish them from images of other diseases, including their unique geometry, spatial structure, and other characteristics. However, in the current field of bladder tumor segmentation research, the amount of relevant work is relatively small, and these characteristics of bladder tumor images have not been well utilized. Currently, there are three key problems in the field of bladder tumor segmentation technology: First, due to the morphological specificity of bladder tumors, tumors may be nodular, cauliflower-like, lobed, or irregular in shape, which increases the complexity of image segmentation. Second, there is the problem of distinguishing tumors from normal tissue; during cystoscopy, it is sometimes difficult to distinguish healthy normal bladder tissue from tumor tissue. Finally, there is the blurring of bladder tumor boundaries: in images of bladder cancer, the boundaries of the tumor may appear unclear. This may be due to insufficient contrast between the tumor and surrounding normal tissue, or due to the complex morphology of the tumor, such as nodular or cauliflower-like shapes, making the boundaries difficult to determine. Additionally, in some cases, bladder cancer may present as a small tumor area or a discontinuous distribution. In such cases, the tumor may be missed or confused with surrounding normal tissue, making the boundaries even more difficult to determine.

[0007] Therefore, how to propose a bladder tumor segmentation method and system based on parallel branch-focused linear attention, and how to distinguish the morphological specificity of bladder tumors, the boundaries between bladder tumors and normal tissues, and the boundaries of blurred bladder tumors during the bladder tumor segmentation process are problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention provides a bladder tumor segmentation method and system based on parallel branch-focusing linear attention, which solves the problems of bladder tumor morphological specificity, differentiation between bladder tumors and normal tissues, and blurred bladder tumor boundaries. To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A bladder tumor segmentation method based on parallel branch-focusing linear attention includes:

[0010] The cystoscopy images were cropped, filtered, and labeled to obtain a labeled bladder tumor dataset;

[0011] A bladder tumor segmentation model is constructed, comprising a backbone extraction network branch for extracting global backbone multi-scale features from cystoscopy images, a fast convolutional branch parallel to the backbone extraction network branch for capturing spatial features, and a result prediction head connected to the backbone extraction network branch and the fast convolutional branch for fusing the features of the two branches to obtain the final prediction result.

[0012] The bladder tumor segmentation model was optimized using a labeled bladder tumor dataset;

[0013] Real-time cystoscopy images are acquired and input into the bladder tumor segmentation model to obtain bladder tumor segmentation results.

[0014] Optional steps include: acquiring cystoscopy surgical video footage and performing non-destructive cropping; filtering images based on quality to obtain a bladder tumor dataset; annotating the images; constructing an annotated bladder tumor dataset; and dividing the dataset into training and testing sets in an 8:2 ratio.

[0015] Optionally, the fast convolution branch comprises three parts: the first part, the encoder part, consists of a branch network of a FasterNet network architecture, which performs progressive spatial downsampling and channel expansion to acquire information; the second part, the intermediate block, consists of two spatial channel reconstruction convolutional residual modules, which are used to extract features and connect the encoder and decoder; the third part, the decoder part, consists of multiple spatial channel reconstruction convolutional residual modules and upsampling layers, which are used to reconstruct the original image from the encoded features. Each decoder contains a spatial channel reconstruction convolutional residual module, which performs upsampling through the upsampling layer to finally obtain the feature map extracted by the fast convolution branch.

[0016] Optionally, the FasterNet network architecture employs partial convolution and partial point convolution. Partial convolution applies regular convolution to some input channels for spatial feature extraction, while the remaining channels remain unchanged. Partial point convolution adds point convolution to the partial convolution.

[0017] Optionally, the fast convolutional branch adopts an upsampling restoration strategy with staggered skip connections. It starts to run in parallel based on the results of the two layers of the intermediate block, connects and upsamples based on the feature pyramid of each layer in the encoder stage, connects the parallel results, and merges them with the original input through residual connections to obtain the final fast convolutional branch prediction feature map.

[0018] Optionally, the spatial channel reconstruction convolutional residual module includes a group normalization layer, a SiLU activation function, a convolutional layer, and a spatial channel reconstruction convolutional layer, wherein the spatial channel reconstruction convolutional layer is composed of spatial reconstruction units and channel reconstruction units.

[0019] Optionally, the loss function of the bladder tumor segmentation model is the sum of the binary cross-entropy loss and the Dice loss, and the loss function is defined as follows:

[0020] L = L BCE +L Dice ;

[0021] Among them, L BCE and L DiceThese represent the binary cross-entropy loss and the Dice loss, respectively.

[0022] Optionally, the method may also include: using qualitative and quantitative analysis, as well as ablation analysis, to evaluate the performance of the bladder tumor segmentation model, wherein the quantitative analysis includes: using the average dice similarity coefficient, average crossover ratio, average precision, and average recall to quantitatively evaluate the performance of the bladder tumor segmentation model.

[0023] Optionally, a bladder tumor segmentation system based on parallel branch-focusing linear attention includes:

[0024] Data construction module: used to crop, filter, and label cystoscopy images to obtain an labeled bladder tumor dataset;

[0025] Segmentation Model Construction Module: Used to construct a bladder tumor segmentation model, which includes a backbone extraction network branch for extracting global backbone multi-scale features from cystoscopy images, a fast convolutional branch parallel to the backbone extraction network branch for capturing spatial features, and a result prediction head connected to the backbone extraction network branch and the fast convolutional branch for fusing the features of the two branches to obtain the final prediction result.

[0026] Model optimization module: used to optimize the bladder tumor segmentation model using a labeled bladder tumor dataset;

[0027] Real-time segmentation module: used to acquire real-time cystoscopy images, input them into the bladder tumor segmentation model, and obtain bladder tumor segmentation results.

[0028] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a bladder tumor segmentation method and system based on parallel branch focused linear attention, which has the following beneficial effects: The present invention utilizes the backbone extraction network branch to extract global backbone multi-scale features, and the parallel part of the fast convolution branch captures spatial details. The prediction result is obtained by combining the feature maps of both through the result prediction head, showing stronger global information extraction capability and better robustness to input interference. The backbone extraction network branch adopts a focused linear attention module, which obtains a more focused attention distribution through a focusing function. A matrix rank recovery module composed of lightweight deep convolution is used as compensation to maintain feature diversity, which is both efficient and has strong model expressive power. By using spatial channel reconstruction convolutional layers to replace some of the original convolutional layers to construct spatial channel reconstruction convolutional residual modules, redundant computation is reduced and representative feature learning is promoted, while significantly reducing complexity and computational cost. The fast convolutional branch of the FasterNet network architecture is mainly constructed using partial convolution and partial point convolution, which greatly reduces the amount of computation while maintaining the extraction of rich local and global features. In the process of bladder tumor segmentation, the morphological specificity of bladder tumors, the boundary between bladder tumors and normal tissues, and blurred bladder tumor boundaries are effectively distinguished. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of a bladder tumor segmentation method based on parallel branch-focused linear attention, provided by the present invention.

[0031] Figure 2 This is a schematic diagram of the result prediction head structure provided by the present invention.

[0032] Figure 3 This is a schematic diagram of the branch structure of the backbone extraction network provided by the present invention.

[0033] Figure 4 This is a schematic diagram of the local activation module structure provided by the present invention.

[0034] Figure 5 A schematic diagram of the fast convolution branch structure provided by the present invention.

[0035] Figure 6 This is a schematic diagram of the FasterNet network architecture provided by the present invention.

[0036] Figure 7 A schematic diagram of the spatial channel reconstruction convolutional residual module provided by the present invention.

[0037] Figure 8 This is a schematic diagram of the SCConv module structure provided by the present invention.

[0038] Figure 9 This is a schematic diagram illustrating the qualitative comparison results provided by the present invention. Detailed Implementation

[0039] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] This invention discloses a bladder tumor segmentation method based on parallel branch-focusing linear attention, comprising:

[0041] The cystoscopy images were cropped, filtered, and labeled to obtain a labeled bladder tumor dataset;

[0042] A bladder tumor segmentation model is constructed, comprising a backbone extraction network branch for extracting global backbone multi-scale features from cystoscopy images, a fast convolutional branch parallel to the backbone extraction network branch for capturing spatial features, and a result prediction head connected to the backbone extraction network branch and the fast convolutional branch for fusing the features of the two branches to obtain the final prediction result.

[0043] The bladder tumor segmentation model was optimized using a labeled bladder tumor dataset;

[0044] Real-time cystoscopy images are acquired and input into the bladder tumor segmentation model to obtain bladder tumor segmentation results.

[0045] Furthermore, such as Figure 1As shown, a bladder tumor segmentation method based on parallel branch-focused linear attention is presented. Here, FTB is a linear Transformer branch, cat. represents channel connectivity, up.352 is nearest neighbor upsampling, and pred. is a 1×1 convolution followed by a sigmoid function. The method utilizes a FLTP backbone extraction network branch, a parallel fast convolutional branch FCB, a residual++ module RB++, and a result prediction head PH. The FTB backbone extraction network branch extracts more specific morphological features from cystoscopy images, and combines this with a novel focused linear attention module to achieve high efficiency and expressiveness, addressing the morphological specificity issue of bladder tumors. The parallel fast convolutional branch FCB captures spatial details, compensating for the limitations of the FTB backbone extraction network branch in full-size prediction, and addressing the distinction between bladder tumors and normal tissue. The residual++ module RB++ mainly consists of modern variants of residual blocks. The system consists of a group normalization layer, SiLU activation function, convolutional layers, spatial channel reconstruction convolutional (SCConv) layers, and partially convolutional (PConv) layers. It also features residual connections, which reduce redundant computation and promote representative feature learning, while significantly reducing complexity and computational cost. The prediction head (PH) obtains the final prediction result by connecting the feature maps obtained from the backbone extraction network branch FTB and the parallel partially fast convolutional branch FCB. It fuses the features from both branches, combining the important but coarse-grained features extracted by TB with the fine-grained features extracted by FCB to predict the segmentation map, thus addressing the problem of blurred bladder tumor boundaries.

[0046] Furthermore, such as Figure 3 , Figure 4 As shown, LE represents local activation, stage X is the Xth stage of the linear transformer, RB++ represents residual block++, and the FTB backbone extraction network branch of the FLTT Transformer adopts a focused linear attention module. This focused linear attention module, unlike the traditional softmax attention, significantly reduces the complexity of the algorithm. It obtains a more focused attention distribution through a focusing function and uses a matrix rank recovery module composed of lightweight deep convolutions as compensation to maintain feature diversity. It is both efficient and has strong model expressive power. Compared with PVTv2, it shows stronger global information extraction ability and better robustness to input interference, avoiding the model performance degradation caused by using linear attention.

[0047] Furthermore, such as Figure 5As shown, FNB is a fast network block, PConv is a partial convolution, BN is batch normalization, ReLU is the ReLU activation function, Conv.1×1 is a 1×1 convolution, Up is nearest neighbor upsampling, Embedding is a 4×4 convolution with a stride of 4, and Merging is a 2×2 convolution with a stride of 2. To achieve this summation, the parallel Fast Convolutional Branch (FCB) comprises three parts. The first part, the encoder, consists of a branch network of a FasterNet architecture. This encoder performs progressive spatial downsampling and channel expansion to acquire more information. The second part, the intermediate block, consists of two residual++ modules (RB++) for further feature extraction and connects the encoder and decoder. The third part, the decoder, consists of multiple residual++ modules (RB++) and upsampling layers to reconstruct the original image from the encoded features. Each decoder contains a residual module (RB++), which is then upsampled through an upsampling layer, ultimately yielding a feature map with more local details extracted by the parallel Fast Convolutional Branch (FCB).

[0048] Furthermore, such as Figure 6 As shown, the FasterNet network architecture in the parallel partial fast convolution branch FCB uses partial convolution PConv and partial point convolution PWConv. Partial convolution PConv applies regular convolution only on a portion of the input channels to extract spatial features, while the remaining channels remain unchanged. Partial point convolution PWConv adds point convolution to partial convolution PConv, which allows information from all channels to be fully and effectively utilized.

[0049] Furthermore, such as Figure 7 As shown, PConv is a partial convolution, SCConv is a spatial and channel reconstruction convolution, GN is group normalization (group number = 32), SiLU is the SiLU activation function, Conv.3×3 is a 3×3 convolution, SRU is a spatial reconstruction unit, CRU is a channel reconstruction unit, and the residual++ module RB++ is mainly composed of a modern variant of residual blocks, including group normalization layers, SiLU activation functions, convolutional layers, and spatial and channel reconstruction convolution SCConv convolutional layers. It also has residual connections, which can reduce redundant computation and promote representative feature learning, while significantly reducing complexity and computational cost.

[0050] Furthermore, the SCConv convolutional layers used in the residual++ module RB++ employ spatial channel reconstruction convolution SCConv and partial convolution PConv, such as Figure 8As shown, the SCConv module consists of a spatial reconstruction unit (SRU) and a channel reconstruction unit (CRU). The SCConv module is designed to effectively limit feature redundancy, which not only reduces the number of model parameters and FLOPs, but also enhances the feature representation capability.

[0051] Furthermore, such as Figure 2 As shown, the prediction head PH obtains the final prediction result by connecting the feature maps obtained from the backbone extraction network branch FTB and the parallel fast convolution branch FCB. Then, it fuses the feature maps from the two branches and uses the prediction head PH to fuse the important but coarse-grained features extracted by TB with the fine-grained features extracted by FCB to predict the segmentation map.

[0052] Furthermore, the loss function is designed to be the sum of the binary cross-entropy (BCE) loss and the Dice loss. The loss function is defined as follows:

[0053] L = L BCE +L Dice ;

[0054] Among them, L BCE and L Dice These represent the binary cross-entropy (BCE) loss and the Dice loss, respectively. The binary cross-entropy (BCE) loss L... BCE And Dice lost L Dice They are described as follows:

[0055]

[0056]

[0057] Among them, L BCE and L Dice y represents the binary cross-entropy loss (BCE) and the Dice loss, respectively. i It is the binary label value of the i-th sample, which is either 0 or 1 (ground truth), p(y i =1) is the model's prediction for the i-th sample, i.e., the probability that the model predicts the label value of the i-th sample to be 1. |X∩Y| represents the intersection of the sets X and Y, where |X| and |Y| represent the number of elements. For the segmentation task, |X| and |Y| represent the ground truth map and the prediction mask for segmentation. Binary cross-entropy loss (BCE) BCE Dice loss is a commonly used loss function in binary classification problems. Dice It is commonly used in semantic segmentation problems, mainly to penalize predictions with low confidence, while predictions with high confidence will receive a lower Dice loss.

[0058] Furthermore, the segmentation performance of bladder tumors was quantitatively evaluated using the average dice similarity coefficient (mDice), average intersection-union ratio (mIoU), average precision (mPrecision), and average recall (mRecall). The specific descriptions of each indicator are as follows:

[0059] The formula for the average dice similarity coefficient mDice is as follows:

[0060] The formula for the average intersection-union ratio (mIoU) is as follows:

[0061] The formula for average precision mPrecision is as follows:

[0062] The formula for average recall mRecall is as follows:

[0063] Where TP represents true positive, FP represents false positive, FN represents false negative, and n represents the number of test images.

[0064] In a specific implementation, a bladder tumor segmentation system based on parallel branch-focusing linear attention includes:

[0065] Data construction module: used to crop, filter, and label cystoscopy images to obtain an labeled bladder tumor dataset;

[0066] Segmentation Model Construction Module: Used to construct a bladder tumor segmentation model, which includes a backbone extraction network branch for extracting global backbone multi-scale features from cystoscopy images, a fast convolutional branch parallel to the backbone extraction network branch for capturing spatial features, and a result prediction head connected to the backbone extraction network branch and the fast convolutional branch for fusing the features of the two branches to obtain the final prediction result.

[0067] Model optimization module: used to optimize the bladder tumor segmentation model using a labeled bladder tumor dataset;

[0068] Real-time segmentation module: used to acquire real-time cystoscopy images, input them into the bladder tumor segmentation model, and obtain bladder tumor segmentation results.

[0069] In a specific embodiment, 110 surgical videos of 51 patients who underwent cystoscopy or TURBT surgery between 2016 and 2023 were used. The surgical video footage was non-destructively cropped to obtain 47,151 images. After screening, 1,948 high-quality images were selected as the bladder tumor BC-DB dataset. These images were then annotated in collaboration with urology experts, constructing a substantial annotated cystoscopy bladder tumor dataset. This dataset was divided into training and testing sets at an 8:2 ratio for neural network training and validation testing.

[0070] To evaluate segmentation and generalization performance, experiments were conducted on two polyp segmentation datasets: Kvasir-SEG and CVC-ClinicDB. The Kvasir-SEG dataset consists of 1000 polyp images and their corresponding ground-truth polyp masks, annotated by endoscopy experts. The images in Kvasir-SEG range in resolution from 332×487 to 1920×1072 pixels. The CVC-ClinicDB dataset, also known as CVC-612, contains 612 open-access images from 25 colonoscopy videos, with a resolution of 384×288. Each image has its associated manually annotated ground-truth label covering the polyp. Training and testing were performed separately on each public dataset; Table 1 shows the dataset information.

[0071] Table 1

[0072]

[0073]

[0074] Three experiments were conducted to validate the performance of the bladder tumor segmentation method based on parallel branch-focused linear attention, including qualitative, quantitative, and ablation experiments. The methods were compared with six different approaches, including UNet, PraNet, HardNet, CaraNet, DCRNet, and FCBFormer, and three evaluation metrics were analyzed and compared on five different polyp segmentation datasets. For fair comparison, the segmentation maps of these methods were generated from the original code published or directly provided by the authors. Figure 9 As shown in the visualization comparison with different competing methods, it can be qualitatively seen that the present invention can segment tumor regions more accurately and performs better in several challenging aspects. Table 2 compares the experimental results of different methods on the BC-DB dataset for bladder tumors.

[0075] Table 2

[0076]

[0077] Tables 2 and 3 present the quantitative results of statistical comparisons with six different competing methods on five different datasets. As shown in Table 2, the model is optimal across all metrics on both datasets, with mDice and mIou at 90% and 84.1%, respectively. Table 3 compares the experimental results of different methods on the CVC-ClinicDB and Kvasir-SEG polyp datasets.

[0078] Table 3

[0079]

[0080]

[0081] To verify the model's transfer performance, tests were conducted on two public datasets (CVC-ClinicDB and Kvasir-SEG). As shown in Table 3, the model of this invention achieved good transfer performance compared to six competing methods. In particular, on the challenging dataset CVC-ColonDB, the transfer capability of this invention achieved a significant improvement. On CVC-ClinicDB, the model's mDice outperformed FCBFormer and DCRNet by 1.3% and 7.2%, respectively, while mIou outperformed FCBFormer and DCRNet by 0.21% and 10%, respectively. On Kvasir-SEG, the model outperformed FCBFormer and DCRNet by 2.3% and 6%, respectively, while mIou outperformed FCBFormer and DCRNet by 3.7% and 8.3%, respectively.

[0082] To verify the effectiveness of each component in the model, ablation experiments were conducted on the Flatten Transformer backbone extraction network branch FTB, the parallel part of the fast convolution branch FCB, and the spatial channel reconstruction convolutional residual module RB++. The backbone network was modified based on the FCBFormer architecture, with FCBFormer as the baseline model. The effectiveness of different modules was evaluated by removing or changing different modules from the standard model. "Baseline" indicates the baseline model FCBFormer, "Flatten" indicates changing the Flatten Transformer backbone extraction network branch FTB in the baseline model, "RB++" indicates changing the residual++ module RB++ in the baseline model, and "Faster" indicates changing the parallel part of the fast convolution branch FCB in the baseline model. The experimental results are shown in Table 4 below. Table 4 shows the ablation experimental results on the bladder tumor BC-DB dataset.

[0083] Table 4

[0084] Baseline FTB FCB RB++ mDice mIoU mPrecision mRecall FLOPs. √ 0.881806 0.816558 0.885005 0.911282 73295.38M √ √ 0.892922 0.828194 0.899305 0.914457 73431.67M √ √ 0.892922 0.82884 0.902339 0.911427 67144.23M √ √ 0.900619 0.839475 0.912442 0.913681 57927.20M √ √ √ 0.889335 0.825342 0.894735 0.907559 63185.00M √ √ √ 0.899941 0.835762 0.912442 0.913681 58063.50M √ √ √ 0.900553 0.837717 0.905223 0.9183 51741.45M √ √ √ √ 0.903875 0.841768 0.91499 0.911656 50834.72M

[0085] Further ablation studies on FTB: To investigate the effectiveness of the FTB branch, a version called "Flatten" was trained, which differed from the baseline model only in the modification of the FTB branch. The experimental results are shown in Table 4. Compared to the baseline model, the model with only the modified FTB branch of the Flatten Transformer backbone extraction network showed significant improvements across all four evaluation metrics. Modifying the FTB improved mDice, mIou, and mPrecision by 1.1%, 0.12%, and 0.11%, respectively. Meanwhile, the computational complexity metric FLOPs remained largely unchanged, indicating no increase in computational cost.

[0086] Ablation Study of RB++: To quantitatively analyze the effectiveness of the modified RB++ module, a version "RB++" with only the RB++ module modified was trained. Table 4 shows that, compared to the baseline model, the model with the modified RB++ module significantly improved across all four evaluation metrics, increasing mDice, mIou, mPrecision, and mRecall by 1.9%, 2.3%, 2.7%, and 0.02%, respectively. Simultaneously, it reduced computational complexity by approximately 23% in FLOPs, significantly decreasing the computational cost.

[0087] Ablation study of FCB: To demonstrate the capabilities of FCB, a version called "Faster" was trained with only the FCB branch modified. As shown in Table 4, compared to the standard model, modifying and adding the FCB branch improved mDice, mIou, and mPrecision by 1.1%, 1.2%, and 1.4%, respectively. Simultaneously, it resulted in a reduction of approximately 14% in FLOPs, a metric for computational complexity, significantly reducing the computational load.

[0088] Combination Ablation Study: Ablation experiments were conducted on different combinations of modules to assess their combined effects. Individual improvements to FTB, FCB, and RB++ resulted in improvements across various metrics, and pairwise combinations also yielded varying degrees of improvement. The FFB-Former, which integrates all improvements, achieved the best results in all metrics except mRecall. These results demonstrate the effectiveness of our improved methods.

[0089] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0090] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bladder tumor segmentation method based on parallel branch focused linear attention, characterized in that, The method comprises the steps of: carrying out cropping, screening and labeling on cystoscope images to obtain a labeled bladder tumor dataset; constructing a bladder tumor segmentation model, wherein the bladder tumor segmentation model comprises a backbone extraction network branch for extracting global backbone multi-scale features from the cystoscope images, a fast convolution branch for capturing spatial features, which is mutually parallel to the backbone extraction network branch, and a result prediction head for fusing features of the two branches to obtain a final prediction result, which is connected to the backbone extraction network branch and the fast convolution branch; the backbone extraction network branch adopts a focal linear attention module; the fast convolution branch comprises three parts, a first part, an encoder part, which is composed of a branch network of a FasterNet network architecture, gradually performs spatial downsampling and channel number expansion to obtain information, a second part, an intermediate block, which is composed of two spatial channel reconstruction convolution residual modules, is used for extracting features and connecting the encoder and the decoder, and a third part, a decoder part, which is composed of a plurality of spatial channel reconstruction convolution residual modules and up-sampling layers, is used for reconstructing an original image from the encoded features, each decoder comprises a spatial channel reconstruction convolution residual module, and up-sampling is performed through the up-sampling layer to finally obtain a feature map extracted by the fast convolution branch; the fast convolution branch adopts an up-sampling restoration strategy of staggered skip connection, starts to be connected in parallel step by step according to the results of two layers of the intermediate block, connects and up-samples according to the feature pyramid of each layer of the encoder stage, connects the parallel results, and finally obtains a final fast convolution branch prediction feature map through residual connection with the original input; the bladder tumor segmentation model is optimized through the labeled bladder tumor dataset; real-time cystoscope images are obtained and input into the bladder tumor segmentation model to obtain a bladder tumor segmentation result.

2. The bladder tumor segmentation method based on parallel branch focused linear attention according to claim 1, characterized in that, The method comprises the steps of: obtaining a surgery video picture and performing lossless cropping, screening the image according to the quality to obtain a bladder tumor dataset, labeling the image, constructing a labeled bladder tumor dataset, and dividing the training set and the test set according to the quantity ratio of 8:

2.

3. The bladder tumor segmentation method based on parallel branch focused linear attention according to claim 1, characterized in that, The FasterNet network architecture adopts partial convolution and partial point convolution, the partial convolution applies a conventional convolution on part of the input channels to extract spatial features, and the remaining channels remain unchanged; and the partial point convolution is additionally provided with a point convolution on the basis of the partial convolution.

4. The bladder tumor segmentation method based on parallel branch focused linear attention according to claim 1, characterized in that, The spatial channel reconstruction convolution residual module comprises a group normalization layer, a SiLU activation function, a convolution layer and a spatial channel reconstruction convolution layer, and the spatial channel reconstruction convolution layer is composed of a spatial reconstruction unit and a channel reconstruction unit.

5. The bladder tumor segmentation method based on parallel branch focused linear attention according to claim 1, characterized in that, The loss function of the bladder tumor segmentation model is the sum of a binary cross entropy loss and a Dice loss, and the loss function is defined as follows: ； where, and respectively represent binary cross-entropy loss and Dice loss.

6. The bladder tumor segmentation method based on parallel branch focused linear attention according to claim 1, characterized in that, The method further comprises the steps of: qualitative analysis, quantitative analysis and ablation analysis are used to evaluate the performance of the bladder tumor segmentation model, and the quantitative analysis comprises the following steps: the performance of the bladder tumor segmentation model is quantitatively evaluated by using an average dice similarity coefficient, an average intersection over union, an average precision and an average recall.

7. A bladder tumor segmentation system based on parallel branch focused linear attention, characterized in that, The method comprises the steps of: a data construction module is used for cropping, screening and labeling the cystoscope images to obtain a labeled bladder tumor dataset; The segmentation model construction module is configured to construct a bladder tumor segmentation model, which includes a backbone extraction network branch for extracting global backbone multi-scale features from cystoscope images, a fast convolution branch for capturing spatial features in parallel with the backbone extraction network branch, and a result prediction head connected with the backbone extraction network branch and the fast convolution branch for fusing features of the two branches to obtain a final prediction result. The backbone extraction network branch adopts a focal linear attention module. The fast convolution branch includes three parts. A first part, an encoder part, is composed of a branch network of a FasterNet network architecture, which performs step-by-step spatial down-sampling and channel number expansion to obtain information. A second part, an intermediate block, is composed of two spatial channel reconstruction convolution residual modules, which are used to extract features and connect the encoder and the decoder. A third part, a decoder part, is composed of multiple spatial channel reconstruction convolution residual modules and up-sampling layers, which are used to reconstruct an original image from the encoded features. Each decoder includes a spatial channel reconstruction convolution residual module, which is up-sampled by an up-sampling layer to finally obtain a feature map extracted by the fast convolution branch. The fast convolution branch adopts an up-sampling restoration strategy of staggered skip connection, which starts to be connected in parallel step by step according to the results of two layers of the intermediate block, connects and up-samples according to the feature pyramid of each layer of the encoder stage, connects the parallel results, and finally obtains a final fast convolution branch prediction feature map by residual connection with the original input. The model optimization module is configured to optimize the bladder tumor segmentation model by using a labeled bladder tumor data set. The real-time segmentation module is configured to obtain real-time cystoscope images, which are input into the bladder tumor segmentation model to obtain a bladder tumor segmentation result.

Images