Craniocerebral injury diagnosis and treatment aid decision-making system based on medical big data

Abstract

The invention discloses a craniocerebral injury diagnosis and treatment auxiliary decision-making system based on medical big data, particularly relates to the field of computer-aided diagnosis, and is used for solving the problem of inaccurate structural evolution interpretation in the existing craniocerebral injury diagnosis and treatment process. Acquiring a craniocerebral injury CT image sequence in a complete diagnosis and treatment period through a trajectory construction module, and extracting brain density values to construct a structure evolution trajectory sample library; identifying an evolution abnormity feedback signal through an abnormity identification module; a conflict triggering module introduces a historical brain function state of the patient, and judges whether a mode conflict between the structure change and the function evaluation exists or not; if the conflict exists, a cross mechanism reasoning chain is reconstructed by a reasoning reconstruction module; the model training module is used for training an injury mechanism identification model based on the reconstructed reasoning chain and the function state score; finally, the decision output module assists in pathological mechanism judgment according to the recognition result, and the diagnosis and treatment decision accuracy under the complex injury condition is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided diagnosis, more particularly, to a craniocerebral injury diagnosis and treatment auxiliary decision system based on medical big data. BACKGROUND

[0002] The existing craniocerebral injury diagnosis and treatment auxiliary method mainly relies on artificial interpretation of CT images and empirical evaluation of pathological changes, lacks systematic modeling of structural evolution process, and is difficult to accurately identify structural abnormalities in the patient's disease course and combine functional state for mechanism reasoning. In clinical practice, the manifestation of craniocerebral injury is complex, and may be caused by the interweaving of multiple pathological mechanisms. Static comparison of a single structural feature is not enough to reveal the potential conflict phenomenon in the complex evolution process, which easily leads to mechanism identification bias. In addition, the current lack of a mechanism identification system that unifies the structural evolution path and brain function score cannot logically model and conflict identify the inconsistency between structural changes and functional performance, limiting the intelligent level of auxiliary decision-making. In the existing method, only the static anatomical region comparison of the image sequence is performed, the time sequence evolution characteristics of the structural density are ignored, the trajectory structure for different pathological mechanism formation paths is not established, and it is also impossible to infer the structural migration trend under the cross action of multiple mechanisms after identifying abnormal evolution. Especially when the structural change direction and the functional state evaluation result do not match, the existing scheme lacks dynamic reasoning ability and is difficult to realize mechanism reconstruction and auxiliary judgment.

[0003] In view of the above problems, it is urgent to build an intelligent diagnosis and treatment auxiliary method based on structural evolution trajectory, which integrates functional score trend for abnormal identification, conflict detection and mechanism reasoning reconstruction. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a craniocerebral injury diagnosis and treatment auxiliary decision system based on medical big data to solve the problems raised in the background art.

[0005] To achieve the above object, the present application provides the following technical scheme: A craniocerebral injury diagnosis and treatment auxiliary decision system based on medical big data, comprising: A trajectory construction module is used to acquire CT image sequences of craniocerebral injury in a complete diagnosis and treatment cycle in medical big data, extract brain region density from the images, and construct a structural evolution trajectory sample library of the brain region based on time sequence; An abnormality identification module is used to acquire CT images of a patient to construct a structural change path, compare with the structural evolution trajectory sample library, and identify evolution abnormal feedback signals in the structural trajectory; a conflict triggering module, based on the structural abnormality feedback signal, extracting patient historical brain function state data, detecting whether there is a pattern conflict between the structural change direction and the function evaluation; a reasoning reconstruction module, when there is a pattern conflict, discarding the single mechanism evolution path hypothesis, and reconstructing a reasoning chain based on multi-mechanism cross; a model training module, based on the cross-mechanism reasoning chain and the brain function state, training an injury mechanism identification model; a decision output module, based on the injury mechanism identification model output result, assisting in decision-making for the actual patient pathological mechanism.

[0006] In a preferred embodiment, the brain region density extraction of the image in the trajectory construction module is based on the time sequence to construct the brain region structural evolution trajectory sample library in the following specific manner: The CT image sequence of the craniocerebral injury is sorted in time sequence, and the continuous sample group with image quality meeting the imaging standard is screened; The density value of the brain region is extracted by registration processing of each image, and the density value is the gray value corresponding to the brain tissue region in the CT image; Based on the density value change of adjacent time points, a density change sequence is constructed, clustered into evolution path clusters according to the same mechanism label, and a structural trajectory sample library with time and mechanism label is formed; The density change sequence that cannot be clustered into any mechanism label is integrated to obtain the historical medical record of the unknown injury patient; The mechanism label is a classification index mark of the evolution path cluster based on medical diagnosis results.

[0007] In a preferred embodiment, in the abnormality recognition module, the process of recognizing the evolution abnormality feedback signal in the structural trajectory is as follows: Obtain the continuous CT image sequence collected by the unknown injury patient during the treatment period, extract the brain region density change sequence through image space registration, and construct a continuous structural change path in time sequence; The structural change path is compared with all the trajectory samples in the structural evolution trajectory sample library point by point, and a similarity score matrix is constructed using three indexes of density change gradient, boundary migration amplitude and center symmetry offset; When the score result of any comparison point is lower than the preset similarity threshold value, and the direction of the density change gradient is reversed or the symmetry is mutated, the evolution abnormality feedback signal is output.

[0008] In a preferred embodiment, in the conflict triggering module, detecting whether there is a pattern conflict between the structural change direction and the function evaluation specifically includes: extracting an unknown injury patient's historical brain function state evaluation score, the evaluation score being a comprehensive score including standardized neurological function, movement disorder and cognitive ability; After obtaining the evolution abnormal feedback signal, the structural change path established in the abnormality recognition module is extracted and converted into a density change direction vector expression, which is compared with the time curve slope direction of the historical brain function state evaluation score; When the density change direction is inconsistent with the expected change direction of the brain function state, it is marked as a conflict, and if the number of conflict states appearing in the treatment period is more than the set number of indicators, it is determined that there is a mode conflict.

[0009] In a preferred embodiment, in the reasoning reconstruction module, the reconstruction of the reasoning chain based on multi-mechanism cross includes: After recognizing the mode conflict, the structural evolution hypothesis based on a single mechanism evolution track is discarded, a number of track samples with the highest similarity to the CT image sequence of the unknown injury patient are extracted from the structural evolution track sample library, and candidate paths corresponding to different mechanisms are constructed according to the mechanism labels; In the candidate path, the track samples are segmented by time stage based on the density change direction inflection point of the functional state change area, and mechanism track stages divided by time are generated; Based on the occurrence time of the mode conflict, the track stages in each candidate path are re-mapped with the structural change path of the patient's CT image at the segment level, and the dynamic linking of the mechanism track stage with the highest mapping similarity is performed; The above linking process is recursively executed, and a cross-mechanism reasoning chain is constructed according to the possible structural combinations of different mechanisms.

[0010] In a preferred embodiment, the mechanism is a pathological mechanism, pathological change process or medically defined symptom model of a specific medical trauma type corresponding to the mechanism label.

[0011] In a preferred embodiment, the model training module, based on the cross-mechanism reasoning chain and the brain function state, trains an injury mechanism recognition model, specifically including: All unknown injury patient corresponding cross-mechanism reasoning chains are fragmented, and a graph structure reasoning network is constructed according to the mechanism label and the time stage; Obtain the brain function state evaluation score sequence of all unknown injury patients within the treatment period, map the score sequence to the structural change node in the reasoning network according to the time stage, and construct a guided path system with the brain function state evaluation score as the edge weight factor; Train the graph structure path recognition model, and the model uses a time series graph based attention mechanism to calculate the path passing probability.

[0012] In a preferred embodiment, the decision output module, based on the injury mechanism identification model output result, assists in decision-making of the actual patient pathological mechanism specifically includes: Input the continuous CT images of the actual patient in the current treatment cycle into the image processing flow, and extract the density change sequence of the brain region; According to the node setting of the graph structure reasoning network, the density change sequence is mapped to the corresponding structure change node in the graph, and path reasoning is performed according to the edge weight factor on the node connection edge; The mechanism label associated with the reasoning path with the highest passing probability is statistically output as the structural injury mechanism identification result.

[0013] The technical effect and advantages of the craniocerebral injury diagnosis and treatment auxiliary decision system based on medical big data of the present application are: By constructing a sample library based on the structural density evolution path, and fusing the brain function state evaluation trend, the potential mode conflict between the structural abnormalities and functional abnormalities of the craniocerebral injury patient can be automatically identified, and after detecting the mode conflict, a reasoning chain is constructed based on the multi-mechanism cross path, realizing multi-reasoning and reconstruction of the pathological mechanism. Compared with the existing method, the present scheme breaks through the limitation of single mechanism modeling, has higher pathological adaptability and structural evolution identification ability. At the same time, by introducing a graph structure reasoning network constructed based on mechanism labels and time stages, and combining with the guided path modeling of the brain function score trend, the modeling ability of the corresponding relationship between structural changes and functional performance is significantly improved, and the identification ability of the model to complex injury mechanism is effectively enhanced. The finally output mechanism identification result not only improves the adaptability of the actual patient structural change path, but also provides mechanism relocation suggestions and examination guidance areas for the clinic, which helps to improve the diagnosis and treatment accuracy and efficiency, has strong practical value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The present application is a schematic diagram of a craniocerebral injury diagnosis and treatment auxiliary decision system based on medical big data. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] Embodiment 1 Figure 1 The present application is a craniocerebral injury diagnosis and treatment auxiliary decision system based on medical big data, which comprises: A trajectory construction module is configured to acquire CT image sequences of brain injury in a complete diagnosis and treatment cycle in medical big data, extract brain region density from the images, and construct a structural evolution trajectory sample library of the brain region based on time series; An anomaly identification module is configured to acquire CT images of a patient to construct a structural change path, compare the structural change path with the structural evolution trajectory sample library, and identify an evolution anomaly feedback signal in the structural trajectory; A conflict triggering module is configured to extract historical brain function state data of the patient based on the structural anomaly feedback signal, and detect whether there is a mode conflict between the structural change direction and the function evaluation; A reasoning reconstruction module is configured to discard a single mechanism evolution path assumption and reconstruct a reasoning chain based on a cross-mechanism when there is a mode conflict; A model training module is configured to train an injury mechanism identification model based on the cross-mechanism reasoning chain and the brain function state; A decision output module is configured to output a result based on the injury mechanism identification model to assist in decision-making for an actual patient pathological mechanism.

[0017] In the trajectory construction module, brain region density is extracted from the images, and a structural evolution trajectory sample library of the brain region is constructed based on time series.

[0018] Through an image retrieval process oriented to a medical image archiving database, CT image sequences related to brain injury of a target patient in a complete diagnosis and treatment cycle are acquired. The sequences need to include all brain CT examination records of the patient from the first visit to the latest discharge or review, and the time span usually covers at least two different visit stages to reflect the structural change trend. The retrieved images need to carry timestamp information to ensure that each image has an accurate shooting time record for subsequent time series construction. After the image retrieval is completed, an image sample quality screening process is performed. The process screens the image samples through standardized image quality control indicators, including but not limited to image resolution (which needs to be greater than 512x512 pixels), signal-to-noise ratio (which needs to be higher than 20 dB), artifact interference score (which cannot be higher than level 1), and brain tissue edge definition score (which needs to be greater than a set threshold of 120 using a gradient amplitude algorithm). If the CT image sample at a certain time point does not meet the above imaging standards, the image is not included in the continuous sample group. Through the evaluation of the above indicators one by one, a continuous image sequence with stable image quality and reasonable time interval is screened out in the diagnosis and treatment cycle of the same patient, usually ensuring not less than 5 time points and not more than 7 days for each time interval.

[0019] After obtaining the continuous sample set, image registration processing is performed on each CT image to ensure the spatial consistency of subsequent structure comparison and evolutionary calculation. In the specific processing flow, first, the CT image with the best imaging quality is selected as the reference base image, and then rigid registration and affine transformation are performed on all the remaining images to ensure the alignment of the brain region in each image. The standard image resampling method is used in the registration process to unify all images to the same spatial scale, and the structural edge overlap rate is used to evaluate the registration accuracy to ensure that the structural alignment error is not higher than 20 pixels. After registration is completed, the brain parenchymal region and the ventricular region are located in the image according to the standard brain atlas coordinate system, and are segmented in the form of a structural region mask. The density value extraction process is performed by region: for each structural region, all pixels in the region are traversed respectively, and the arithmetic mean of their gray values is calculated. That is, the density value is the average of the gray values of the pixels covered by the brain tissue region (such as the frontal lobe, temporal lobe, and ventricle) in the CT image, which reflects the tissue density performance of the region at the current time point. For example, in the brain parenchymal region, if about 3500 valid pixels are extracted, and the mean of their gray values is 123.7, this value is taken as the density index of the structural region of the current image. This process is repeated in multiple time point images to obtain the sequence of changes in the density values of the structural regions over time.

[0020] After the density value sequence is constructed, the density change trend of adjacent time dimensions needs to be quantitatively modeled to form the density change sequence. Specifically, for each structural region, the average density values at consecutive time points are arranged in chronological order to form the original density sequence. Then, the density difference between adjacent time points is calculated, and the difference is normalized to eliminate the interference of individual differences of patients on the density change amplitude. The normalization process uses the maximum density change amplitude of each patient in the whole cycle as a standard reference to map all density differences to the [-1, 1] interval, thereby generating a density change vector sequence. The vector sequence represents the change trend of the structural region over time and is used to evaluate whether the brain tissue evolution has a pattern dominated by a specific pathological mechanism. Based on the sample trajectories with labeled pathological mechanism labels in the historical big data, clustering analysis of the density change vectors is performed. The clustering method uses a morphological similarity matching algorithm based on the density change vector, combined with time series dynamic time warping to calculate the distance measure between samples. If multiple density change sequences are similar in trend morphology, they are clustered into the same evolutionary path cluster, and the cluster is assigned a mechanism label corresponding to it. For example, if a density sequence trend is closest to a known brain contusion mechanism path sample, it is labeled as a "brain contusion" mechanism label.

[0021] The density change sequence that fails to fall into any mechanism label is uniformly integrated and processed with medical background tracing to ensure the integrity of the overall sample system and provide identification basis for multiple injury combination or new pathological mechanism. After completing sample integration, the historical medical records of the corresponding unknown injury patient are called, including the patient's preliminary diagnosis, follow-up records, imaging description, and standardized neurological function, movement disorder and cognitive ability evaluation records during the current diagnosis and treatment cycle.

[0022] The mechanism label is an index generated after artificial medical review and statistical merging, used to define the pathological type difference between sample clustering clusters. In the management system of structural trajectory sample library, the mechanism label is not defined by image features or algorithm classification standard, but by actual clinical medical diagnosis as the classification main basis, ensuring that the trajectory clustering result has interpretability and traceability in subsequent clinical auxiliary diagnosis and mechanism reasoning.

[0023] In the abnormality recognition module, the evolution of abnormal feedback signals in the structural trajectory is recognized.

[0024] The continuous CT image sequence of each unknown injury patient during the complete treatment cycle is obtained. The image acquisition time interval needs to meet the dynamic monitoring requirements during the diagnosis and treatment cycle, and at least contains four and more time nodes. The images are first subjected to uniform registration processing to eliminate image space errors caused by patient posture, scanning angle and other factors at different time points. The registration method adopts a combined processing method of rigid registration and elastic registration to ensure that the brain structure is uniform in position and boundary overlap in different time point images. Subsequently, brain structure region segmentation operation is performed to extract brain parenchyma, ventricle and related structure regions, and the average gray value of each region in each time point image is calculated as the density expression index to form a time sequence density change data set.

[0025] After the density change sequence is constructed, a continuous structural change path is established according to the time sequence. The definition of the path is: taking the image time node as the horizontal axis and the average density value of the brain structure region as the vertical axis, connecting the time sequence density change trajectory formed. The structural change path, as an important indicator of expressing the evolution trend of the lesion, needs to have integrity and continuity, and any image quality substandard node in the construction process will be excluded to ensure that the sequence truly reflects the structural change process.

[0026] The structure change path is compared with all trajectory samples in the structure evolution trajectory sample library point by point. The point-by-point comparison does not refer to pixel-by-pixel comparison at the image pixel level, nor does it refer to absolute system timestamp time comparison, but refers to matching and comparing the structure density values of each time node under the unified diagnosis and treatment cycle alignment benchmark. Due to individual differences in CT image acquisition time of different patients, in order to ensure the time consistency of the comparison, the structure evolution path in all trajectory sample libraries is time normalized, and a normalized time coordinate system is constructed based on the diagnosis and treatment cycle proportional time axis. Each structure trajectory is mapped to the closed interval of 0-1 according to its time node, and the unknown patient structure path also performs the same normalization processing, so as to ensure that the point-by-point comparison is completed under the unified diagnosis and treatment cycle reference system. Specifically, the regular follow-up cycle of brain injury patients (such as 0 days, 3 days, 7 days, 14 days, 30 days defined in the traumatic brain injury clinical pathway) can be used as the basic time reference point. In the sample collection process, the time point at which the patient first appears in the CT image of the brain and the density reaches the minimum contrast requirement for lesion recognition is taken as the starting point of the normalized time axis, and the subsequent image time points are proportionally mapped according to their positions in the standard cycle. For example, if a patient's CT image obtained on the second day of the first visit shows a recognizable low-density area of the brain parenchyma, this time point is set as the normalized starting point; if the patient completes subsequent CT examination on the 5th day, 10th day and 16th day, these time points correspond to normalized coordinates 0.25, 0.5 and 0.8 respectively, reflecting the relative time progress in the 30-day observation cycle.

[0027] In the comparison process, for the brain structure density value corresponding to each time node, three key indicators are calculated respectively: density change gradient (i.e. the density change rate between adjacent two points), boundary migration amplitude (i.e. the spatial distance of the change of the edge position of the structure region), and center symmetry deviation (i.e. whether the density distribution of the brain structure deviates from the midline). The change trend of the above three indicators is calculated in the time dimension, and a comparison matrix with each path in the sample library is established. Each comparison unit calculates the similarity score of the node in the three indicators, and then merges it into one item in the overall score matrix. The score range is set to 0 to 1, and a full score of 1 represents complete consistency. In the score matrix analysis, the similarity threshold is set to 0.6. If the overall score result at any time node is lower than this threshold, and at the same time accompanied by any of the following conditions, it is considered as an abnormal feedback event triggering condition: (1) the density change gradient direction is reversed at the node, that is, it changes from density decline to rapid rise, or from stable state to sudden and severe decline. Such changes are often seen in the interruption of brain absorption process in the later stage of cerebral hemorrhage, or re-hemorrhage after surgery; (2) the structure boundary symmetry index mutates, that is, the structure symmetry index decreases by more than 40% between two consecutive nodes. Such phenomenon often reflects the brain midline shift caused by structure displacement or edema expansion. If any of the above abnormalities simultaneously appears below the score threshold, an evolution abnormal feedback signal is immediately output. This signal will be called by the subsequent module as an important basis for judging the conflict between structure abnormal evolution and functional state.

[0028] In the conflict triggering module, it is detected whether there is a pattern conflict between the structure change direction and the function evaluation.

[0029] The detection time range is set to be consistent with the entire length of the patient's visit period, covering all assessment periods from the initial admission to the current state before the patient's brain function changes. The extracted brain function assessment scores should include three main indicators: standardized neurological function scores, motor disorder scores, and cognitive ability scores. In the construction of standardized neurological function scores, first identify the assessment items containing electroencephalogram analysis, evoked potential assessment, reflex examination, and consciousness level score from the patient's medical records. Convert the scores into quantitative values according to the international scoring system (such as NIHSS or modified Rankin scale). The extraction of motor disorder scores should focus on gait analysis, muscle tension assessment, and limb coordination test records. Select data sources that are recorded multiple times and have continuity. These indicators should be converted into corresponding scores according to the established reference standard, such as converting muscle strength scores (MRC scoring method) to a 0-20 score interval. The construction of cognitive ability scores should be based on the patient's language comprehension, attention, short-term memory, and executive function dimensions. Extract the results of standard cognitive assessment tables such as MoCA and MMSE. If there are multiple versions, convert them to a common scale, set the standard total score to 30 points, and perform linear mapping to convert to a unified score scale. After completing the extraction and quantification of the above three types of sub-scores, a multi-dimensional comprehensive scoring system is constructed. The comprehensive score structure uses a weighted superposition method, with a weight ratio of 50% for neurological function scores, 30% for motor disorder scores, and 20% for cognitive ability scores. The weight setting is based on the degree of dependence of brain function on neurophysiological integrity, and also refers to the order of brain function disorders in previous clinical research. The comprehensive score is 100 points, and the lower the score, the more severe the brain function impairment.

[0030] After detecting the evolution anomaly feedback signal, the established structural change path in the anomaly recognition module is extracted, which is a continuous curve of brain tissue density value change in CT images over time. In a specific implementation, the average gray value of the corresponding region is extracted for each image node in the time series to form a time series density vector. This density vector is based on structural changes and reflects the direction of tissue morphology evolution. Then, the first order difference of the density time series is calculated to obtain the density change direction vector, where a positive value indicates an increase in tissue density and a negative value indicates a decrease in density. Correspondingly, the time series data of the historical brain function state score is extracted, and the same method is used to calculate the score slope direction to form a function change direction vector. In the present application, the prerequisite for defining directional conflict is that the trends of the two directions are opposite, that is, when the structural density continuously decreases but the brain function score does not decrease but increases or remains stable, or when the structural density significantly increases but the function score shows a downward trend, it is determined that there is a directional conflict. Taking a real clinical situation as an example, if the density in the frontal lobe region of a patient continuously decreases, but the neurological function score increases from 15 to 20, it indicates that structural degradation does not lead to functional loss, and there may be mechanism mismatch; for example, if an abnormal density increase trend is found in the basal ganglia region, and the motor function score decreases from 90 to 60, it indicates that the abnormal density increase does not bring about functional recovery, and may be a pathological density increase such as edema or hemorrhage, which also constitutes a conflict.

[0031] After completing the above directional judgment, all structural-function directional conflict events are counted according to the time nodes to form a conflict event time sequence. Mode conflict is defined as the situation where the frequency of conflict events in the treatment period exceeds a certain set threshold. To ensure the clinical credibility of the judgment, the number of times is set to refer to the medical clinical standard period, and a 28-day observation window is set. If the number of conflicts in the period is ≥3 times (adjustable to 2-5 times to adapt to different diagnosis and treatment scenarios), the mode conflict flag is triggered. In the conflict judgment, each event must meet two conditions: first, the structural density change direction is significant (the change rate exceeds 5%), and second, the function score slope direction has statistical significance (such as the score change amplitude exceeding 2 times the evaluation standard error). All judgments are based on the cross analysis of the structural change path and the function score time sequence to avoid misjudging the conflict state due to only single score fluctuation or image artifact. When the directional deviation occurs at three consecutive time nodes, or there are multiple discontinuous but cumulative conflict nodes that reach the set threshold, it is determined that the mode conflict is established.

[0032] In the reasoning reconstruction module, the reconstruction is based on a multi-mechanism cross reasoning chain.

[0033] After the mode conflict is formally confirmed, the single mechanism path is no longer interpreted as the dominant structure of the injury mechanism. The inference chain reconstruction process based on cross-mechanism is started. The first step is to retrieve several trajectory samples from the structural evolution trajectory sample library that are most similar to the current patient's structural change path. The retrieval method is based on the overall similarity score of the structural change path. In the comparison process, the structural density trend, boundary migration trend, and symmetry preservation are used as joint similarity evaluation indicators. The final score is a weighted combination of the three, with the structural density change trend weight set to 0.5, the boundary migration amplitude weight to 0.3, and the symmetry deviation weight to 0.2. After scoring and sorting, the top five trajectory samples with the highest similarity scores are selected as candidate trajectories. Here, the trajectory sample refers to the standardized structural change path data in the structural evolution path sample library, which has been normalized by time and organized by mechanism label. Each trajectory sample is accompanied by a corresponding mechanism label, which is established based on the medical attribution of clinical diagnosis records, clearly reflecting the potential injury mechanism type carried by the trajectory sample, such as "brain hemorrhage edema absorption period mechanism", "ischemic brain injury functional compensation mechanism", "early evolution of vascular dementia mechanism", etc. These trajectory samples are classified according to their mechanism labels and independently form mechanism candidate paths.

[0034] After obtaining the different mechanism candidate path sets, structural time period division operations are performed on each path. The core basis for division is the key turning points in the density change trend within the functional state change zone. Taking a specific implementation as an example, first extract the brain region segment corresponding to the patient's functional state change zone in each mechanism candidate path, and analyze the density change curve of the region over time. In this change curve, identify the inflection points of density change, i.e., the critical points where the direction of the density curve changes from decreasing to increasing or from increasing to decreasing. Each inflection point serves as a time period division boundary. In clinical settings, common mechanism trajectory stages include: acute stress period (such as rapid density decrease), subacute transition period (density decrease slows down or oscillates), compensation adaptation period (density rises or tends to be stable), etc. For example, in a hemorrhage absorption mechanism trajectory sample, the density change trend may show rapid density increase (hematoma formation) in the 0-0.3 diagnosis and treatment cycle, gradual density decrease (absorption stage) in the 0.3-0.6 cycle, and stable density in the 0.6-1.0 cycle (absorption completion period). Based on this, three mechanism trajectory stages can be divided. All time stage segmentation is completed on a unified diagnosis and treatment cycle reference system corresponding to the time axis, ensuring that the stage division results are comparable across patients. After stage division, each mechanism path is assigned multiple stage sub-paths, providing basic units for subsequent stage-based structure matching and dynamic linking operations.

[0035] After the phase division of the mechanism trajectory, a remapping comparison process of the structural change path and the mechanism trajectory phase is performed to determine which phase of which mechanism may dominate the explanation of the conflict point. First, the occurrence time of the pattern conflict is projected into the corresponding time axis of the diagnosis and treatment cycle reference system, and the normalized phase position of the time point is determined. It is worth noting that the "time" here no longer refers to the actual calendar time, but to the time positioning based on the CT image time normalization and the mechanism trajectory phase division result in the diagnosis and treatment cycle. For example, with the normalized cycle [0, 1] as the axis, if the conflict event occurs at the 0.65 position point, it should fall in the "compensation adaptation period" phase marked in the candidate path. Then, at this normalized time position, the data segment of the corresponding time period in the patient's CT image structural change path is extracted to form a sub-path of the actual structural evolution. Compare this structural change sub-path with the corresponding phase path in all mechanism trajectories one by one, score using structural trend consistency and density inflection point overlap, select the mechanism trajectory phase with the highest score, and take this mechanism phase as the link node in the candidate mechanism. The linking operation includes replacing the discarded segment in the original structural evolution path with the new mechanism phase path, and recording the mechanism source and time period on which the link depends to ensure the traceability of the reasoning chain. This process is time-advanced, centered on the conflict event, and recursively extended forward and backward, and in the subsequent stage, it continues to find matching mechanism trajectories to build a combined path dominated by multiple mechanisms. The final constructed cross-mechanism reasoning chain will contain trajectory phases with different mechanism labels and cover the CT image evolution path in the entire treatment cycle, forming a structural reasoning chain with high explainability and composite mechanism support.

[0036] In addition, the engineering semantics of the term "mechanism" is not an abstract behavior model or algorithm control path, but a medical pathology essence bound to each structural evolution trajectory sample. Specifically, each mechanism label represents a disease evolution pattern or symptom process recognized by the medical community and has a clear pathological logic chain in the actual diagnosis and treatment process, such as "hematoma absorption mechanism after brain contusion and laceration", "cerebral ischemia-reperfusion mechanism after cerebral infarction", "self-limiting absorption mechanism of chronic subdural hematoma", or "secondary hydrocephalus mechanism after cerebral hemorrhage". The naming of the above mechanisms is based on clinical trauma types and combined with the evolution trend of anatomical structure changes and the process of symptom group composition.

[0037] The model training module trains an injury mechanism recognition model based on the cross-mechanism reasoning chain and the brain function state.

[0038] The cross-mechanism reasoning chain of all unknown injury patients is standardized and fragmented. Each cross-mechanism reasoning chain usually contains multiple mechanism trajectory stages, and each mechanism trajectory stage corresponds to a set of continuous structural change nodes. In the fragmentation process, the time stage division criterion adopted in the reasoning reconstruction module is used to ensure consistency between the two modules. The time stage is divided according to the normalized coordinate axis of the treatment cycle, and the time point at which the structural density mean in the first CT image reaches the set threshold is used as the normalized starting point. The subsequent time period is based on the fluctuation period of brain function score and the structural density inflection point change in the mechanism trajectory, and the representative time period intervals such as acute phase, subacute phase, pre-recovery phase, and recovery phase are divided.

[0039] Each mechanism fragment is taken as a node in the graph structure network, and the connection mode of the edge in the graph is set according to the causal connection relationship in the original reasoning chain. Each node is attached with a mechanism label to identify its pathological mechanism attribution. To enhance the coherence of the graph structure, the system further fuses the shared nodes appearing in multiple mechanism trajectory stages of the same patient, retains the highest frequency mechanism label identification, and records the source mechanism trajectory stage information for subsequent backtracking. Finally, a complete cross-patient mechanism graph structure reasoning network is constructed in the mechanism fragments of all patients, ensuring that the connection between the nodes in the network has directionality and time sequence relationship, which constitutes the basic constraint condition of the edges in the graph.

[0040] The brain function state evaluation score sequence of all unknown injury patients during the treatment cycle is obtained. The score sequence is mapped synchronously according to the aforementioned time stage, and is labeled to the corresponding structural change node in the graph structure reasoning network, forming the coupling relationship between the node and the score. When constructing the guide path system, the function state score of each patient at each stage is mapped to the structural change node in the graph structure at the corresponding time stage, and a weight is assigned to each connection edge. The edge weight factor is defined as the slope value of the score change between the two structural change nodes connected by the edge. The rising slope of the score represents the improvement trend of the function, and the falling score represents the deterioration trend. The size of the edge weight quantifies the guiding strength of each path to the function evolution direction.

[0041] Based on the constructed graph structure reasoning network and the edge weight system mapped with the guide path, a training process of a graph structure path recognition model is performed. The model adopts a time sequence graph-based attention mechanism to infer the structure, that is, under the premise that the input sequence is a time stage directed graph, the importance and passing probability of each path in the graph are dynamically evaluated using a multi-head attention mechanism. The training target is to identify the mechanism combination sequence with the highest path passing probability in the graph under the condition that the functional score evolution trend best matches. The training process uses a standard cross-entropy loss function, and combines node access frequency and score change difference as optimization factors to improve the model's path recognition ability in complex mechanism combination. The path recognition model finally trained will be called as a basic model in subsequent actual patient mechanism recognition reasoning.

[0042] The decision output module assists in decision-making for the actual patient's pathological mechanism based on the damage mechanism recognition model output result.

[0043] The continuous CT images of the actual patient in the current treatment cycle are centrally processed to ensure that all input images meet the pre-requisites of spatial registration and image quality standards. The image registration process performs three-dimensional repositioning of the brain structure in the continuous image frames to remove non-structural changes caused by factors such as patient position, angle deviation, etc. After registration, the brain key areas are extracted using a method based on brain region template division, and the brain parenchymal region, ventricle region, and marginal edema zone are obtained, and the density value of the corresponding region in each image is calculated. The density value is represented by the average gray intensity of the unit voxel in the CT image, reflecting the radioactivity attenuation characteristics of the tissue. According to the time axis order of the image sequence, the sequence of brain density changes over time is constructed to form a complete density change sequence data structure as the basis for expressing structural changes.

[0044] After establishing the density change sequence, it needs to be mapped to the previously constructed graph structure reasoning network. The nodes of the reasoning network are composed of a "mechanism label-time stage" binary tuple, and each node represents the structural evolution performance of a specific damage mechanism at a certain time stage in the treatment cycle. Therefore, the input of the density change sequence needs to be structured and converted to form a node input feature format that the graph structure model can recognize. The specific operation includes: segmenting the density change sequence according to the same time stages as in the reasoning network, such as acute stage, subacute stage, pre-recovery stage, and post-recovery stage; within each stage, extract statistical features of the density change sequence in that time period, such as mean change rate, gradient trend, symmetry deviation index, etc., to construct a structural change vector for that time stage. Each structural change vector serves as a feature vector for the input node, corresponding to the node features in the graph structure.

[0045] The call graph structure path recognition model is used to evaluate the path passing probability of the current input node sequence of the patient. Since the model introduces an edge weight factor in the training process, the edge weight of the path essentially represents the trend of the brain function state score between adjacent nodes in the historical sample. Therefore, in the actual reasoning stage, to avoid the bias caused by directly applying the historical score results, the brain function state evaluation score sequence of the current patient needs to be introduced for comparison. By calculating the matching degree of the current patient's score trend and the historical score trend in the edge weight (such as the consistency of the slope direction and the numerical similarity), the edge weight factor is dynamically adjusted to form a patient-specific edge weight mapping. The model then performs temporal graph reasoning calculation based on the adjusted edge weight for all possible paths, dynamically evaluates the path passing probability through the attention mechanism, and outputs the passing scores of all paths. The reasoning path with the highest passing probability is identified, and all the mechanism label information passed through the path is extracted. Since the mechanism label of each node in the graph structure has been labeled during construction, the frequency distribution and combination structure of the mechanisms covered by the path can be counted. The mechanism label with the highest frequency or dominant in the path is taken as the output result of the structural damage mechanism recognition, which supports more explanatory mechanism judgment on the pathological state of the current patient and provides mechanism-level basis for the next treatment plan and function prediction. The whole process ensures the collaborative evolution logic between the input structure information, the function state change and the mechanism label, and provides high-reliability auxiliary decision for multi-mechanism structure recognition under complex damage conditions.

[0046] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large number of data to obtain a formula of the latest real situation. The preset parameters and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0047] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0048] Those of ordinary skill in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0049] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device, and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0050] In several embodiments provided in the present application, it should be understood that the disclosed system, device, and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed ones can be indirect coupling or communication connection through some interfaces, devices, or modules, which can be electrical, mechanical, or other forms.

[0051] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0052] In addition, the functional modules in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0053] The functions, if realized in the form of software function modules and sold or used as independent products, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products, and the computer software products are stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various program code storage media.

[0054] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0055] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A decision support system for the diagnosis and treatment of traumatic brain injury based on medical big data, characterized in that, include: The trajectory construction module is used to acquire CT image sequences of craniocerebral injury within a complete diagnosis and treatment cycle in medical big data, extract brain region density from the images, and construct a sample library of structural evolution trajectory of brain regions based on time series. The anomaly recognition module is used to acquire the patient's CT images to construct the structural change path, compare it with the structural evolution trajectory sample library, and identify the evolutionary anomaly feedback signal in the structural trajectory. The conflict triggering module extracts historical brain function data of patients based on structural abnormality feedback signals and detects whether there is a pattern conflict between the direction of structural changes and functional assessment. The reasoning reconstruction module discards single-mechanism evolution path assumptions and reconstructs reasoning chains based on multi-mechanism intersections when pattern conflicts exist. The model training module trains a damage mechanism identification model based on cross-mechanism inference chains and brain functional states. The decision output module, based on the output results of the damage mechanism identification model, assists in decision-making regarding the pathological mechanisms of actual patients.

2. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 1, characterized in that, The trajectory construction module extracts brain region density from the image and constructs a sample library of brain region structural evolution trajectories based on time series data. The CT image sequences of traumatic brain injury were sorted in chronological order, and continuous sample groups with image quality meeting the imaging standards were selected. For each image, a registration process is performed to extract the density value of the brain region. The density value is the gray value corresponding to the brain tissue region in the CT image. Density change sequences are constructed based on density value changes at adjacent time points, and clustered into evolutionary path clusters according to the same mechanism label, forming a structural trajectory sample library with time and mechanism labels; For density change sequences that failed to cluster into any mechanism label, sample integration was performed to obtain the historical medical records of patients with corresponding unknown injuries. The mechanism label is a classification index identifier for evolutionary path clusters based on medical diagnostic results.

3. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 1, characterized in that, The anomaly identification module identifies the process of feedback signals for evolving structural trajectories as follows: We acquire continuous CT image sequences of patients with unknown injuries during their medical visits, extract brain region density change sequences through image spatial registration, and construct continuous structural change paths in chronological order. The structural change path is compared point by point with all trajectory samples in the structural evolution trajectory sample library, and a similarity scoring matrix is ​​constructed using three types of indicators: density change gradient, boundary migration amplitude, and central symmetry offset. In the scoring matrix, when the score of any comparison point is lower than the preset similarity threshold, and the direction of the density change gradient is reversed or a symmetric mutation occurs, an evolutionary anomaly feedback signal is output.

4. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 3, characterized in that, The conflict triggering module specifically includes detecting whether there is a pattern conflict between the direction of structural change and the functional evaluation: Historical brain function status assessment scores were extracted from patients with unknown injuries. These assessment scores were comprehensive scores that included standardized neurological function, motor impairment, and cognitive ability. After obtaining the evolutionary anomaly feedback signal, the structural change path established in the anomaly identification module is extracted and converted into a density change direction vector expression, which is then compared with the slope direction of the time curve of the historical brain function state assessment score. When the direction of density change is inconsistent with the expected direction of brain function change, it is marked as a conflict. If the number of times a conflict occurs within the treatment period exceeds the set number of times, it is considered that there is a pattern conflict.

5. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 1, characterized in that, In the inference reconstruction module, reconstructing the inference chain based on the intersection of multiple mechanisms specifically includes: After identifying pattern conflicts, the structural evolution hypothesis based on a single mechanism evolution trajectory is abandoned. Several trajectory samples with the highest similarity to CT image sequences of patients with unknown injuries are extracted from the structural evolution trajectory sample library. Candidate paths corresponding to different mechanisms are constructed according to the mechanism labels. In the candidate path, the trajectory samples are segmented into time stages based on the inflection point of the density change direction in the functional state change area to generate mechanism trajectory stages divided by time. Based on the occurrence time of the pattern conflict, segment-level remapping is performed on the trajectory stages in each candidate path and the structural change paths in the patient's CT image, and dynamic linking is performed on the mechanism trajectory stages with the highest mapping similarity. The above linking process is executed recursively, and a cross-mechanism inference chain is constructed based on the possible structural combinations of different mechanisms.

6. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 5, characterized in that, The mechanism is the pathological mechanism, pathological change process, or medically defined symptom pattern corresponding to a specific type of medical trauma, as indicated by the mechanism label.

7. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 1, characterized in that, The model training module, based on the cross-mechanism inference chain and brain functional state, specifically trains the injury mechanism identification model, including: The cross-mechanism inference chains corresponding to all patients with unknown injuries are broken down into segments, and a graph-structured inference network is constructed according to mechanism labels and time stages. Obtain the brain function status assessment score sequence of all patients with unknown damage during their medical treatment cycle, map the score sequence to the structural change nodes in the inference network according to time stage, and construct a guiding path system with brain function status assessment score as the edge weight factor. The training graph structure path recognition model uses a temporal graph-based attention mechanism to calculate the path probability.

8. The traumatic brain injury diagnosis and treatment auxiliary decision-making system based on medical big data according to claim 1, characterized in that, The decision output module, based on the output results of the damage mechanism identification model, provides auxiliary decision-making for the actual patient's pathological mechanism, specifically including: The continuous CT images of the actual patient during the current medical cycle are input into the image processing flow to extract the density change sequence of the brain region; According to the node setting of the graph structure inference network, the density change sequence is mapped to the corresponding structure change node in the graph, and path passage inference is performed based on the edge weight factor on the node connection edge. The mechanism labels associated with the inference path with the highest probability of occurrence are statistically output as the results of structural damage mechanism identification.

Images