CN-122023757-A - Auxiliary mechanical arm system for guiding positioning of hematoma under neuroendoscopy based on deep learning

CN122023757ACN 122023757 ACN122023757 ACN 122023757ACN-122023757-A

Abstract

The invention discloses an auxiliary mechanical arm system for guiding positioning of hematoma under a neuroendoscopy based on deep learning, and belongs to the field of computer vision and medical equipment. The invention utilizes YOLOv deep learning algorithm and medical image three-dimensional reconstruction and depth fusion of robot control technology, innovatively solves the difficult problems of automatic and accurate calculation of cerebral hemorrhage focus volume, remarkably improves the planning efficiency, operation precision and overall safety of brain surgery, provides a brand new technical scheme for realizing accurate and minimally invasive intelligent surgery, and has extremely high clinical application value and wide market potential.

Inventors

XIE ZONGYI
HU JIANGPING
He Shengwan
DENG JIACHENG
ZHANG ZHONGYI
WU HAOZE

Assignees

电子科技大学
重庆医科大学附属第二医院

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (4)

1. An auxiliary mechanical arm system for guiding the positioning of hematoma under a neuroendoscopy based on deep learning mainly comprises a medical image acquisition module, a focus detection module, a three-dimensional reconstruction module, an operation planning module and a mechanical arm control and execution unit; the medical image acquisition module acquires brain CT images in real time; the focus detection module identifies the accurate position, the spatial morphology and the total volume of the whole hematoma and the edema according to the acquired brain CT image, and the specific method comprises the following steps: The method comprises the steps of 1, marking focus features in CT images by adopting CT image data containing a large number of cerebral hemorrhage cases as a training set, training YOLOv a deep learning model based on the data set, so that the focus regions in the craniocerebral CT images and the boundary and range of peripheral oedema can be accurately and rapidly identified; Step 2, inputting CT tomographic images of the brain of a patient into a trained YOLOv model to analyze each two-dimensional CT image, identifying hematoma focus and edema areas in the CT images in real time, and outputting accurate boundary frame coordinates and size information in the two-dimensional images; the specific method of the three-dimensional reconstruction module comprises the following steps: Step 3, accumulating continuous two-dimensional CT image sequences, namely accumulating the volumes layer by layer according to a focus boundary frame identified by YOLOv on each slice and combining layer thickness parameters of CT, so as to reconstruct the accurate position, spatial shape and total volume of the whole hematoma and edema in a three-dimensional space; The operation planning module plans the optimal operation path according to the accurate position, the spatial morphology and the total volume of the whole hematoma, and the specific method comprises the following steps: step 4, adopting a three-dimensional focus model to automatically plan an optimal operation path at the cost of minimizing damage to healthy brain tissues; Step 4.1, calculating a comprehensive risk cost function for the mechanical arm by using the three-dimensional digital focus model generated in the step 3 through an optimization algorithm Surgical path with optimal value ; ; Wherein P represents a candidate surgical path, Is the cost of the geometric length of the path, Is the risk cost of the path through or near advanced brain function areas, which increases dramatically as the path approaches these areas, the detailed calculation is: ; In the formula, To identify the total number of key brain function regions, Is the first The risk weight coefficients of the individual functional areas, For sampling points on path P From the functional area The shortest euclidean distance of the boundary, Is a preset tiny positive number to avoid zero denominator, Representing the distance differentiation; The method is characterized in that the path has the convenience and accessibility cost for completely eliminating the focus, and the calculation method comprises the following steps: ; In the formula, By minimizing the included angle between the path vector and the major long axis direction of the hematoma three-dimensional model Making the path as deep as possible longitudinally along the long axis of hematoma to maximize the operative field under the neuroendoscopy; Is the end point of the path Geometric centroid of hematoma Is used for ensuring that the puncture channel directly reaches the focus core; The weight is adjusted for the proportion of the corresponding sub-item; Weighting coefficients for each cost: the control path is simple, and unnecessary wounds are prevented from being increased due to overlong paths; Representing the security weight, and determining the importance degree of the system on avoiding the functional area; Representing the effectiveness weight, determining whether the path is convenient for the doctor to perform hematoma removal operation, and performing iterative search by the algorithm until a cost function is found Paths that reach minimum or converge to an optimal solution The path is determined as the optimal surgical approach, which precisely defines the locations of the scalp incision, the skull opening, and the trajectory and depth of the surgical instrument into the brain parenchyma; Step 4.2, in determining the optimal Path Then, the mechanical arm control system converts the coordinate point sequence of the path in the Cartesian space into an angle sequence required to rotate each joint of the mechanical arm through inverse kinematics calculation, and adopts the following controller to correct the motion error in real time and control the motion instruction of the mechanical arm end effector Is generated by the following formula: ; Wherein, the Representing the error vector, being the desired position And the actual position Deviation between, i.e , From the surgical planning module in step 4.1, Sensor feedback and real-time calculation from the robotic arm itself; Representing control parameters in a PID control method; The mechanical arm control and execution unit is used for executing the operation according to the surgical path and the instruments required by the current operation.
2. The auxiliary mechanical arm system for guiding the positioning of the hematoma under the neuroendoscopy based on the deep learning according to claim 1, wherein the deep convolution neural network in the step 1 is a YOLOv11 deep learning model, CT image data containing a large number of cerebral hemorrhage cases is adopted as a training set, marking tools such as Labelimg are adopted, the positions and the ranges of cerebral hemorrhage focuses are accurately marked in CT images by rectangular bounding boxes, the deep convolution neural network is constructed, network parameters are optimized by adopting a loss function, and a YOLOv model for identifying cerebral hemorrhage and edema area ranges by the CT images is trained.
3. The system of claim 1, wherein the focus detection module in step 2 analyzes the images one by one after receiving the CT image sequence, the pre-trained YOLOv model in step 1, rapidly identifies the region of the cerebral hemorrhage focus and outputs the position bounding box thereof, and the system only adopts the detection result higher than the preset confidence threshold to ensure the reliability, and further eliminates the artifact interference by verifying the high-density imaging characteristics of the region, thereby improving the accuracy and the reliability of the positioning.
4. The auxiliary mechanical arm system for guiding the positioning of the neuroendoscopic hematoma based on the deep learning according to claim 1, wherein the specific method of the step 3 is as follows: step 3.1, firstly, carrying out serialization processing on all continuous two-dimensional CT images of the focus boundary frame identified by the step 2, and reading the following geometric parameters from DICOM metadata of each image: 、 Scanning the layer thickness T; Representing the physical width of a single pixel, Representing the physical height of a single pixel, step 3.2: for any slice in the sequence Calculating the physical area of the model on a two-dimensional image by using the pixel coordinates of the boundary frame output by the trained YOLOv model This calculation is done by the following formula: ; Wherein, the And (3) with The pixel width and height of the bounding box, respectively; Step 3.3 after calculating the physical areas of all relevant slices, the areas are integrated and accumulated and multiplied by the scanned layer thickness Finally, the total volume of the target focus is obtained The process is represented by the following formula: ; Wherein, the The total number of the sections of the focus is detected, and the complete space shape, position and total volume of the target focus are finally reconstructed in a three-dimensional space coordinate system to form a digital three-dimensional model which can be used for planning a subsequent operation path.

Description

Auxiliary mechanical arm system for guiding positioning of hematoma under neuroendoscopy based on deep learning Technical Field The invention belongs to the field of computer vision and medical instruments, and particularly relates to a system and a method for realizing accurate positioning and three-dimensional reconstruction of a focus by analyzing a human brain CT image based on YOLOv11 deep learning algorithm and guiding a mechanical arm to assist in completing intracranial hematoma removal under a cerebral hemorrhage neuroendoscopy. Background YOLOv11 has high efficiency and real-time performance, and can identify the target in the image with high precision in a short time through a single neural network architecture. The algorithm has great application value in various directions such as intelligent manufacturing, image recognition, intelligent traffic and the like. Cerebral hemorrhage, also called intracranial hemorrhage (Intracerebral Hemorrhage, ICH), is a disease in which blood vessels in brain parenchyma are ruptured, blood overflows and accumulates inside brain tissue, and is an acute cerebrovascular disease with high disability rate and high mortality rate in neurosurgery. Patients meeting the operation instruction need to go through intracranial hematoma removal operation as early as possible, the success rate of surgery and prognosis of patients are indistinguishable from the clinical experience of the doctor's home doctor and interpretation ability of CT images. The neuroendoscopic hematoma removal is a modern minimally invasive operation, has the characteristics of small wound and quick recovery, but has higher technical requirements on operators, doctors need to construct the three-dimensional form and the spatial position of a focus through two-dimensional CT images and draw an operation path, subjective deviation exists in the process, and the peripheral healthy brain tissues can be damaged in the operation or the hematoma is incompletely removed, so that the prognosis of a patient is influenced. In recent years, although neural navigation and surgical robot techniques have been applied to clinics, the accuracy of surgery is improved to some extent, most of the existing auxiliary systems still rely on preoperative planning, and the capability of real-time registration and coping with emergency situations such as brain tissue displacement in surgery is limited. Furthermore, the precise definition of the extent of the lesion and its surrounding oedema remains a key and difficult point of surgical success. Therefore, it is important to deeply blend the high precision surgical needs with advanced visual recognition techniques. The realization of full-flow automatic assistance from two-dimensional image analysis, three-dimensional focus modeling to mechanical arm path planning and operation is still a technical problem to be solved urgently in the field. How to accurately read medical images by using artificial intelligence and convert the information into an executable accurate operation instruction of an operation mechanical arm is a key for improving the accuracy of preoperative judgment of hematoma removal under a neuroendoscope and improving the safety and effectiveness of the operation. Disclosure of Invention In order to solve the problems of insufficient positioning accuracy, excessive dependence on experience of a surgeon and the like in the prior art, the invention provides an auxiliary mechanical arm system and an auxiliary mechanical arm method for guiding a neuroendoscopic hematoma removal operation based on deep learning. The invention aims at intelligently analyzing brain CT images by utilizing YOLOv-11 deep learning model, accurately identifying and three-dimensionally reconstructing hematoma focus, further guiding a mechanical arm to complete a series of high-precision operation, and improving the accuracy, safety and minimally invasive level of the operation to a new height. The technical scheme of the invention is that the auxiliary mechanical arm system for guiding the positioning of hematoma under the neuroendoscopy based on deep learning mainly comprises a medical image acquisition module, a focus detection module, a three-dimensional reconstruction module, an operation planning module and a mechanical arm control and execution unit; the medical image acquisition module acquires brain CT images in real time; the focus detection module identifies the accurate position, the spatial morphology and the total volume of the whole hematoma and the edema according to the acquired brain CT image, and the specific method comprises the following steps: The method comprises the steps of 1, marking focus features in CT images by adopting CT image data containing a large number of cerebral hemorrhage cases as a training set, training YOLOv a deep learning model based on the data set, so that the focus regions in the craniocerebral CT images and the boundary and range of peripheral oed