Intelligent prioritization system for radiology emergency imaging
By combining multimodal data acquisition, evidence-based deep learning, and reinforcement learning, the uncertainty of the model is quantified, which solves the problem of misjudgment in intelligent sorting systems under the interference of rare lesions or artifacts, and realizes safe and efficient sorting and resource optimization of emergency images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANAN UNIV AFFILIATED HOSPITAL
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-19
AI Technical Summary
Existing intelligent sorting technologies are prone to high-confidence misjudgments when faced with rare lesions or artifact interference, leading to sorting chaos and medical risks. Furthermore, traditional models lack the ability to quantify their own cognitive uncertainty.
A multimodal data acquisition module is used to acquire medical images and electronic application form texts. Features are extracted and uncertainty is quantified through evidence deep learning network. Combined with clinical semantic features and environmental status, a reinforcement learning dynamic scheduling module is used for sorting. Uncertainty perception and visualization mechanisms are introduced to prevent misjudgment.
It effectively identifies out-of-distribution samples that are easily misjudged by traditional models, prevents incorrect sorting, improves the safety and robustness of the emergency triage system, dynamically balances the allocation of medical resources, and improves the accuracy of patient criticality assessment and the clinical interpretability of sorting decisions.
Smart Images
Figure CN122245673A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to an intelligent sorting system for prioritizing radiological emergency images. Background Technology
[0002] The radiology emergency department is a crucial link in hospital life-saving efforts; the speed of imaging diagnosis often determines the survival rate of patients with acute and critical illnesses such as cerebral hemorrhage, aortic dissection, and severe trauma. Faced with the contradiction between the ever-increasing demand for imaging examinations and limited medical resources, the traditional first-come, first-served queuing model is no longer sufficient to meet the demands of clinical timeliness. In recent years, artificial intelligence technology has been widely applied in the field of medical image analysis. Deep learning algorithms automatically identify critical signs in images and assist in triage, becoming a core means to improve the efficiency of emergency department operations and ensure patient safety. Various auxiliary diagnostic systems are gradually transforming from single lesion detection to end-to-end workflow optimization, aiming to achieve precise allocation of medical resources through intelligent means.
[0003] However, existing intelligent ranking technologies mainly rely on deterministic deep learning models, which lack the ability to quantify the uncertainty of the model's own cognition. When dealing with rare diseases or artifact interference, they are prone to high-confidence misjudgments, leading to ranking chaos and medical risks. Summary of the Invention
[0004] To overcome the above shortcomings, this invention provides an intelligent sorting system for prioritizing radiological emergency images, aiming to improve the problem that existing technologies are prone to high-confidence misjudgments when faced with rare lesions or artifact interference.
[0005] This invention provides the following technical solution: an intelligent sorting system for prioritizing emergency radiology images, comprising:
[0006] The multimodal data acquisition module is used to acquire medical images of patients to be diagnosed and related electronic application form texts;
[0007] The uncertainty-aware reasoning module is used to extract features from the medical images through an evidence deep learning network and generate a predicted probability representing the lesion category and a score representing the uncertainty of the model using a parameterized Dirichlet distribution.
[0008] The clinical semantic feature extraction module is used to perform semantic encoding on the electronic application form text and output a clinical feature vector that represents the severity of the disease information contained in the text.
[0009] The environment state construction module is used to aggregate the current queue length, the historical average reading time of the doctor currently performing the reading task, and the cumulative patient stay time to construct an environment state vector that represents the load on reading resources.
[0010] The reinforcement learning dynamic scheduling module is used to take the uncertainty score, clinical feature vector and environmental state vector as input states, and use the policy network trained based on the reward function including the waiting time penalty term to perform mapping calculation and output the target insertion position index.
[0011] The queue execution and visualization module is used to reconstruct the order of the reading list based on the target insertion position index, and to highlight entries with uncertainty scores higher than a preset threshold.
[0012] Preferably, in the multimodal data acquisition module, the step of acquiring the medical images of the patient to be diagnosed and the associated electronic application form text includes:
[0013] Configure a data monitoring interface to monitor data flow changes in the image archiving system in real time. By parsing the unique identifier in the image header file, establish an index association between image data and patient electronic requisition forms in the radiology information system.
[0014] The text parsing algorithm is called to extract the unstructured clinical description text from the electronic application form, and the sensitive fields are de-identified in combination with the preset privacy dictionary to meet data compliance requirements;
[0015] The heterogeneous medical image and text data are time-aligned and encapsulated using a data buffer queue to construct a multimodal data packet containing complete diagnostic information, which is then transmitted to the inference memory pool for processing.
[0016] Preferably, in the uncertainty-aware reasoning module, the step of extracting features from the medical image using an evidence deep learning network includes:
[0017] Using a pre-built deep convolutional neural network backbone model, multi-scale convolution and downsampling operations are performed on the input medical images to extract high-dimensional feature maps containing lesion morphology, texture and spatial location;
[0018] A spatial attention mechanism is introduced to calculate the saliency weight of each pixel in the feature map, suppress background noise areas, and enhance the feature response value of suspected lesion areas to improve the signal-to-noise ratio of the features.
[0019] The weighted high-dimensional feature map is compressed into a one-dimensional feature vector by global pooling, and the feature vector is mapped to a non-negative evidence space by a fully connected layer to generate an evidence output vector corresponding to each lesion category.
[0020] Preferably, in the uncertainty-aware reasoning module, the step of generating the predicted probability representing the lesion category and the score representing the uncertainty of the model using the parameterized Dirichlet distribution includes:
[0021] The evidence output vector is transformed using a nonlinear activation function to ensure that the generated pseudo-count parameters satisfy the positive definiteness constraint of the Dirichlet distribution, thereby constructing a parameter set describing the multi-class probability distribution.
[0022] Based on evidence theory, the ratio of pseudo-count parameters of each category to the total evidence strength is calculated to obtain the predicted probability of each lesion category, quantifying the possibility that the current image belongs to a specific critical disease.
[0023] The reciprocal of the total number of categories and the total strength of evidence is calculated to quantify the cognitive uncertainty caused by the lack of training data distribution or the ambiguity of sample features, resulting in a normalized uncertainty score.
[0024] Preferably, in the clinical semantic feature extraction module, the step of semantically encoding the electronic application text includes:
[0025] We used a pre-trained natural language processing model to segment and vectorize the electronic application text, capturing the contextual semantic dependencies of clinical terms in the text.
[0026] Perform entity recognition and negation detection, extract key symptom entities and identify their affirmative or negative states, and reverse the feature polarity described as exclusionary symptoms to correct semantic expression;
[0027] The extracted symptom feature vectors are fused and encoded with emergency triage labels to output clinical feature vectors that can quantitatively reflect the urgency of the patient's current chief complaint and the stability of vital signs.
[0028] Preferably, in the environment state construction module, the step of constructing an environment state vector representing the load on viewing resources includes:
[0029] Set a sliding time window to statistically analyze the average reading time and processing variance of doctors currently performing image reading tasks within the window period, and quantify the current work efficiency and fatigue fluctuation characteristics of doctors.
[0030] Traverse the current image queue, count the number of images backlogged in the queue and the distribution of their severity levels, and generate load characteristics that reflect the complexity and urgency of the current task to be processed.
[0031] The waiting time of all patients in the queue is normalized, and the work efficiency feature, load feature and waiting time feature are concatenated to construct an environment state vector that is dynamically updated over time.
[0032] Preferably, in the reinforcement learning dynamic scheduling module, the step of performing mapping calculation using a policy network trained based on a reward function that includes a waiting time penalty term includes:
[0033] Define a composite reward function that includes a positive reward term proportional to the accuracy of prioritizing critical patients and a negative penalty term that increases non-linearly with the length of stay of ordinary patients.
[0034] The policy gradient algorithm or value iteration algorithm is used as the training framework. The historical state transition samples are stored using the experience replay mechanism. The weight parameters of the policy network are updated by maximizing the cumulative discounted reward.
[0035] The current environment state vector is input into a pre-trained policy network, and the value assessment or probability distribution of taking different insertion actions in the current state is calculated through forward propagation.
[0036] Preferably, in the reinforcement learning dynamic scheduling module, the step of inserting the output target position index includes:
[0037] A mask vector is generated based on the soft constraint of the first-in-first-out principle to block out illegal insertion positions that violate the first-in-first-out principle and reduce the action search space.
[0038] The values of the legal actions output by the policy network are sorted or sampled, and the position with the highest expected return is selected as the optimal decision, generating the corresponding target insertion position index.
[0039] Based on the target insertion position index, the estimated waiting time parameters of all patients after that position in the queue are dynamically adjusted to complete the closed loop of a single scheduling decision.
[0040] Preferably, in the queue execution and visualization module, the step of highlighting entries with uncertainty scores higher than a preset threshold includes:
[0041] The uncertainty score calculated in real time is compared with the preset anomaly detection threshold to identify out-of-distribution samples or high-noise samples that the model cannot be certain of.
[0042] Visual warning labels are overlaid on the identified abnormal samples in the image review work list, the automatic sorting function of the entry is locked, and a prompt signal requiring manual review is generated to prevent model misjudgment.
[0043] An interpretability algorithm is used to generate an uncertainty heatmap, which is then overlaid and displayed when doctors review images, providing a visual indication of image regions that cause high uncertainty in the model.
[0044] The present invention has the following beneficial effects:
[0045] 1. In this invention, by introducing an uncertainty perception and reasoning module based on evidence-based deep learning, the uncertainty of the model’s cognition of rare lesions or artifacts is quantified while outputting the lesion classification probability. This effectively identifies out-of-distribution samples that are easily misjudged by traditional models and triggers a manual review mechanism, avoiding the risk of incorrect sorting caused by the model’s blind confidence, and significantly improving the safety and robustness of the emergency triage system.
[0046] 2. In this invention, the sorting problem is modeled as a global sequence optimization decision by using a reinforcement learning dynamic scheduling module and a reward function that includes a waiting time penalty term. This breaks through the limitations of traditional greedy algorithms, ensuring priority treatment for critically ill patients while dynamically balancing the waiting rights of ordinary patients. This effectively prevents the backlog of low-priority patients and achieves the global optimal allocation of medical resources under variable load environments.
[0047] 3. In this invention, the clinical semantic feature extraction module deeply integrates electronic application form text and medical image features, uses a negation detection mechanism to accurately correct semantic misunderstandings of simple keyword matching, and combines real-time environmental features of doctor efficiency and queue load to significantly improve the accuracy of patient criticality assessment and the clinical interpretability of ranking decisions. Attached Figure Description
[0048] Figure 1 This is an architecture diagram of the intelligent sorting system for prioritizing radiological emergency images proposed in this invention. Detailed Implementation
[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0050] In embodiments of the present invention, the present invention provides an intelligent sorting system for prioritizing radiological emergency images, such as... Figure 1 As shown, it includes:
[0051] The multimodal data acquisition module is used to acquire medical images of patients to be diagnosed and related electronic application form texts;
[0052] Furthermore, in the multimodal data acquisition module, the steps for acquiring the medical images of the patient to be diagnosed and the associated electronic application form text include: configuring a data monitoring interface to monitor the data flow changes of the image archiving system in real time; establishing an index association between the image data and the patient's electronic application form in the radiology information system by parsing the unique identifier in the image header file; calling a text parsing algorithm to extract the unstructured clinical description text in the electronic application form, and performing desensitization processing on sensitive fields in conjunction with a preset privacy dictionary to meet data compliance requirements; using a data buffer queue to perform time-series alignment and encapsulation of heterogeneous medical image and text data, constructing a multimodal data packet containing complete diagnostic information, and transmitting it to the inference memory pool for processing.
[0053] Specifically, the multimodal data acquisition module is deployed on a server interconnected with the hospital's image archiving and communication system and radiology information system. This module first initializes a DICMC-STORE server daemon, which listens on a preset TCP port to capture medical image data streams broadcast by the image archiving system in real time. When a new image instance transmission request is detected, the system calls the DICOM protocol parser to read the image file header information, extracts the AccessionNumber (with tag group (0008, 0050)) as a unique identifier, and uses this AccessionNumber as an index key to send an SQL query command or HL7 query message to the radiology information system's database to retrieve and download the corresponding electronic application form text data for that patient. The system uses a regular expression engine to locate the unstructured clinical description text field in the electronic application form, which typically contains the chief complaint, present illness, and past medical history. To meet privacy compliance requirements, the module loads a pre-built privacy dictionary containing common surnames, administrative division names, and ID card number verification rules.
[0054] For the anonymization of text data, the system executes a dictionary-based filtering algorithm. Let the original clinical text sequence in the electronic application form be... ,in Indicates the first in the text Each word segmentation unit, the pre-defined privacy dictionary set is defined as follows: The system uses the following mapping function. Perform word-by-word scanning and replacement on the text:
[0055] ;
[0056] In the formula, The input is the current word to be detected. Preset desensitization placeholders, such as asterisks or generic pronouns, This involves creating a sensitive word feature library containing patient names, ID numbers, and contact information. After the above calculations, the desensitized text sequence is output. This allows for the removal of identity-sensitive information while preserving clinical pathology information.
[0057] After data extraction and anonymization, the module utilizes an in-memory database to construct a first-in-first-out (FIFO) primary data buffer queue. Since the transmission time for medical image data is typically much longer than that for text data, the system establishes hash mapping buckets using the examination number as the key. When data for any modality arrives, the system checks if the hash bucket already contains data for the corresponding modality. If not, the current data is temporarily stored and a timeout timer is started; if it exists, the medical image matrix is... With the desensitized text sequence Execute timing alignment and encapsulate into multimodal data packets. ,in This is the check number. The packet is then pushed to the inference memory pool, and the hash bucket space is released.
[0058] This implementation process enables automated association and standardized cleaning of heterogeneous medical data, ensuring that the data input into downstream models has spatiotemporal consistency and privacy security.
[0059] The uncertainty-aware reasoning module is used to extract features from medical images through an evidence deep learning network and to generate predicted probabilities representing lesion categories and scores representing the uncertainty of the model using a parameterized Dirichlet distribution.
[0060] Furthermore, in the uncertainty-aware reasoning module, the steps for feature extraction from medical images using an evidence deep learning network include: utilizing a pre-built deep convolutional neural network backbone model to perform multi-scale convolution and downsampling operations on the input medical images to extract high-dimensional feature maps containing lesion morphology, texture, and spatial location; introducing a spatial attention mechanism to calculate the saliency weight of each pixel in the feature map, suppressing background noise regions, and enhancing the feature response values of suspected lesion regions to improve the signal-to-noise ratio of the features; compressing the weighted high-dimensional feature map into a one-dimensional feature vector through a global pooling operation, and using a fully connected layer to map the feature vector to a non-negative evidence space to generate evidence output vectors corresponding to each lesion category.
[0061] Furthermore, in the uncertainty perception reasoning module, the steps of generating predicted probabilities representing lesion categories and scores representing model uncertainty using parameterized Dirichlet distribution include: transforming the evidence output vector using a nonlinear activation function to ensure that the generated pseudo-count parameters satisfy the positive definiteness constraint of the Dirichlet distribution, and constructing a parameter set describing the multi-class probability distribution; calculating the ratio of pseudo-count parameters of each category to the total evidence strength based on evidence theory to obtain the predicted probability of each lesion category, quantifying the possibility that the current image belongs to a specific critical disease; calculating the reciprocal of the total number of categories to the total evidence strength to quantify the cognitive uncertainty caused by the model due to missing training data distribution or fuzzy sample features, and obtaining a normalized uncertainty score.
[0062] Specifically, the uncertainty-aware inference module first receives a preprocessed two-dimensional medical image data matrix transmitted by the multimodal data acquisition module. This module loads a pre-trained ResNet50 deep convolutional neural network as the backbone architecture for feature extraction, removes the fully connected layers at the ends of the original ResNet50 network, and retains all convolutional layers and residual block structures from Conv1 to Conv5_x. Input image After forward propagation computation using ResNet50, and through a series of convolutions, batch normalization, and ReLU activation operations, a feature map tensor containing high-dimensional semantic information of the lesions is output. Let the feature map tensor be... The dimension is ,in and These represent the height and width of the feature map, respectively. Representing the number of channels, this feature map encodes the morphological and textural features of potential hemorrhages, fracture lines, or space-occupying lesions in the image.
[0063] To suppress background noise and focus on key lesion areas, the module then processes the feature map tensor. A spatial attention mechanism is introduced. First, the feature maps are processed along the channel axis. Max pooling and average pooling operations are performed to generate two 2D statistical feature maps, which are then concatenated along the channel dimension. The concatenated feature map is then processed by a convolutional kernel of size [size missing]. Spatial features are fused using convolutional layers, and then a spatial attention weight matrix with values ranging from 0 to 1 is generated using a sigmoid activation function. The calculation formula is as follows:
[0064] ;
[0065] In the formula, This represents the Sigmoid activation function. Indicates the kernel size as Convolution operation, and These represent the average pooling and max pooling operations at the channel level, respectively. This indicates a concatenation operation, resulting in the weight matrix. Then, it is compared with the original feature map tensor. Perform element-wise multiplication to obtain the weighted feature map. This enhances the characteristic response values of suspected lesion areas.
[0066] Subsequently, the module processes the weighted feature map. Perform a global average pooling operation, spatial dimensions compressed to The length is obtained as One-dimensional feature vector This feature vector It is fed into a fully connected layer, and the number of output nodes of that layer is... This corresponds to a preset number of emergency lesion categories, such as cerebral hemorrhage, cerebral hemorrhage, cerebral infarction, skull fracture, and normal. To convert the neural network output into non-negative evidence values in evidence theory, the module uses the Softplus activation function on the linear output of the fully connected layer. Perform a nonlinear transformation to calculate the first... Evidence value of lesion-like lesions The calculation formula is as follows:
[0067] ;
[0068] In the formula, For the fully connected layer corresponding to the first The output logical values for each category, The evidence value is non-negative. Based on evidence-based deep learning theory, this module constructs a parameterized Dirichlet distribution to simulate the uncertainty of classification prediction. First, the parameter vector of the Dirichlet distribution is calculated based on the evidence value. , of which Parameters for each category Defined as:
[0069] ;
[0070] Next, the total intensity of the Dirichlet distribution is calculated. This value equals the sum of all category parameters, i.e. Based on this, the module calculates that the current image belongs to the [number]th [image]. Predictive probability of lesions and the model's overall cognitive uncertainty score regarding the prediction result. The calculation formulas are as follows:
[0071] ;
[0072] In the formula, The total number of categories. This represents the total intensity of the Dirichlet distribution. The system ultimately outputs the maximum probability value. The corresponding lesion category is used as the preliminary diagnostic result and output. As an uncertainty score, the total evidence strength is calculated when the input image contains rare lesions or severe artifacts that prevent the network from matching known features. It will approach This leads to an uncertainty score. Approaching 1.
[0073] This implementation step enables the system not only to provide lesion classification results, but also to quantitatively assess the model's confidence level in the results, effectively identify out-of-distribution samples, and prevent AI models from giving erroneous diagnoses with high confidence levels under unknown circumstances.
[0074] The clinical semantic feature extraction module is used to perform semantic encoding on the electronic application form text and output a clinical feature vector that represents the severity of the disease information contained in the text.
[0075] Furthermore, in the clinical semantic feature extraction module, the step of semantically encoding the electronic application form text includes: using a pre-trained natural language processing model to segment and vectorize the electronic application form text to capture the contextual semantic dependencies of clinical terms in the text; performing entity recognition and negation detection to extract key symptom entities and identify their affirmative or negative states, and reversing the feature polarity described as exclusionary symptoms to correct semantic expression; and fusing and encoding the extracted symptom feature vector with emergency triage labels to output a clinical feature vector that can quantitatively reflect the urgency of the patient's current complaint and the stability of vital signs.
[0076] Specifically, the clinical semantic feature extraction module first receives the electronic application form text data sequence from the multimodal data acquisition module. ,in Represents the first in the text Each character or word segmentation unit. This module loads a pre-trained BERT-based deep neural network architecture as the base model for semantic encoding. The BERT-based model contains a 12-layer Transformer encoder with a hidden layer dimension of 768. The input is a text sequence. First, the data is processed by the WordPiece word segmenter, with special classification markers [CLS] and separator markers [SEP] added to the beginning and end, respectively. Then, it is input into the BERT model. After contextual feature aggregation via a multi-head self-attention mechanism, the model outputs the corresponding contextual semantic vector sequence. ,in The calculation process is described as follows:
[0077] ;
[0078] In the formula, Each vector in They all incorporate the semantic dependencies between the words at that position and the context of the entire sentence.
[0079] To perform entity recognition and negation detection, the module connects a bidirectional long short-term memory network layer and a conditional random field layer above the BERT output layer. The BiLSTM layer processes vector sequences. To capture long-range sequence dependency features, the CRF layer is used to predict the biomedical entity label for each character. The system identifies the set of entities belonging to the symptom category. and the set of entities belonging to the category of negation words. For each identified symptom entity The system searches for the existence of a negation word entity within a preset context window. If it exists, the feature polarity of the symptom entity is inverted. Let the symptom entity... The corresponding vector in the BERT output is represented as follows: Its corrected semantic vector The calculation formula is as follows:
[0080] ;
[0081] In the formula, For the indicator function, when the symptom entity Located in the scope of negation words The value is 1 if the chest pain is present, and 0 otherwise. Mathematically, this step projects the chest pain feature vector in the absence of chest pain into the opposite direction of the semantic space, thereby correcting the semantic expression.
[0082] Subsequently, the module extracts the emergency triage labels from the electronic application form, such as Level I critical, Level II severe, Level III emergency, and Level IV non-emergency, and maps them to fixed-dimensional one-hot encoded vectors. To integrate unstructured text features with structured hierarchical features, the module first processes all corrected symptom vectors. Perform max pooling to extract globally salient text feature vectors Then it is compared with the emergency triage vector. The data are concatenated and then passed through a fully connected layer to generate the final clinical feature vector. The calculation formula is as follows:
[0083] ;
[0084] In the formula, and These are the weight matrix and bias term of the fully connected layer, respectively. This represents a vector concatenation operation. This is the activation function. The output is... This refers to a high-dimensional feature vector that can quantitatively reflect the urgency of the patient's current chief complaint and the stability of vital signs.
[0085] This implementation step corrects the semantic ambiguity of simple keyword matching by introducing a negation detection mechanism and integrates structured hierarchical data, effectively solving the clinical semantic understanding problem of misjudging bleeding when there is no bleeding, and improving the accuracy of feature vectors in representing the true severity of the condition.
[0086] The environment state construction module is used to aggregate the current queue length, the historical average reading time of the doctor currently performing the reading task, and the cumulative patient stay time to construct an environment state vector that represents the load on reading resources.
[0087] Furthermore, in the environmental state construction module, the steps for constructing an environmental state vector representing the load of image reading resources include: setting a sliding time window, statistically analyzing the average image reading time and processing variance of the doctor currently performing the image reading task within the window period, and quantifying the current doctor's work efficiency and fatigue fluctuation characteristics; traversing the current image reading queue, statistically analyzing the number of images backlogged in the queue and the distribution of criticality levels, and generating load characteristics reflecting the complexity and urgency of the current pending tasks; normalizing the waiting time of all patients in the queue, and cascading the work efficiency characteristics, load characteristics, and waiting time characteristics to construct an environmental state vector that is dynamically updated over time.
[0088] Specifically, the environment state construction module first connects to the backend database of the radiation information system via an RPC remote procedure call interface, and initializes a sliding time window data structure based on the first-in, first-out principle. The capacity of this window is set to... Used to store the most recently completed tasks by the currently logged-in doctor. Example of log data for film review tasks. The system captures the start timestamp of each historical task in real time. With report review timestamp Calculate the time consumed by a single image reading session. Based on the time-consuming data set within this sliding window. The module calculates the average time spent by doctors reviewing images. and the variance of viewing time These are used to quantify the doctor's baseline work efficiency and current fatigue fluctuation characteristics, respectively. The calculation formula is as follows:
[0089] ;
[0090] In the formula, For the first The time taken for each historical task This represents the size of the sliding window. Larger variance values typically indicate decreased physician attention or increased fatigue.
[0091] Next, the module iterates through the PACS system's list of images awaiting review, and counts in real time the total number of images backlogged in the current queue. Simultaneously, the module reads the emergency triage label for each patient in the queue. This label is generated by the clinical semantic feature extraction module, and it is used to statistically determine the patient's level of criticality. number of patients ,in These correspond to critically ill, severely ill, acute, and non-acute cases, respectively. The system calculates the distribution vector of the proportion of patients in each category. ,in This vector reflects the overall urgency structure of the tasks currently pending.
[0092] Based on the cumulative patient stay time, the module obtains the registration and check-in timestamp of each patient in the queue. Calculate its time relative to the current system time. The waiting time To eliminate the influence of different time scales on the measurement, the system uses the min-max normalization method to handle waiting time and calculates the average normalized waiting time of the queue. The calculation formula is as follows:
[0093] ;
[0094] In the formula, This is the preset maximum allowable waiting time constant for emergency services. For the first The actual waiting time for each patient.
[0095] Finally, the module will use the doctor efficiency characteristics calculated above. and Queue load characteristics and and waiting time characteristics By cascading and splicing, a high-dimensional environment state vector is constructed. The calculation formula is expressed as follows:
[0096] ;
[0097] The vector It will serve as the environmental observation input for the reinforcement learning agent.
[0098] This implementation step, by quantifying the real-time working status of doctors and the global load distribution of the queue, constructs a digital environmental representation that dynamically reflects the supply and demand relationship of medical resources, providing a comprehensive decision-making context for subsequent intelligent scheduling strategies.
[0099] The reinforcement learning dynamic scheduling module takes the uncertainty score, clinical feature vector, and environmental state vector as input states, and uses a policy network trained based on a reward function that includes a waiting time penalty term to perform mapping calculations and output the target insertion position index.
[0100] Furthermore, in the reinforcement learning dynamic scheduling module, the step of using a policy network trained based on a reward function containing a waiting time penalty term for mapping calculation includes: defining a composite reward function, which includes a positive reward term proportional to the accuracy of prioritizing critical patients and a negative penalty term that increases non-linearly with the dwell time of ordinary patients; using a policy gradient algorithm or a value iteration algorithm as the training framework, storing historical state transition samples using an experience replay mechanism, and updating the weight parameters of the policy network by maximizing the cumulative discounted reward; inputting the current environmental state vector into the pre-trained policy network, and calculating the value assessment value or probability distribution of taking different insertion actions in the current state through forward propagation.
[0101] Furthermore, in the reinforcement learning dynamic scheduling module, the step of outputting the target insertion position index includes: generating a mask vector based on the soft constraint of the first-in-first-out principle to block illegal insertion positions that violate the first-in-first-out principle and reduce the action search space; sorting or sampling the values of legal actions output by the policy network, selecting the position with the highest expected reward as the optimal decision, and generating the corresponding target insertion position index; and dynamically adjusting the estimated waiting time parameters of all patients after that position in the queue based on the target insertion position index to complete the closed loop of a single scheduling decision.
[0102] Specifically, the reinforcement learning dynamic scheduling module first vectorizes and concatenates the features output by the preceding modules to construct the state vector at the current time step. The state vector is the normalized uncertainty score output by the uncertainty-aware inference module. Clinical feature vectors output by the clinical semantic feature extraction module and the environment state vector output by the environment state construction module Composition. Let... The dimension is , The dimension is Then the input state vector The dimension is The vector is input into a pre-trained deep Q-network, which uses a standard multilayer perceptron architecture, consisting of an input layer, three hidden layers, and an output layer. The number of nodes in the input layer corresponds to the dimension of the state vector, the hidden layers all use the ReLU activation function, and the number of nodes is set to 256. The number of nodes in the output layer is... Corresponding to the maximum allowed capacity of the image reading queue, the value of each node in the output layer represents the position at which the current patient is inserted into the queue. action value .
[0103] During the training phase of the policy network, the system defines a composite reward function. This is to guide the agent in learning the optimal strategy. The reward function aims to balance the priority of critically ill patients with the waiting time benefits of ordinary patients; its mathematical expression is defined as:
[0104] ;
[0105] In the formula, The overall priority score is calculated for the current patient based on clinical characteristics and imaging urgency. The insertion position index chosen for the agent, where 0 indicates the head of the queue. This is the current queue length. For the first in the queue The patient's length of stay , and These are preset weighting coefficients and non-linear growth factors. The first term of this function is a positive return term, indicating that high-risk patients are ranked higher. The first item is a positive reward for shorter stays; the second item is a negative penalty, which increases with the length of stay of other patients in the queue. The increase is exponential, punishing behaviors that lead to excessively long wait times for ordinary patients. An experience playback mechanism is used during training to transfer historical state samples. Store the data in a buffer, and update the network weights by randomly sampling batches of data from the buffer and minimizing the mean squared error loss function. :
[0106] ;
[0107] In the formula, As a discount factor, These are the parameters of the target network. Through the above training, the policy network can be trained in a given state. The output can maximize the distribution of action value that accumulates rewards over a long period of time.
[0108] During the inference and output target insertion position index stage, the system first generates a string of length [length missing]. mask vector Based on the soft constraint of the first-in, first-out principle, patients are only allowed to be inserted into valid positions within the current queue length, i.e., for indexes... ,when hour, ,otherwise The system will output the original action value vector from the policy network. With mask vector By adding elements one by one, we obtain the value vector of legal actions. :
[0109] ;
[0110] This operation masks the Q-value of illegal positions as negative infinity, thereby reducing the action search space. Subsequently, the system... Perform the Argmax operation to select the index with the highest expected return as the target insertion position. :
[0111] ;
[0112] Determine the insertion position Then, the system inserts the current patient into the image reading list based on the location, and automatically updates the estimated waiting time parameters for all patients after that location, completing the closed loop of a single scheduling operation.
[0113] This implementation step, by introducing a reward mechanism that includes waiting time penalties and action masking technology, achieves the goal of ensuring priority treatment for critically ill patients while dynamically suppressing the unfair allocation of medical resources caused by ordinary patients being frequently cut in line, thus optimizing the overall allocation efficiency of emergency resources.
[0114] The queue execution and visualization module is used to reconstruct the order of the reading list based on the target insertion position index and to highlight entries with uncertainty scores higher than a preset threshold.
[0115] Furthermore, in the queue execution and visualization module, the step of highlighting entries with uncertainty scores higher than a preset threshold includes: comparing the real-time calculated uncertainty score with a preset anomaly judgment threshold to identify out-of-distribution samples or high-noise samples that the model cannot be certain of; overlaying visual warning labels on the identified abnormal samples in the image reading work list, locking the automatic sorting function of the entry, and generating a prompt signal requiring manual review to prevent model misjudgment; calling the interpretability algorithm to generate an uncertainty heatmap, and overlaying the heatmap when the doctor views the images to intuitively indicate the image areas that cause high uncertainty in the model.
[0116] Specifically, the queue execution and visualization module is deployed between the front-end rendering layer and the back-end data service layer of the PACS image reading workstation. The module first receives the target insertion position index output by the reinforcement learning dynamic scheduling module. and the normalized uncertainty score output by the uncertainty perception and reasoning module. The system calls the list update interface of the radiology information system, based on... The system inserts the current patient's examination entry into the corresponding row of the image reading list and triggers a WebSocket message push to refresh the front-end interface of all online doctors. Simultaneously, the system loads preset anomaly detection thresholds. This threshold is determined based on the statistical distribution of historical misjudgment data. The module executes comparison logic, and the uncertainty score is calculated in real time. If the image data is found to be an out-of-distribution sample or a high-noise sample, the system will immediately update the status field of the entry in the database to "locked" and render a prominent red warning icon next to the patient's name in the front-end list. At this point, the automatic sorting permission for the entry is suspended, ignoring subsequent reinforcement learning scheduling instructions, until the doctor manually clicks the "Review Complete" button, thus generating a prompt signal indicating that manual review is required.
[0117] To visually highlight image regions that cause high uncertainty in the model, the module asynchronously invokes the Grad-CAM algorithm, a gradient-based class activation mapping algorithm, to generate an uncertainty heatmap when a doctor double-clicks to open the high-risk image. This algorithm is based on the ResNet50 backbone network used in the preceding module and selects the feature map set output from the last convolutional layer of ResNet50. ,in Indicates the first Feature maps of each channel. Let... For the network output corresponding to the predicted category The system first calculates the original logic value before Softplus transformation. Relative to feature map Each pixel gradient And calculate the first by global average pooling operation Importance weights of each feature map :
[0118] ;
[0119] In the formula, This represents the total number of pixels in the feature map. After obtaining the weights, the module performs a weighted linear combination of all feature maps and applies the ReLU activation function to filter out features that negatively impact the prediction results, generating a coarse-grained heatmap. :
[0120] ;
[0121] Subsequently, the bilinear interpolation algorithm was used to... Upsampling to the original medical image With the same spatial resolution, and after normalizing the heatmap values to the range of 0 to 255, the Jet pseudo-color map was applied to convert them into RGB format heatmap images. Finally, the thermal image and the original grayscale image are compared using a preset transparency factor. The images are then overlaid and blended to generate the final display image. It will be displayed in the auxiliary window of the viewing interface.
[0122] This implementation step prevents uncertain samples from interfering with the stability of automated sorting through a visual locking mechanism, and uses interpretable algorithms to transform abstract algorithmic uncertainties into visual cues that doctors can perceive, helping doctors to quickly locate artifacts or lesion areas.
[0123] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An intelligent sorting system for prioritizing emergency radiology images, characterized in that, include: The multimodal data acquisition module is used to acquire medical images of patients to be diagnosed and related electronic application form texts; The uncertainty-aware reasoning module is used to extract features from the medical images through an evidence deep learning network and generate a predicted probability representing the lesion category and a score representing the uncertainty of the model using a parameterized Dirichlet distribution. The clinical semantic feature extraction module is used to perform semantic encoding on the electronic application form text and output a clinical feature vector that represents the severity of the disease information contained in the text. The environment state construction module is used to aggregate the current queue length, the historical average reading time of the doctor currently performing the reading task, and the cumulative patient stay time to construct an environment state vector that represents the load on reading resources. The reinforcement learning dynamic scheduling module is used to take the uncertainty score, clinical feature vector and environmental state vector as input states, and use the policy network trained based on the reward function including the waiting time penalty term to perform mapping calculation and output the target insertion position index. The queue execution and visualization module is used to reconstruct the order of the reading list based on the target insertion position index, and to highlight entries with uncertainty scores higher than a preset threshold.
2. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the multimodal data acquisition module, the step of acquiring the medical images of the patient to be diagnosed and the associated electronic application form text includes: Configure a data monitoring interface to monitor data flow changes in the image archiving system in real time. By parsing the unique identifier in the image header file, establish an index association between image data and patient electronic requisition forms in the radiology information system. The text parsing algorithm is called to extract the unstructured clinical description text from the electronic application form, and the sensitive fields are de-identified in combination with the preset privacy dictionary to meet data compliance requirements; The heterogeneous medical image and text data are time-aligned and encapsulated using a data buffer queue to construct a multimodal data packet containing complete diagnostic information, which is then transmitted to the inference memory pool for processing.
3. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the uncertainty-aware reasoning module, the step of extracting features from the medical image using an evidence deep learning network includes: Using a pre-built deep convolutional neural network backbone model, multi-scale convolution and downsampling operations are performed on the input medical images to extract high-dimensional feature maps containing lesion morphology, texture and spatial location; A spatial attention mechanism is introduced to calculate the saliency weight of each pixel in the feature map, suppress background noise areas, and enhance the feature response value of suspected lesion areas to improve the signal-to-noise ratio of the features. The weighted high-dimensional feature map is compressed into a one-dimensional feature vector by global pooling, and the feature vector is mapped to a non-negative evidence space by a fully connected layer to generate an evidence output vector corresponding to each lesion category.
4. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the uncertainty-aware reasoning module, the step of generating the predicted probability representing the lesion category and the score representing the uncertainty of the model using the parameterized Dirichlet distribution includes: The evidence output vector is transformed using a nonlinear activation function to ensure that the generated pseudo-count parameters satisfy the positive definiteness constraint of the Dirichlet distribution, thereby constructing a parameter set describing the multi-class probability distribution. Based on evidence theory, the ratio of pseudo-count parameters of each category to the total evidence strength is calculated to obtain the predicted probability of each lesion category, quantifying the possibility that the current image belongs to a specific critical disease. The reciprocal of the total number of categories and the total strength of evidence is calculated to quantify the cognitive uncertainty caused by the lack of training data distribution or the ambiguity of sample features, resulting in a normalized uncertainty score.
5. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the clinical semantic feature extraction module, the step of semantically encoding the electronic application text includes: We used a pre-trained natural language processing model to segment and vectorize the electronic application text, capturing the contextual semantic dependencies of clinical terms in the text. Perform entity recognition and negation detection, extract key symptom entities and identify their affirmative or negative states, and reverse the feature polarity described as exclusionary symptoms to correct semantic expression; The extracted symptom feature vectors are fused and encoded with emergency triage labels to output clinical feature vectors that can quantitatively reflect the urgency of the patient's current chief complaint and the stability of vital signs.
6. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the environment state construction module, the step of constructing an environment state vector representing the load on viewing resources includes: Set a sliding time window to statistically analyze the average reading time and processing variance of doctors currently performing image reading tasks within the window period, and quantify the current work efficiency and fatigue fluctuation characteristics of doctors. Traverse the current image queue, count the number of images backlogged in the queue and the distribution of their severity levels, and generate load characteristics that reflect the complexity and urgency of the current task to be processed. The waiting time of all patients in the queue is normalized, and the work efficiency feature, load feature and waiting time feature are concatenated to construct an environment state vector that is dynamically updated over time.
7. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the reinforcement learning dynamic scheduling module, the step of performing mapping calculations using a policy network trained based on a reward function that includes a waiting time penalty term includes: Define a composite reward function that includes a positive reward term proportional to the accuracy of prioritizing critical patients and a negative penalty term that increases non-linearly with the length of stay of ordinary patients. The policy gradient algorithm or value iteration algorithm is used as the training framework. The historical state transition samples are stored using the experience replay mechanism. The weight parameters of the policy network are updated by maximizing the cumulative discounted reward. The current environment state vector is input into a pre-trained policy network, and the value assessment or probability distribution of taking different insertion actions in the current state is calculated through forward propagation.
8. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the reinforcement learning dynamic scheduling module, the step of inserting the output target position index includes: A mask vector is generated based on the soft constraint of the first-in-first-out principle to block out illegal insertion positions that violate the first-in-first-out principle and reduce the action search space. The values of the legal actions output by the policy network are sorted or sampled, and the position with the highest expected return is selected as the optimal decision, generating the corresponding target insertion position index. Based on the target insertion position index, the estimated waiting time parameters of all patients after that position in the queue are dynamically adjusted to complete the closed loop of a single scheduling decision.
9. The intelligent sorting system for prioritizing radiological emergency images according to claim 1, characterized in that, In the queue execution and visualization module, the step of highlighting entries with uncertainty scores higher than a preset threshold includes: The uncertainty score calculated in real time is compared with the preset anomaly detection threshold to identify out-of-distribution samples or high-noise samples that the model cannot be certain of. Visual warning labels are overlaid on the identified abnormal samples in the image review work list, the automatic sorting function of the entry is locked, and a prompt signal requiring manual review is generated to prevent model misjudgment. An interpretability algorithm is used to generate an uncertainty heatmap, which is then overlaid and displayed when doctors review images, providing a visual indication of image regions that cause high uncertainty in the model.