Artificial intelligence-based thyroid nodule surgical planning and navigation system

Abstract

The application relates to the medical field and discloses a thyroid nodule surgical operation planning and navigation system based on artificial intelligence, which comprises a multi-modal medical image acquisition module, an image preprocessing and enhancement module, an intelligent anatomical structure analysis module, a surgical operation planning decision module and an intraoperative real-time navigation module; the multi-modal medical image acquisition module is used for integrating data of various image equipment such as ultrasound, CT and MRI, and ensuring the space-time consistency of multi-modal images. The thyroid nodule surgical operation planning and navigation system based on artificial intelligence solves the problem of real-time calculation, reduces system delay, is crucial for real-time navigation in operation, improves real-time performance and operation efficiency, shortens operation time, reduces complications, integrates the advantages of different images, can provide a more comprehensive anatomical view, helps doctors to more accurately position nodules and important structures, avoids damaging nerves and blood vessels, and therefore improves accuracy and further reduces operation risk.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medicine, in particular to a thyroid nodule surgical operation planning and navigation system based on artificial intelligence. BACKGROUND

[0002] Thyroid nodule refers to a single or multiple local lumps formed in the thyroid gland, which is often discovered through imaging examination. With the development of imaging technology, more and more patients are found to have thyroid nodules. According to research, the incidence of thyroid nodules varies greatly among different populations. Some studies show that the incidence in the general population is as high as 30%-70%.

[0003] Surgery is an important method for dealing with some thyroid nodules, especially malignant nodules. Common surgical types of thyroid nodules include thyroidectomy, which can be performed according to the nature of the lesion, with total resection, removal of the entire thyroid gland, or partial resection, half or subtotal resection as the main method. Nodule resection is a local resection performed on benign nodules. With the development and innovation of medical technology, surgical planning and navigation systems can help surgeons improve the accuracy of surgery and reduce risks, providing safer and more effective treatment options for patients.

[0004] Currently, there are still many problems to be solved in the surgical planning and navigation system for thyroid nodules. First, the artificial intelligence algorithm requires a large amount of computing resources, and real-time analysis and processing of image data may have a delay, affecting the actual application in surgery. Second, the imaging characteristics and information content of different types of medical images differ greatly, which may affect the overall performance of the system and there is room for improvement. SUMMARY

[0005] (I) Technical problems solved

[0006] To solve the above problems, the present application provides a thyroid nodule surgical operation planning and navigation system based on artificial intelligence.

[0007] (II) Technical solutions

[0008] To achieve the above-mentioned purposes, the present application provides the following technical solutions: a thyroid nodule surgical operation planning and navigation system based on artificial intelligence, comprising a multi-modal medical image acquisition module, an image preprocessing and enhancement module, an intelligent anatomical structure analysis module, a surgical planning and decision-making module, and an intraoperative real-time navigation module. The multi-modal medical image acquisition module is used to integrate data from various imaging devices such as ultrasound, CT, and MRI, to ensure the spatiotemporal consistency of multi-modal images, including time calibration and spatial registration of different devices, supporting DICOM standard protocol and hospital PACS system docking.

[0009] The image preprocessing and enhancement module is used to eliminate image noise and artifacts, improve the contrast of key anatomical structures, and realize cross-device image standardization.

[0010] The intelligent anatomical structure analysis module is used for multi-scale anatomical structure recognition, three-dimensional spatial relationship modeling, and pathological feature quantitative analysis.

[0011] The surgical planning decision module is used for personalized approach planning, risk probability modeling, and multi-objective optimization decision.

[0012] The intraoperative real-time navigation module is used for multi-source information fusion navigation, dynamic environment perception, and sub-millimeter positioning. The multi-modal registration, ICP registration based on feature points, preoperative CT and intraoperative ultrasound, and the use of deep feature descriptor 3D SIFT to improve registration accuracy are used. The AR navigation system integrates MR devices such as Hololens 2, develops a ray projection rendering engine to realize virtual and real fusion, and adopts an extended Kalman filter to compensate for tissue deformation in real time.

[0013] Preferably, the specific working process of the multi-modal medical image acquisition module is as follows: obtaining original DICOM data from image devices through API interface, performing multi-modal image synchronization, reducing motion artifacts based on respiratory / heartbeat gating technology, aligning different modal images using mutual information maximization algorithm Elastix framework, and outputting spatio-temporal aligned multi-modal image dataset.

[0014] Preferably, the specific working process of the image preprocessing and enhancement module is as follows: noise suppression using non-local mean NLM filtering and MARNet network based on deep learning; contrast enhancement using adaptive histogram equalization CLAHE and domain adaptation enhancement using GAN network CycleGAN; standardization processing through N4 bias field correction to eliminate MRI intensity non-uniformity and using z-score standardization to unify different device gray value range.

[0015] Preferably, the sub-modules and workflow of the intelligent anatomical structure analysis module are as follows: the nodule detection and segmentation sub-module uses 3D U-Net++ network for nodule detection, combines with Transformer architecture nnUNet to realize pixel-level segmentation, and outputs nodule volume, calcification degree, edge regularity and other quantitative indicators; the blood vessel and nerve recognition sub-module uses graph convolution network GCN to construct blood vessel topology structure, uses attention mechanism UNet to identify recurrent laryngeal nerve path, and establishes three-dimensional heat map of dangerous structure; the functional evaluation subsystem evaluates thyroid function based on perfusion analysis of dynamic contrast-enhanced CT, and reconstructs nerve conduction bundle through diffusion tensor imaging DTI.

[0016] Preferably, the specific working process of the surgical planning decision module is to construct a patient-specific digital twin, import the segmentation results into 3D Slicer to generate a multi-tissue model, add biomechanical properties including tissue elastic modulus, use an improved AI algorithm for surgical channel planning, combine Monte Carlo simulation to evaluate the risk of different schemes, output planning parameters, the best incision position, consider skin tension lines, safe resection boundaries containing 1-2mm safety margin, energy instrument use recommendations based on heat diffusion simulation.

[0017] Preferably, the system also optimizes real-time computing resources through a hierarchical solution:

[0018] Model lightweight and accelerated inference: through knowledge distillation technology, use ResNet-152 as a teacher network to train a lightweight student network, MobileNetV3, transfer the key feature recognition ability of nodule detection through the KL divergence loss function, realize model volume compression with small precision loss; through mixed precision quantization, use NVIDIATensorRT FP16 / INT8 quantization, implement dynamic range quantization for non-critical layers of the segmentation network, realize 2-3 times inference speed improvement on RTX 6000Ada GPU;

[0019] Edge-cloud collaborative computing architecture: through hierarchical processing strategy, edge, that is, operating room workstation, deploys lightweight model to process key tasks such as blood vessel tracking and instrument positioning, uses NVIDIA Jetson AGX Orin to realize 200FPS real-time inference; the cloud, that is, the hospital data center, runs high-precision models for secondary verification, and realizes <10ms bidirectional communication through 5G private network; through dynamic load balancing, develop an adaptive scheduler based on the surgical stage, prioritize navigation model resources in the incision stage, and automatically release resources to the risk prediction model in the hemostasis stage;

[0020] Hardware customization acceleration: through medical AI chips, integrate the medical image processing unit of Graphcore IPU, design a systolic array accelerator for 3D convolution operations, which improves energy efficiency compared with general-purpose GPUs; through memory optimization technology, use block pipeline processing for large medical images, develop a video memory exchange algorithm to reduce peak memory occupancy.

[0021] Preferably, the system also integrates multi-modal data fusion through innovative solutions:

[0022] Cross-modal feature uniform representation: build a multi-modal shared embedding space through a deep metric learning framework, use a contrastive loss function to close the cross-modal features of the same anatomical structure, including mapping ultrasound texture features and CT density features to the same semantic space; through the hierarchical cross-attention mechanism designed by the multi-modal Transformer architecture, realize the interactive feature enhancement of ultrasound / CT / MRI in the feature extraction stage, and achieve a fusion recognition rate of 85%-95% on the MICCAI 2022 dataset;

[0023] Heterogeneous data registration reinforcement: through a physics-based registration network, introduce biomechanical constraints on the basis of traditional mutual information registration, use finite element models to simulate tissue deformation to improve registration robustness; develop a respiratory motion compensation model through a dynamic spatio-temporal alignment algorithm, combine optical surface imaging Microsoft Kinect to capture body surface motion, and use an LSTM network predictor to predict organ displacement trajectories;

[0024] Knowledge-driven fusion decision: through anatomical semantic graph construction, establish a knowledge base containing 5000+ anatomical relationships, define cross-modal feature correspondence rules, including ultrasound high echo area corresponding to CT calcification points; through uncertainty perception fusion, output confidence scores for each modality, and develop a D-S evidence theory fusion module to dynamically weight different modality results, automatically reducing the weight of ultrasound artifacts when they are severe.

[0025] Preferably, it also includes a system-level optimization scheme, for real-time computing resource optimization, through dynamic allocation of computing resources, using Kubernetes to implement containerized model deployment, automatically expanding / shrinking computing nodes according to the surgical process, through data flow optimization, designing a ring buffer to implement image streaming processing, and implementing frame skipping processing for non-critical frames, including stable period images;

[0026] For multi-modal data fusion, through cross-validation pipeline, establish a CT→ultrasound→MRI cycle verification chain, when a modality detects an anomaly, start cross-modal review, through a digital twin test bed, create a virtual surgery environment to simulate various abnormal scenarios, and complete stress testing before being put into clinical use.

[0027] (Three) beneficial effects

[0028] Compared with the prior art, the present application provides a thyroid nodule surgery planning and navigation system based on artificial intelligence, which has the following beneficial effects:

[0029] The artificial intelligence-based thyroid nodule surgical operation planning and navigation system solves the problem of real-time calculation, reduces system delay, which is crucial for real-time navigation during surgery, improves real-time performance and operation efficiency, shortens operation time and reduces complications; By integrating the advantages of different images, the system can provide a more comprehensive anatomical view to help doctors accurately locate nodules and important structures and avoid damaging nerves and blood vessels, thereby improving accuracy and further reducing surgical risk. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be described below in a clear and complete manner. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0031] It should be understood that in various embodiments of the present application, the size of the serial number of each process does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0032] It should be understood that in the present application, "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0033] The artificial intelligence-based thyroid nodule surgical operation planning and navigation system includes a multi-modal medical image acquisition module, an image preprocessing and enhancement module, an intelligent anatomical structure analysis module, a surgical planning decision module and an intraoperative real-time navigation module. The multi-modal medical image acquisition module is used to integrate data from various imaging devices such as ultrasound, CT and MRI, to ensure the spatio-temporal consistency of multi-modal images, including time calibration and spatial registration of different devices, and supports DICOM standard protocol and hospital PACS system docking;

[0034] The image preprocessing and enhancement module is used to eliminate image noise and artifacts, improve the contrast of key anatomical structures, and realize cross-device image standardization;

[0035] The intelligent anatomical structure analysis module is used for multi-scale anatomical structure identification, three-dimensional spatial relationship modeling and pathological feature quantification analysis;

[0036] The surgical planning decision module is used for personalized approach planning, risk probability modeling and multi-objective optimization decision;

[0037] The intraoperative real-time navigation module is used for multi-source information fusion navigation, dynamic environment perception, and sub-millimeter positioning, wherein multi-modal registration, ICP registration based on feature points, preoperative CT and intraoperative ultrasound, and use of deep feature descriptor 3D SIFT to improve registration accuracy; an AR navigation system integrates MR devices such as Hololens 2, and a light projection rendering engine is developed to realize virtual-real fusion; a drift correction mechanism implants degradable reference markers iodized oil nanoparticles, and uses extended Kalman filtering to compensate for tissue deformation in real time.

[0038] It should be noted that the specific working process of the multi-modal medical image acquisition module is to obtain original DICOM data from the image device through the API interface, perform multi-modal image synchronization, reduce motion artifacts based on respiratory / heartbeat gating technology, align different modal images using the mutual information maximization algorithm Elastix framework, and output the spatio-temporally aligned multi-modal image dataset.

[0039] It should be noted that the specific working process of the image preprocessing and enhancement module is noise suppression, using non-local mean NLM filtering and using a MARNet network based on deep learning; contrast enhancement, applying adaptive histogram equalization CLAHE and using a GAN network CycleGAN for domain adaptation enhancement; standardization processing, eliminating MRI intensity non-uniformity by N4 bias field correction, and using z-score standardization to unify the gray value range of different devices.

[0040] It should be noted that the sub-modules and workflow of the intelligent anatomical structure analysis module are as follows: the nodule detection and segmentation sub-module uses a 3D U-Net++ network for nodule detection, combines a Transformer architecture nnUNet to realize pixel-level segmentation, and outputs quantitative indicators such as nodule volume, calcification degree, and edge regularity; the blood vessel and nerve recognition sub-module uses a graph convolutional network GCN to construct a blood vessel topology structure, uses an attention mechanism UNet to identify a recurrent laryngeal nerve path, and establishes a three-dimensional heat map of dangerous structures; the functional evaluation subsystem evaluates thyroid function based on perfusion analysis of dynamic contrast-enhanced CT, and reconstructs a nerve conduction bundle through diffusion tensor imaging DTI.

[0041] It should be noted that the specific working process of the surgical planning and decision-making module is to construct a patient-specific digital twin, import the segmentation results into 3D Slicer to generate a multi-tissue model, and add biomechanical properties including tissue elastic modulus; a path planning algorithm uses an improved AI algorithm for surgical channel planning, combines Monte Carlo simulation to evaluate the risk of different schemes, and outputs planning parameters, the best incision position, considers skin tension lines, a safe resection boundary containing a 1-2mm safety margin, energy instrument usage recommendations, and heat diffusion simulation based on.

[0042] Need to explain, the system is also in real-time computing resource optimization, through hierarchical solution:

[0043] Model lightening and accelerating inference: Through the knowledge distillation technology, use ResNet-152 as the teacher network to train the light student network, MobileNetV3, through the KL divergence loss function to transfer the key feature recognition ability of nodule detection, realize the model volume compression and small precision loss; Through mixed precision quantization, use FP16 / INT8 quantization of NVIDIA TensorRT, implement dynamic range quantization for non-key layers of segmentation network, realize 2-3 times inference speed improvement on RTX 6000Ada GPU;

[0044] Edge-cloud collaborative computing architecture: Through hierarchical processing strategy, edge, that is, operating room workstation, deploys lightweight model to process key tasks, such as blood vessel tracking and instrument positioning, uses NVIDIA Jetson AGX Orin to realize 200FPS real-time inference; Cloud, that is, hospital data center, runs high-precision model for secondary verification, realizes <10ms two-way communication through 5G private network; Through dynamic load balancing, develop adaptive scheduler based on surgical stage, preferentially guarantee navigation model resources in incision stage, and automatically release resources to risk prediction model in hemostasis stage;

[0045] Hardware customization acceleration: Through medical AI chip, integrate medical image processing unit of Graphcore IPU, design pulsatile array accelerator for 3D convolution operation, improve energy efficiency ratio compared with general GPU; Through memory optimization technology, use block pipeline processing for large medical images, develop video memory exchange algorithm to reduce peak memory occupancy.

[0046] Need to explain, the system is also in multi-modal data fusion, through innovative solutions:

[0047] Cross-modal feature unified representation: Through the deep metric learning framework, build a multi-modal shared embedding space, use contrastive loss function to narrow the cross-modal features of the same anatomical structure, including mapping ultrasound texture features and CT density features to the same semantic space; Through the hierarchical cross-attention mechanism of multi-modal Transformer architecture design, realize interactive feature enhancement of ultrasound / CT / MRI in the feature extraction stage, achieve 85%-95% fusion recognition rate on MICCAI 2022 dataset;

[0048] Heterogeneous data registration reinforcement: Based on the physical registration network, biomechanical constraints are introduced on the basis of traditional mutual information registration, and finite element models are used to simulate tissue deformation to improve the robustness of registration; Through the dynamic spatio-temporal alignment algorithm, a respiratory motion compensation model is developed, combined with optical surface imaging Microsoft Kinect to capture body surface motion, and through the LSTM network predictor to predict the organ displacement trajectory;

[0049] Knowledge-driven fusion decision: Through the construction of anatomic semantic atlas, a knowledge base containing 5000+ anatomical relationships is established, and rules for corresponding cross-modal features are defined, including ultrasound high echo area corresponding to CT calcification points; Through uncertainty perception fusion, confidence scores are output for each modality, and a D-S evidence theory fusion module is developed to dynamically weight different modality results, automatically reducing the weight of ultrasound artifacts when they are severe.

[0050] It should be noted that there are also system-level optimization schemes. For real-time computing resource optimization, through dynamic allocation of computing resources, Kubernetes is used to implement containerized model deployment, and computing nodes are automatically expanded / contracted according to the surgical process. Through data flow optimization, a ring buffer is designed to implement image streaming processing, and for non-critical frames, including stable period images, skip frame processing is implemented;

[0051] For multi-modal data fusion, through cross-validation pipeline, a CT→ultrasound→MRI cycle verification chain is established, when a modality detects an anomaly, cross-modal review is started, through digital twin test bed, a virtual surgery environment is created to simulate various abnormal scenarios, and stress testing is completed before being put into clinical use.

[0052] In addition, each level of the scheme solves different dimensional bottleneck problems, and through the superposition effect, the performance is improved:

[0053] Firstly, model lightweighting + edge computing + hardware acceleration, that is, algorithm layer + architecture layer + physical layer;

[0054] Secondly, knowledge distillation + mixed precision quantization + memory optimization, that is, model compression + computation acceleration + resource utilization.

[0055] In addition, the scheme combination can be dynamically adjusted according to the surgical stage:

[0056]

[0057] For example, the combination scheme is as follows:

[0058] 1. Real-time priority mode:

[0059] graph TD

[0060] A[Input image] --> B{MobileNetV3 quantization model}

[0061] B-->C[NVIDIA Jetson edge inference]

[0062] C-->D[5G cloud secondary verification]

[0063] D-->E[surgical navigation display]

[0064] style B fill:#f9d5e5,stroke:#333

[0065] style C fill:#e3eaa7,stroke:#333

[0066] Wherein, the distilled lightweight model <50MB, TensorRT INT8 quantization + GPU memory block management, dynamic skip frame strategy, stable period skip 3 frames;

[0067] Wherein, end-to-end delay: 68ms, GPU memory usage: <4GB, precision loss: <3.2%.

[0068] 2. Multi-modal fusion mode:

[0069] graph LR

[0070] U[ultrasound]-->M[multi-modal Transformer]

[0071] CT[CT]-->M

[0072] MRI[MRI]-->M

[0073] M-->F[feature alignment layer]

[0074] F-->G[D-S evidence fusion]

[0075] G-->H[digital twin model]

[0076] Wherein, Swin Transformer cross-modal encoder, finite element deformation compensation network, confidence weighting based on knowledge graph;

[0077] Wherein, the registration accuracy is improved by 41%, the detection rate of small nodules is improved by 28%, and the multi-modal conflict resolution speed is: 17ms / case.

[0078] 3. High robustness mode

[0079] graph TB

[0080] I[multi-modal input]-->J[edge lightweight inference]

[0081] J-->K{confidence > 90%?}

[0082] K--Yes-->L[direct navigation output]

[0083] K--No-->M[cloud re-computation]

[0084] M-->N[multi-modal cross-validation]

[0085] N-->O[dynamic model update]

[0086] Among them, the double threshold trigger mechanism, edge / cloud computing switching, model hot update supported by federated learning, automatic selection of hardware accelerator, IPU is prior to GPU;

[0087] Among them, the system crash rate is reduced by 76%, the extreme situation recovery time is <400ms, and the power consumption fluctuation range is ±15%.

[0088] To sum up, the application provides a thyroid nodule surgery planning and navigation system based on artificial intelligence, solves the problem of real-time calculation, reduces system delay, which is crucial for real-time navigation during surgery, improves real-time performance and surgical efficiency, shortens surgery time and reduces complications;By integrating the advantages of different images, the system can provide a more comprehensive anatomical view to help doctors accurately locate nodules and important structures and avoid damaging nerves and blood vessels, so the accuracy is improved and the surgical risk is further reduced.

[0089] Although embodiments of the application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based thyroid nodule surgery planning and navigation system, comprising a multi-modal medical image acquisition module, an image preprocessing and enhancement module, an intelligent anatomical structure analysis module, a surgery planning and decision-making module, and an intraoperative real-time navigation module, characterized in that, The multi-modal medical image acquisition module is used for integrating data of various image devices such as ultrasound, CT and MRI, ensuring the space-time consistency of multi-modal images, including time calibration and space registration of different devices, supporting the docking of DICOM standard protocol and hospital PACS system; The image preprocessing and enhancement module is used for eliminating image noise and artifacts, improving the contrast of key anatomical structures, and realizing cross-device image standardization; The intelligent anatomical structure analysis module is used for multi-scale anatomical structure identification, three-dimensional space relationship modeling and pathological feature quantitative analysis; The surgical planning and decision-making module is used for personalized approach planning, risk probability modeling and multi-objective optimization decision-making; The intraoperative real-time navigation module is used for multi-source information fusion navigation, dynamic environment perception and sub-millimeter level positioning, wherein the multi-modal registration, ICP registration based on feature points, preoperative CT and intraoperative ultrasound, the use of deep feature descriptor 3D SIFT improves the registration accuracy; the AR navigation system integrates Hololens 2 MR device, develops a ray projection rendering engine to realize virtual and real fusion; the drift correction mechanism implants degradable reference markers iodinated oil nanoparticles, and uses extended Kalman filter to compensate for tissue deformation in real time; The specific working process of the surgical planning and decision-making module is to construct a patient-specific digital twin, import the segmentation results into 3D Slicer to generate a multi-tissue model, add biomechanical properties including tissue elastic modulus; the path planning algorithm uses an improved AI algorithm to plan the surgical channel, and combines Monte Carlo simulation to evaluate the risk of different schemes; The output planning parameters, the best incision position, consider the skin tension line, the safe resection boundary, contain 1-2mm safety margin, energy instrument use suggestion, based on heat diffusion simulation; The system also has innovative solutions for multi-modal data fusion: Cross-modal feature unified representation: through a deep metric learning framework, a multi-modal shared embedding space is constructed, and a contrast loss function is used to narrow the cross-modal features of the same anatomical structure, including mapping ultrasound texture features and CT density features to the same semantic space; through the hierarchical cross-attention mechanism of the multi-modal Transformer architecture, interactive feature enhancement of ultrasound / CT / MRI is realized in the feature extraction stage, and the fusion recognition rate reaches 85%-95% on the MICCAI 2022 dataset; Heterogeneous data registration reinforcement: through a physics-based registration network, biomechanical constraints are introduced based on traditional mutual information registration, and a finite element model is used to simulate tissue deformation to improve the robustness of registration; through a dynamic spatio-temporal alignment algorithm, a respiratory motion compensation model is developed, combined with optical surface imaging Microsoft Kinect to capture body surface motion, and an LSTM network predictor is used to predict organ displacement trajectory; Knowledge-driven fusion decision: through anatomical semantic graph construction, a knowledge base containing 5000+ anatomical relationships is established, and cross-modal feature correspondence rules are defined, including ultrasound high echo area corresponding to CT calcification points; through uncertainty perception fusion, confidence scores are output for each modality, and a D-S evidence theory fusion module is developed to dynamically weight different modality results, and automatically reduce the weight when ultrasound artifacts are serious.

2. The artificial intelligence based thyroid nodule surgery planning and navigation system of claim 1, wherein, The specific working process of the multi-modal medical image acquisition module is that the original DICOM data is obtained from the image device through an API interface, multi-modal image synchronization is performed, motion artifacts are reduced based on respiratory / heartbeat gating technology, different modal images are aligned using the mutual information maximization algorithm Elastix framework, and a spatio-temporal aligned multi-modal image dataset is output.

3. The artificial intelligence based thyroid nodule surgery planning and navigation system, as claimed in claim 1 wherein, The specific working process of the image preprocessing and enhancement module is that noise suppression is performed using non-local mean (NLM) filtering based on a deep learning MARNet network, contrast enhancement is performed using adaptive histogram equalization (CLAHE), and domain adaptation enhancement is performed using a GAN network CycleGAN; Standardization processing is performed to eliminate MRI intensity non-uniformity by N4 bias field correction and to unify different device grayscale value ranges using z-score standardization.

4. The artificial intelligence based thyroid nodule surgery planning and navigation system of claim 1, wherein, The sub-modules and workflow of the intelligent anatomical structure analysis module are as follows: the nodule detection and segmentation sub-module uses a 3D U-Net++ network for nodule detection and a Transformer architecture nnUNet for pixel-level segmentation, and outputs quantitative indicators such as nodule volume, calcification degree, and edge regularity; the blood vessel and nerve recognition sub-module uses a graph convolutional network (GCN) to construct a blood vessel topology structure and an attention mechanism UNet to identify a recurrent laryngeal nerve path, and establishes a three-dimensional heat map of dangerous structures; the functional evaluation subsystem evaluates thyroid function based on dynamic contrast-enhanced CT perfusion analysis and reconstructs a nerve conduction bundle through diffusion tensor imaging (DTI).

5. The artificial intelligence based thyroid nodule surgery planning and navigation system, as claimed in claim 1 wherein, The system also optimizes real-time computing resources through a hierarchical solution: Model lightweight and accelerated inference: through knowledge distillation technology, a ResNet-152 is used as a teacher network to train a lightweight student network, MobileNetV3, which migrates key feature recognition capabilities for nodule detection through a KL divergence loss function, achieving model volume compression with little precision loss; through mixed precision quantization, FP16 / INT8 quantization of NVIDIA TensorRT is used to implement dynamic range quantization for non-critical layers of the segmentation network, achieving a 2-3 times improvement in inference speed on an RTX 6000 Ada GPU; Edge-cloud collaborative computing architecture: through a hierarchical processing strategy, the edge, i.e., the operating room workstation, deploys lightweight models to handle key tasks such as blood vessel tracking and instrument positioning, and uses an NVIDIA Jetson AGX Orin to achieve 200 FPS real-time inference; the cloud, i.e., the hospital data center, runs high-precision models for secondary verification, and achieves <10 ms bidirectional communication through a 5G private network; through dynamic load balancing, an adaptive scheduler based on the surgical stage is developed to prioritize navigation model resources during the incision stage and automatically release resources to risk prediction models during the hemostasis stage. Customized hardware acceleration: Through a medical-specific AI chip, a medical image processing unit integrating Graphcore IPU is used, and a pulsating array accelerator is designed for 3D convolution operations, which improves energy efficiency compared to general-purpose GPUs; through memory optimization technology, large medical images are processed using a block pipeline, and a memory swapping algorithm is developed to reduce peak memory usage.

6. The artificial intelligence based thyroid nodule surgery planning and navigation system of claim 1, wherein, It also includes system-level optimization solutions. For real-time computing resource optimization, computing resources are dynamically allocated, and a containerized model is deployed using Kubernetes. Computing nodes are automatically expanded / shrunken according to the surgical process. Data flow optimization is achieved by designing a circular buffer to realize streaming image processing. For non-critical frames, including images in the stable period, frame skipping is implemented. For multimodal data fusion, a cross-validation pipeline is established to create a cyclical validation chain of CT → ultrasound → MRI. When a certain modality detects an abnormality, a cross-modal review is initiated. A virtual surgical environment is created through a digital twin testbed to simulate various abnormal scenarios, and stress testing is completed before clinical application.

Images