CN-122023974-A - Auditory neuron target extraction and positioning method based on vertical model

CN122023974ACN 122023974 ACN122023974 ACN 122023974ACN-122023974-A

Abstract

The invention discloses an auditory neuron target extraction and positioning method based on a vertical model. The method comprises the following steps of 1) preprocessing multi-mode data, 2) analyzing clinical symptom descriptions in the current medical history by means of a Gemini vertical model, 3) constructing a multi-mode feature fusion module, and 4) dynamically analyzing auditory neurons of neuron images of different age groups. The invention screens the neuron characteristics of the specific area in the image and identifies the information, overcomes the difficulties that the positioning precision of the auditory neurons in the image is low and the individual difference of the patient cannot be dynamically adapted in the prior art, and remarkably improves the accuracy and the efficiency of analysis. Meanwhile, the dynamic adaptation capability of the technology can optimize the analysis strategy aiming at the image signal characteristics of special crowds such as young, old and the like, and the fine analysis of auditory nerve mechanisms is promoted.

Inventors

WANG SHUYUE
ZHU ZHIJING
WANG RUIWEN
Hu Gelang
ZHANG YIHONG
Han Tianzhe
ZHANG HANYING

Assignees

浙大城市学院

Dates

Publication Date: 20260512
Application Date: 20251010

Claims (10)

1. The method for extracting and positioning the auditory neuron target based on the vertical model is characterized by comprising the following steps of: 1) The multi-mode data preprocessing comprises the steps of constructing a U-Net network containing a residual error attention module, realizing the accurate segmentation of neurons in a target area, further integrating a thousand-query hybrid reasoning model into a loss function in a characteristic weight output mode, and optimizing the weight distribution of the hybrid loss function; 2) Analyzing the clinical symptom description in the current medical history by means of a Gemini vertical model, namely carrying out structural analysis on the symptom description in the clinical history text by using a natural language processing technology of a Gemini large language model, establishing a space-time association relationship among symptoms by using a time sequence attention mechanism, and generating a standardized clinical term code; 3) Constructing a multi-modal feature fusion network, namely realizing the spatial alignment of clinical text keywords in a segmented region by adopting a cross-modal attention mechanism; 4) Dynamically analyzing auditory neuropathy conditions of neuron images of different age groups, namely acquiring images containing dynamic data issued by auditory cortex neurons, dividing the neuron images into three types of modes of young, adult and elderly according to the identification result of the step 2), adopting specific pretreatment for the images in different modes, and then combining the step 1) to the step 3), and finally outputting positions of pathological change neurons in the images, pathological change reasons and clinical symptoms of the pathological change neurons.
2. The method for extracting and locating an auditory neuron target based on a vertical model according to claim 1, wherein in the step 1), the specific steps of the multi-modal data preprocessing are as follows: Step 1.1) building a U-Net network consisting of an encoder-decoder The encoder path adopts three-level hierarchical feature extraction modules, each module comprises two-level 3X 3 convolution layers, batch normalization, ELU activation and Dropout regularization are sequentially carried out after each convolution layer, and then the space dimension of the feature map is gradually compressed to the original input 1/4 through 2X 2 maximum pooling of step length 2, and the channel dimension is synchronously expanded to 256, so that the gradual extraction of the context information is realized; The decoder path realizes feature reconstruction by combining bilinear interpolation and transposition convolution, wherein in each stage of up-sampling stage, the space dimension of a low-resolution feature map is expanded by 2X 2 transposition convolution and then is connected with a high-resolution feature jump of a corresponding layer of an encoder, and the fused features after jump connection are sequentially processed by 3X 3 convolution kernels of 16 channels and 8 channels and then are activated by 1X 1 convolution affine transformation and sigmoid to generate a segmentation result; Step 1.2) calculating a segmentation error according to real-time segmentation output by the U-Net network, dynamically adjusting a loss function by combining a thousand-number hybrid reasoning model, and optimizing weight distribution of the hybrid loss function; Step 1.3) calculating all neurons in the neuron image subjected to the segmentation denoising treatment one by using the following formula, so as to obtain an activity curve corresponding to each neuron: Wherein, the Representing the accumulation of all pixel values in the marked region of the current neuron over the current frame number, Representation of Average value of (2); step 1.4) comparing the emission of all the neuron calcium signals in the image, and judging whether the neurons in the image are lesion neurons according to the emission degree of the neuron calcium signals.
3. The method for extracting and locating an auditory neuron target based on a vertical model according to claim 2, wherein the step 1.4) specifically comprises: if the calcium signal emission degree of one neuron in the image is more than 50% weaker than that of other neurons, judging that the calcium signal is weakened, judging that the neuron is a pathological change neuron, and judging that the pathological change reason is that the signal emission is weakened; If the calcium signal emission degree of one neuron in the image is more than 50% of the calcium signal emission degree of other neurons, judging that the calcium signal is enhanced, judging that the neuron is a pathological change neuron, and judging that the pathological change reason is that the signal emission is over strong; If one neuron calcium signal issuing period in the image and other neuron issuing signal issuing periods are increased and are more than 2 times and more, determining that the neuron activity is inhibited, and determining that the neuron is a pathological change neuron, wherein the pathological change reason is that the neuron activity is inhibited; If one neuron calcium signal emission period in the image is obviously shortened from that of other neurons, and is less than 2 times or more, the neuron is judged to be over excited, and the neuron is judged to be a pathological neuron, and the pathological cause is that the neuron is over excited.
4. The method for extracting and locating an auditory neuron target based on a vertical model according to claim 1, wherein said step 2) specifically comprises: Step 2.1) adopting a Gemini vertical model to carry out intelligent processing on the current medical history text of a patient, relying on a medical knowledge base pre-trained by the model and the context understanding capability, automatically identifying clinical entities in the text, and establishing a three-dimensional semantic association network of symptoms, signs and disease processes; The medical knowledge base comprises rules for associating symptoms and signs of the acoustic nerve diseases, wherein the clinical entities comprise symptoms, signs and disease courses; Step 2.2) utilizing a multi-head self-attention mechanism to realize deep semantic deconstructment of the current medical history text through parallel multi-dimensional dynamic weight distribution.
5. The method for extracting and positioning the auditory neuron target based on the vertical model according to claim 4, wherein the step 2.2) is characterized in that each text establishes a dynamic connection weight through a query-key value interaction system, the obtained dynamic connection weight obtains a position vector through sine and cosine function calculation, absolute time sequence information is injected through the position vector, so that the Gemini vertical model accurately perceives a time span in symptom description; The functions of the different attention heads are respectively to concentrate on the symptom and sign modification relation, track the semantic jurisdiction of the negative word and establish the cross-sentence logic association; The query-key value interaction system has the functions that each text is mapped into a unique key based on a hash function, an associated dynamic weight value is obtained through dot product attention calculation of a query vector and a pre-stored key vector and is stored in a memory database, a bloom filter is further adopted to optimize query efficiency and integrate an LRU cache strategy to manage high-frequency access data, instantaneity and accuracy of weight calculation are ensured through a distributed consistency protocol, and finally dynamic connection weights are established.
6. The method for extracting and locating an auditory neuron target based on a vertical model according to claim 1, wherein the constructing a multi-modal feature fusion network in the step 3) specifically comprises: performing spatial alignment and weight distribution on the auditory calcium signals preprocessed in the step 1) and the clinical text semantic features subjected to entity identification and feature normalization in the step 2) by adopting a cross-modal attention mechanism, wherein the preliminary alignment of the calcium signal time sequence features and the text symptom time sequence is realized through a dynamic time warping algorithm, and then the attention score is utilized to optimize the spatial mapping precision; if the neuron calcium signal in the image is a juvenile neuron calcium signal, the spatial alignment and the spatial mapping between the juvenile neuron calcium signal and the clinical text semantic features are enhanced by adjusting the cross-modal attention mechanism parameter.
7. The method for extracting and locating auditory neuron targets based on vertical model according to claim 1, wherein in the step 4), the auditory neuropathy conditions of different age groups are dynamically analyzed, and the following steps are performed The steps for dynamically analyzing the auditory neuropathy of the neuron images of different age groups are as follows: acquiring an image containing dynamic data issued by auditory cortex neurons, and dividing the neuron images in the image into three types of modes of young, adult and old according to the identification result of the step 2); For the young auditory cortex neuron release image, firstly denoising, then dividing and identifying and extracting a middle ear neuron calcium signal by adopting the method of the step 1), obtaining a structured clinical text by adopting the text identification and the feature normalization of the step 2), obtaining the association rule of the neuron calcium signal and the clinical text based on the step 3), and finally outputting the positions of pathological change neurons, pathological change reasons of the pathological change neurons and clinical symptoms of the pathological change neurons; for the adult auditory cortex neuron release image, segmenting and identifying and extracting a middle ear neuron calcium signal by adopting the method of the step 1), obtaining a structured clinical text by text identification and feature normalization of the step 2), obtaining a correlation rule of the neuron calcium signal and the clinical text based on the step 3), and finally outputting the positions of pathological change neurons, pathological change reasons of the pathological change neurons and clinical symptoms of the pathological change neurons; for the aged auditory cortex neuron issuing image, adopting a time sequence alignment correction network, extracting the time-space characteristics of a calcium signal by utilizing a three-dimensional convolution layer, maintaining the time sequence integrity through space pooling of a reserved time dimension, capturing the context dependency relationship of signals at a bidirectional LSTM layer, finally outputting a time sequence corrected by a time distributed full-connection layer, then adopting the method of the step 1) to segment and identify the calcium signal of the middle ear neuron of the extracted image, obtaining a structured clinical text through the text identification and the characteristic normalization of the step 2), further obtaining the association rule of the calcium signal of the neuron and the clinical text based on the step 3), and finally outputting the positions of pathological neurons in the image, the pathological cause and the clinical symptoms of the pathological neurons.
8. The method for extracting and locating auditory neuron targets based on vertical model according to claim 7, wherein for the young auditory cortex neuron emission image, the denoising process is specifically: The method comprises the steps of constructing a double-light-path imaging frame, taking an image obtained on a light path at one side of a low signal-to-noise ratio as training source data, taking an interval frame as input and an intermediate frame as a supervision target by adopting a discontinuous frame jump sampling strategy, establishing a self-supervision training paradigm without manual marking, inputting the training source data into a convolution neural network based on a symmetrical topological structure to perform preliminary denoising treatment, sequentially forming each unit of the convolution neural network based on the symmetrical topological structure by two convolution layers adopting 3X 3 convolution kernels and a 2X 2 maximum pooling layer in a downsampling stage, correspondingly arranging the two 3X 3 convolution layers in the upsampling stage to realize resolution improvement by matching with transposed convolution, introducing nonlinear characteristic transformation between all levels through a ReLU activation function, and using a Dropout layer with 0.2 probability to prevent overfitting, and further constructing an evaluation system through a reference image obtained by a reference light path with a high signal-to-noise ratio, calculating peak signal-to-noise ratio between the reference image and the preliminary denoising result, so as to perform quantization verification and realize noise suppression of a neuron image.
9. An auditory neuron target extraction and positioning device based on a vertical model is characterized by comprising a memory and a processor; the memory is used for storing a computer program; The processor is configured to implement the vertical model-based auditory neuron lesion localization method according to any one of claims 1 to 8 when the computer program is executed.
10. A computer readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the method for extracting and locating an auditory neuron target based on a vertical model according to any one of claims 1 to 8 is implemented.

Description

Auditory neuron target extraction and positioning method based on vertical model Technical Field The invention belongs to the technical field of intersection of neuroscience and artificial intelligence, and particularly relates to an auditory neuron target extraction and positioning method based on a vertical model. Background At present, the vertical model has remarkable advantages in the field of medical prediction, namely the nonlinear characteristics of multi-mode data can be automatically extracted through a deep neural network architecture, the limitation that the traditional statistical method depends on artificial characteristic engineering is overcome, meanwhile, the vertical model is added into a loss function to dynamically adjust the loss weight according to individual heterogeneity and noise interference of the medical data, the deviation of prediction of key clinical indexes is effectively reduced, the problem of insufficient suitability of the traditional fixed loss function is solved, the capturing capability of the model to weak correlation characteristics in the medical data is further enhanced, the reliability of prediction results under the scenes of rare diseases, complex complications and the like is improved, and better technical support is provided for clinical accurate diagnosis and treatment decision. In addition, the feature fusion analysis of special crowds is dynamically optimized, and compared with the traditional analysis method, the generalization performance is remarkably improved. Furthermore, the continuous learning mechanism of large language models enables them to dynamically adapt to new clinical data distributions, whereas traditional approaches often require re-modeling. One of the core pathogenic mechanisms of nerve deafness is impaired acoustic nerve conduction, and acoustic nerve conduction function depends on normal electrophysiological activity of the Spiral Ganglion Neurons (SGNs) in the cochlea—when a specific SGNs has a functional error (such as abnormal discharge frequency, delayed conduction, and interruption of signal transmission), an acoustic signal cannot normally pass into the central nervous system, and deafness is induced. In the prior art, auditory Brainstem Response (ABR), cochlear electrogram (ECochG) and the like can only judge whether cochlear nerve conduction is damaged, which or which group SGNs is wrong cannot be accurately positioned, while the traditional artificial intelligent model (such as CNN and LSTM) attempts to position abnormal neurons through electrophysiological signals, the traditional artificial intelligent model has two defects that firstly, the model adopts a fixed loss function and cannot adapt to individual differences of electrophysiological signals of different auditory neurons (such as signal-to-noise ratio, conduction delay fluctuation and weak degree of abnormal signals are different), so that the positioning accuracy of the model to part of auditory neurons is insufficient, and secondly, the model cannot capture complex correlation characteristics between SGNs electrophysiological signals and auditory dysfunction, and cannot realize accurate mapping of pathological neuron-hearing abnormal symptoms. Disclosure of Invention In order to solve the problems in the background technology, the invention provides an auditory neuron target extraction and positioning method based on a vertical model, which dynamically adapts to individual data of a patient and accurately positions pathological auditory neurons of an image by fusing a large language model and a deep learning network. Therefore, the invention adopts the following technical scheme: Auditory neuron target extraction and positioning method based on vertical model The method comprises the following steps: 1) The multi-mode data preprocessing comprises the steps of constructing a U-Net network containing a residual error attention module, realizing the accurate segmentation of neurons in a target area, further integrating a thousand-query hybrid reasoning model into a loss function in a characteristic weight output mode, and optimizing the weight distribution of the hybrid loss function; 2) Analyzing the clinical symptom description in the current medical history by means of a Gemini vertical model, namely utilizing a natural language processing technology of a Gemini large language model to carry out structural analysis (automatically identifying key elements such as symptom duration, severity and the like) on the symptom description in the clinical current medical history text, establishing a time-space association relationship among symptoms by adopting a time sequence attention mechanism, and generating a standardized clinical term code; 3) Constructing a multi-modal feature fusion network, namely realizing the spatial alignment of clinical text keywords in a segmented region by adopting a cross-modal attention mechanism; 4) Dynamically analyzing auditory neuropathy