Bladder tumor microscopic examination auxiliary diagnosis and operation planning system based on AI

Abstract

The invention discloses an AI-based bladder tumor microscopic examination auxiliary diagnosis and operation planning system, and relates to the technical field of medical artificial intelligence. The system comprises an image acquisition module, an AI diagnosis module, an operation planning module and a human-computer interaction interface, real-time high-precision analysis and diagnosis of cystoscope images are realized through an improved HRNetV2 neural network and multi-resolution feature fusion, and exosome spectral analysis is integrated to assist in early screening. The system can automatically generate a personalized surgical scheme based on a diagnosis result, supports three-dimensional reconstruction and surgical path planning, and can be in butt joint with a flexible surgical robot to execute accurate operation. The bladder tumor diagnosis accuracy and the operation safety are effectively improved, dependence on external labeled data is reduced, and good clinical applicability and popularization value are achieved.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to an AI-based bladder tumor endoscopic-assisted diagnosis and surgical planning system. Background Technology

[0002] Bladder cancer is one of the most common malignant tumors of the urinary system, ranking 7th in incidence among all malignant tumors. Currently, cystoscopy is the gold standard for diagnosing bladder cancer, but traditional cystoscopy mainly relies on the doctor's visual observation and experience, which has certain limitations. Doctors are easily influenced by subjective factors during observation, such as fatigue and differences in experience, which may lead to the missed diagnosis of small or atypical tumor lesions.

[0003] Statistics show that the survival rate for early-stage bladder cancer can reach over 90%. However, many patients still experience delayed treatment due to misdiagnosis, missed diagnosis, or cumbersome examinations. Because bladder tumors exhibit diverse morphologies, such as villous, follicular, or flat erythematous lesions, they often resemble various inflammatory lesions. This can lead doctors to mistakenly classify tumors as benign, thus missing the opportunity for biopsy and resulting in a missed diagnosis.

[0004] In recent years, artificial intelligence technology has made significant progress in the field of medical image analysis. Studies have shown that deep learning-based algorithms exhibit good performance in bladder tumor detection, but the following problems still exist in practical clinical applications: Most existing systems are limited to image analysis and fail to be fully integrated with clinical diagnosis and treatment processes; Lack of multimodal data fusion capabilities; The surgical planning function is inadequate. The system has limited generalization capabilities and is difficult to adapt to the differences in equipment among different medical institutions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based bladder tumor endoscopic-assisted diagnosis and surgical planning system. This system can provide full-process assistance from diagnosis to surgical planning, thereby improving the diagnostic accuracy and surgical treatment outcomes of bladder cancer.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-based bladder tumor endoscopy-assisted diagnosis and surgical planning system, comprising: an image acquisition module, an AI diagnosis module, a surgical planning module, and a human-computer interaction interface; The image acquisition module is used to acquire and preprocess image frames in the cystoscopy video stream, and integrate the patient's clinical structured data (such as age, gender, medical history, etc.) to form multimodal input data; The AI ​​diagnostic module receives image data and performs the following analyses in sequence: Tumor detection submodule: Locates suspected tumor regions in the image and outputs their bounding box coordinates; Segmentation submodule: Performs pixel-level precise segmentation of the detected tumor region and outlines its contour; Grading submodule: Predicts the malignancy of the segmented tumor based on its morphological characteristics (such as texture, edge regularity, and vascular distribution); This module employs an improved HRNetV2 (High-Resolution Net Version 2) as its core network architecture, which includes parallel-connected multi-resolution sub-networks and a multi-layered feature fusion mechanism to ensure the simultaneous preservation of rich spatial details and high-level semantic information. Its forward propagation process can be described as follows: F_out=F_fusion(F_1(I),F_2(I),F_3(I),F_4(I)); Where I is the input cystoscopy image, F_1 ​​to F_4 represent the feature extraction functions of four different resolution branches, and F_fusion is the feature fusion function, which is responsible for effectively aggregating multi-scale features.

[0007] To further address the scarcity of labeled medical data, the tumor detection and segmentation submodule introduces a semi-supervised learning mechanism based on uncertainty estimation, employing a teacher-student model structure. This mechanism utilizes unlabeled data to enhance the model's generalization ability, and its overall loss function is designed as follows: L_total = α × L_supervised + β × L_unsupervised + γ × L_regularization; Where L_supervised is the supervised loss based on labeled data (such as cross-entropy loss, Dice loss), L_unsupervised is the unsupervised consistency loss based on unlabeled data (such as mean squared error MSE, KL divergence), L_regularization is the regularization term to prevent overfitting, and α, β, γ are the weighting coefficients used to balance the various losses.

[0008] The surgical planning module receives the output from the AI ​​diagnostic module, including the tumor's location, size, shape, and grade information, and automatically generates a personalized surgical plan based on this. This plan specifically includes: Resection extent planning: Based on tumor grading and clinical guidelines, the extent of tissue to be resected and the safe boundaries are automatically calculated and marked on the 3D model.

[0009] Path planning: Plan the optimal path for surgical instruments to reach the tumor area from the entrance, and use algorithms (such as A* algorithm) to avoid important blood vessels and nerve structures to minimize surgical damage.

[0010] Risk prediction: Based on the relationship between the tumor location and surrounding key anatomical structures, predict potential risks during surgery and provide early warnings.

[0011] This module integrates three-dimensional reconstruction technology, which can reconstruct a three-dimensional model of the bladder interior from continuous two-dimensional cystoscopy images through techniques such as Structure from Motion (SfM) or stereo matching, thereby accurately visualizing the spatial relationship between the tumor and surrounding tissues.

[0012] The human-computer interaction interface is used to graphically and intuitively display diagnostic results (such as tumor localization box, segmentation contour, grading results) and surgical plans (3D model, resection range, surgical path) to the doctor, and receive feedback, confirmation or manual adjustment instructions from the doctor, ensuring that the doctor always has the final decision-making power.

[0013] As a further improvement of the present invention, an exosome analysis module may also be included. This module detects specific Raman spectral characteristics of exosomes in urine, uses a convolutional neural network (CNN) to classify and analyze the spectral data, generates exosome-based diagnostic indicators, and cross-references these indicators with cystoscopy image diagnostic results to jointly assist in the early diagnosis and risk assessment of bladder cancer.

[0014] As a further improvement of the present invention, the surgical planning module can be integrated with a flexible surgical robot system, which can directly convert the generated surgical plan (especially the path planning) into control instructions that the robot can execute, driving the robotic arm to complete precise surgical operations.

[0015] As a further improvement of the present invention, the system employs a federated learning framework for model training. Multiple medical institutions' local systems, coordinated by a central server, collaboratively train the AI ​​model, but the original patient data from each center remains locally, with only model parameter updates exchanged. This allows for the utilization of multi-source data to improve model performance while protecting patient privacy.

[0016] As a further improvement of the present invention, the system includes a dynamic update mechanism. The system can continuously collect information on doctors' confirmation and modification of AI diagnostic results during clinical use, as well as patients' postoperative pathological results and follow-up data. By using this real-world data, the AI ​​model can be incrementally learned or fine-tuned, thereby achieving continuous optimization and performance improvement of the model.

[0017] This invention provides an AI-based endoscopic-assisted diagnosis and surgical planning system for bladder tumors. Compared with existing technologies, it has the following advantages: (1) The AI-based bladder tumor endoscopy-assisted diagnosis and surgical planning system significantly improves the detection sensitivity and accuracy of bladder tumors through the improved HRNetV2 neural network architecture and multi-resolution feature fusion mechanism, and improves sensitivity under high-resolution images.

[0018] (2) The AI-based bladder tumor endoscopy-assisted diagnosis and surgical planning system effectively reduces the dependence on a large amount of labeled data by introducing a semi-supervised learning mechanism based on uncertainty estimation and a teacher-student model structure, thereby improving the model's adaptability and generalization ability in the case of scarce labeled data.

[0019] (3) The AI-based bladder tumor endoscopy-assisted diagnosis and surgical planning system integrates an exosome analysis module and combines cystoscopy images and exosome spectral data to achieve multimodal data fusion, providing more comprehensive diagnostic information, assisting in the early diagnosis of bladder cancer, and improving accuracy.

[0020] (4) The AI-based bladder tumor endoscopy-assisted diagnosis and surgical planning system generates personalized surgical plans through the three-dimensional reconstruction, resection range planning and path planning functions of the surgical planning module, and integrates with the flexible surgical robot system to achieve sub-millimeter precision operation and improve surgical safety and effect.

[0021] (5) The AI-based bladder tumor endoscopy-assisted diagnosis and surgical planning system allows multiple medical institutions to collaboratively train the AI ​​model without sharing the original data through a federated learning framework and dynamic update mechanism, thus protecting patient privacy. It can also continuously optimize the model based on doctor feedback and postoperative results, thereby improving system performance and clinical application value. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a neural network structure diagram of the AI ​​diagnostic module of the present invention; Figure 3 This is a flowchart of the surgical planning module of the present invention; Figure 4 This is a comparison chart of the system performance evaluation results of the present invention.

[0023] In the diagram: 1. Image acquisition module; 2. AI diagnosis module; 3. Surgical planning module; 4. Human-computer interaction interface; 5. Exosome analysis module. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention have been clearly and completely described. 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.

[0025] Please see Figure 1 This invention provides an AI-based bladder tumor endoscopic-assisted diagnosis and surgical planning system, comprising four core modules: image acquisition module 1, AI diagnosis module 2, surgical planning module 3, and human-computer interaction interface 4.

[0026] Image acquisition module 1 is responsible for acquiring cystoscopy images and patient clinical data. This module supports image input at multiple resolutions, including high-resolution (above 1920×1080 pixels) and low-resolution (640×480 pixels) images. Simultaneously, this module integrates exosome analysis data, providing multimodal diagnostic information to the system by detecting the Raman spectral characteristics of exosomes in urine.

[0027] AI diagnostic module 2 employs an improved HRNetV2 neural network architecture, including parallel multi-resolution sub-networks and feature fusion mechanisms. This module is specifically implemented as three functional sub-modules: The tumor detection submodule is responsible for identifying suspected tumor areas in images. Segmentation submodule: Performs precise pixel-level contour division of the tumor region; Grading submodule: Assess the degree of malignancy based on tumor morphological characteristics; The AI ​​diagnostic module 2 is implemented based on a high-end GPU computing workstation and the PyTorch / TensorFlow deep learning framework. Its core network adopts an improved HRNetV2 architecture. The forward propagation process is defined by the formula: F_out = F_fusion(F_1(I), F_2(I), F_3(I), F_4(I)), where I is the input cystoscopy image, F_1 ​​to F_4 are four parallel sub-networks, respectively processing feature extraction at 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image resolution, and F_fusion is the feature fusion function. Through cross-resolution exchange and aggregation operations, it generates refined features F_out containing rich spatial details and high-level semantic information, such as... Figure 2 As shown.

[0028] The tumor detection and segmentation submodule employs a semi-supervised learning mechanism based on uncertainty estimation, using a teacher-student model structure. Its training process is achieved by optimizing the overall loss function: L_total = α × L_supervised + β × L_unsupervised + γ × L_regularization In practice: For supervised loss items, the following approach is adopted: L_supervised=CrossEntropyLoss(P_pred,P_gt)+0.5×DiceLoss(P_pred,P_gt), where P_pred is the model's predicted value and P_gt is the true label, ensuring that the model learns the accurate features of the labeled data.

[0029] Unsupervised loss terms are adopted as follows: L_unsupervised=0.8×MSE(Teacher(P_input),Student(P_input))+0.2×KLDiv(Teacher(P_input),Student(P_input)) fully utilizes unlabeled data by constraining the consistency between the teacher model (using exponential moving average weights) and the student model predictions.

[0030] Dynamic adjustment of weight coefficients: In the early stage of training (α=0.8, β=0.1, γ=0.1), and in the later stage (α=0.4, β=0.5, γ=0.1), the regularization term L_regularization adopts L2 regularization to prevent overfitting.

[0031] The network training process is as follows: First, the improved HRNetV2 network is trained in a supervised manner using a large dataset of labeled cystoscopy images (including tumor bounding boxes, segmentation masks, and pathological grading labels) to optimize the supervised part (L_supervised) of the loss function. Simultaneously, a large number of unlabeled cystoscopy images are introduced for semi-supervised learning through a teacher-student model structure. The teacher model uses exponential moving average (EMA) to update weights from the student model and provides a consistency constraint objective for the student model's predictions on unlabeled data, thereby optimizing the unsupervised part (L_unsupervised) of the loss function. The weight coefficients α, β, and γ are dynamically adjusted according to the training stage, initially focusing on supervised learning (with a larger α), and gradually increasing the weight of unsupervised learning (with an increased β) in the later stages. Regularization terms (L_regularization), such as weight decay, are used to prevent overfitting.

[0032] During real-time diagnosis, the system analyzes the cystoscopy video stream frame by frame or keyframe by frame. The input image, after normalization preprocessing, is fed into a trained network for forward inference. The output results undergo post-processing (Non-maximum suppression (NMS) for bounding box deduplication and contour smoothing algorithm for optimizing segmentation boundaries) to generate the final diagnostic result.

[0033] Surgical planning module 3 generates personalized surgical plans based on AI diagnostic results. The workflow is as follows: Figure 3As shown, this module integrates three-dimensional reconstruction technology to reconstruct two-dimensional cystoscopy images into a three-dimensional model, accurately marking the location, size, and relationship of the tumor with surrounding tissues. It also considers key structures such as blood vessel distribution and nerve tissue to optimize the surgical path.

[0034] The specific implementation of 3D reconstruction is as follows: Using consecutive frames of cystoscopy video, the camera pose is recovered and a sparse point cloud is generated through the Structure for Motion Restoration (SfM) algorithm. Then, a dense point cloud is generated using the Multi-View Stereo Vision (MVS) algorithm. Finally, the Moving Cubes algorithm is used for surface reconstruction to generate a 3D mesh model of the bladder's interior. The 2D tumor contour segmented by AI is mapped to the corresponding position in the 3D model based on camera parameters.

[0035] The resection range is planned using the formula to calculate the safe boundary: S = K × (1 + G × 0.5), where S is the safe boundary distance (mm), K is the basic safe distance (5mm), and G is the tumor grade coefficient (low grade = 1, high grade = 2). For example, the safe boundary for a high-grade tumor (G = 2) is S = 5 × (1 + 2 × 0.5) = 10mm. The calculation results are converted into a three-dimensional voxel range and visualized in the model.

[0036] For example, for low-grade non-muscle-invasive bladder cancer, the system automatically extends 5mm outward from the tumor outline on the 3D model to generate a safe boundary; for high-grade tumors or muscle-invasive bladder cancer, it extends 10mm.

[0037] The path planning uses the A* algorithm, with the evaluation function being: f(n) = g(n) + h(n), where g(n) is the actual cost (Euclidean distance) from the starting point to node n, and h(n) is the estimated cost from node n to the target point. The optimal surgical path is searched with the surgical instrument inlet as the starting point, the tumor core as the target point, the bladder wall as the constraint surface, and important blood vessels and nerve structures as obstacles.

[0038] Specifically, it uses the entry point of the surgical instrument through the urethra as the starting point, the core of the tumor region as the target point, the surface of the bladder wall as the constraint surface, and areas with important blood vessels and nerve structures (which need to be marked in advance on preoperative images or preliminarily identified by AI) as obstacles. Within this space, it searches for the shortest and safest path for the instrument to reach it. This path is also visualized in a 3D model.

[0039] The human-computer interface 4 provides a visual interface that displays diagnostic results and surgical plans, supporting physician feedback and adjustments. The interface design conforms to clinical workflows and supports touch and voice interaction for easy intraoperative operation.

[0040] The exosome analysis module 5 provides auxiliary diagnosis by detecting the Raman spectral characteristics of urinary exosomes. Exosomes are separated using a gold nanopore array, and the spectral signal is enhanced by surface-enhanced Raman scattering (SERS). The acquired spectral data undergoes Savitzky-Golay filtering for noise reduction, multinomial baseline correction, and SNV normalization preprocessing before being input into a CNN classifier for analysis. The module outputs the probability of tumor-related exosomes and generates a quantitative diagnostic report.

[0041] The specific implementation process of exosome analysis module 5 is as follows: Patient urine samples are collected and preprocessed to extract exosomes. The extracted samples are placed on a surface-enhanced Raman scattering (SERS) substrate (such as a gold nanopore array) to capture exosomes. The spectral signals of the captured exosomes are acquired using a Raman spectrometer. The acquired raw spectral data first undergoes a preprocessing process, including Savitzky-Golay filtering for noise reduction, polynomial fitting baseline correction, and standard normal variable transformation (SNV) normalization. The processed spectral data is then input into a pre-trained convolutional neural network (CNN, such as ResNet or a custom network) for classification analysis, outputting the probability that it is a tumor-related exosome, and generating a count and proportion report of tumor-related exosomes for each patient as a quantitative indicator for auxiliary diagnosis.

[0042] The system supports a federated learning framework and employs a federated averaging algorithm: w_global = Σ_{k=1}^{K}(n_k / n) × w_k, where w_global represents the global model parameters, w_k is the k-th client parameter, n_k is the amount of client data, and n is the total amount of data. Each medical institution trains its model locally and then encrypts and uploads its parameters. The server aggregates and updates the global model, protecting patient privacy while improving model performance, as detailed below: Multiple hospitals' local systems participate in model training under the coordination of a central server. Each center trains its local model using anonymized local data, and after training, encrypts the model parameters before uploading them to the central server. The server uses a federated averaging algorithm (FedAvg) to securely aggregate all uploaded model parameters, generating a globally improved model, and then distributes the updated global model parameters to all participating hospitals. This iterative process enables collaborative optimization of the AI ​​diagnostic model without sharing the original patient data.

[0043] The system also includes a dynamic update mechanism. Based on physician feedback and postoperative pathology data collected during clinical use, it uses the objective function: L_update=L_total+λ×L_new for incremental learning, where λ is a balancing hyperparameter (0.3-0.5) to control the intensity of new knowledge introduction and achieve continuous model optimization, as detailed below: During clinical use, the system records doctors' confirmation and modification actions regarding AI diagnostic results, as well as the final postoperative pathological diagnosis. This data, after anonymization, forms a continuously growing validation dataset. The system periodically or as needed uses this new dataset to incrementally learn or fine-tune the AI ​​model, enabling it to continuously adapt to new clinical scenarios and achieve continuous performance optimization and iteration. System performance evaluation: To evaluate system performance, we collected 102 cystoscopy videos, including high-resolution and low-resolution images. Tumor detection was performed using the HRNetV2 model, and the results are as follows... Figure 4 As shown in Table 1.

[0044] Table 1: Performance comparison of the system under images of different resolutions:

[0045] At high resolution images, the system exhibits excellent performance, with a sensitivity of 94.8%, accuracy of 94.4%, and an mDice score of 84.7%. Even at low resolution images, the system maintains an acceptable performance level, with a sensitivity of 75.6% and an accuracy of 74.8%.

[0046] Compared with traditional methods, this system significantly improves the accuracy of bladder tumor detection. In a multicenter clinical trial, the accuracy of bladder tumor detection with system assistance was 25.3% higher than that of traditional methods, while the false negative rate was reduced by 18.7%.

[0047] The following example illustrates the clinical application process of this system.

[0048] The patient was a 62-year-old male who presented with painless gross hematuria. Cystoscopy was performed first, and the image acquisition module captured images of the bladder interior, which were then transmitted to the AI ​​diagnostic module.

[0049] The AI ​​diagnostic module analyzed the image in real time and detected a papillary tumor approximately 1.2 cm in diameter on the left wall of the bladder. The segmentation submodule precisely divided the tumor area, and the grading submodule assessed it as high-grade urothelial carcinoma based on the tumor's morphological characteristics.

[0050] The exosome analysis module analyzed the patient's urine sample and found a significant increase in the number of tumor-related exosomes, further confirming the diagnosis.

[0051] The surgical planning module generates a personalized surgical plan based on the diagnostic results. 3D reconstruction shows the tumor is located near the left ureteral orifice, requiring special care to protect it. The system plans an appropriate resection area and surgical path to ensure complete tumor removal while preserving surrounding healthy tissue.

[0052] Finally, the surgical plan incorporated a fully flexible dual-arm endoscopic surgical robot system, allowing the surgeon to perform a transurethral en bloc resection of the bladder tumor with the system's assistance. The surgery took approximately 15 minutes, with minimal bleeding. The patient recovered well post-operatively and was discharged the following day.

[0053] Postoperative pathological examination confirmed high-grade urothelial carcinoma with negative surgical margins, verifying the accuracy of the systematic diagnosis and planning.

[0054] System advantages and innovations The system of this invention has the following advantages compared with the prior art: (1) High diagnostic accuracy: The improved HRNetV2 network architecture is adopted, combined with the multi-resolution feature fusion mechanism, achieving a sensitivity of 94.8% and an accuracy of 94.4% in high-resolution images.

[0055] (2) Reduce dependence on labeled data: Introduce a semi-supervised learning mechanism based on uncertainty estimation and adopt a teacher-student model structure, which can achieve good performance with a small amount of labeled data.

[0056] (3) Multimodal data fusion: Integrating cystoscopy images and exosome spectral data to provide more comprehensive diagnostic information, with a diagnostic accuracy of 97.37%.

[0057] (4) Full-process assistance: to achieve full-process assistance from diagnosis to surgical planning, and improve the efficiency and consistency of diagnosis and treatment.

[0058] (5) High surgical precision: It is integrated with a flexible surgical robot system to achieve sub-millimeter precision operation and reduce surgical complications.

[0059] (6) Privacy protection: The federated learning framework allows multiple medical institutions to collaboratively train AI models without sharing raw data, thus protecting patient privacy.

[0060] (7) Continuous optimization: It has a dynamic update mechanism and can continuously optimize the AI ​​model based on doctor feedback and postoperative results. This invention provides an AI-based endoscopic-assisted diagnosis and surgical planning system for bladder tumors, which can effectively improve the diagnostic accuracy and surgical treatment outcomes of bladder cancer. The system integrates AI diagnosis, exosome analysis, and surgical planning functions, providing end-to-end assistance from diagnosis to treatment. Clinical trials have shown that the system can significantly reduce the missed diagnosis and misdiagnosis rates of bladder cancer, improve surgical precision and safety, and has significant clinical application value.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based cystoscopy assisted diagnosis and surgery planning system, characterized in that: Comprising: An image acquisition module, an AI diagnosis module, a surgical planning module, and a human-computer interaction interface; The image acquisition module is used to collect cystoscope images and patient clinical data; The AI diagnosis module includes a tumor detection submodule, a segmentation submodule, and a grading submodule, which are used for real-time analysis and diagnosis of cystoscope images; The surgical planning module generates personalized surgical plans based on AI diagnosis results, including resection range planning, path planning, and risk prediction; The human-computer interaction interface is used to display diagnosis results and surgical plans, and to receive feedback and adjustments from doctors. 2.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 1, wherein: The AI diagnosis module uses an improved HRNetV2 neural network architecture, which contains parallel multi-resolution subnetworks and feature fusion mechanisms, and its forward propagation formula is: F_out=F_fusion(F_1(I),F_2(I),F_3(I),F_4(I)); Where I is the input image, F_1 to F_4 are the feature extraction functions of the four resolution branches, and F_fusion is the feature fusion function. 3.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 1, wherein: The tumor detection submodule uses a semi-supervised learning mechanism based on uncertainty estimation, which reduces the dependence on labeled data through a teacher-student model structure, and its loss function is: L_total=α×L_supervised+β×L_unsupervised+γ×L_regularization; Where L_supervised is the supervised loss, L_unsupervised is the unsupervised loss, L_regularization is the regularization term, and α, β, γ are weight coefficients. 4.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 1, wherein: The system also includes an exosome analysis module that detects the Raman spectral characteristics of exosomes in urine and uses a convolutional neural network for classification analysis to assist in early diagnosis of bladder cancer. 5.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 1, wherein: The surgical planning module integrates three-dimensional reconstruction technology to reconstruct two-dimensional cystoscope images into three-dimensional models, accurately labeling tumor location, size, and relationship with surrounding tissues. 6.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 5, wherein: The surgical planning module is integrated with a flexible surgical robot system, which can directly convert the planning scheme into robot control instructions to achieve precise surgical operations. 7.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 1, wherein: The system uses a federated learning framework that allows multiple medical institutions to collaboratively train AI models without sharing raw data, protecting patient privacy. 8.The AI-based cystoscopy aided diagnosis and surgery planning system according to claim 1, wherein: The system contains a dynamic updating mechanism that can continuously optimize the AI model based on doctor feedback and postoperative results to improve diagnosis and planning accuracy.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the functions of the system as claimed in any one of claims 1-8.

10. A method of bladder tumor diagnosis and surgical planning, characterized by, Comprising: Obtaining images of the patient's bladder through a cystoscope; Using an AI model to perform real-time analysis and diagnosis of the images; Generating personalized surgical plans based on the diagnosis results; Importing the surgical plans into a surgical robot system to perform surgical operations.

Images