An assisted treatment system and method for refractory suppurative periapical periodontitis

Abstract

This invention relates to the field of intelligent assisted dental treatment technology, and discloses an assisted treatment system and method for refractory purulent periapical periodontitis. The assisted treatment system includes: a multimodal image fusion module: fusing CT / OCT images to generate a 3D map, segmenting the infected area using a lightweight CNN and outputting a probability map; a reinforcement learning debridement navigation module: based on the DDPG algorithm, dynamically planning the path based on biofilm thickness to optimize dentin damage risk; an intelligent drug control module: locating the catheter according to the infection map, triggering drug release with a pH / enzyme sensor, and adjusting the delivery pressure with a PID controller; and a biosensor monitoring module: detecting inflammatory factors, predicting recurrence risk with an LSTM, and triggering image rescanning and path updates when thresholds are exceeded. This invention achieves precise removal of periapical periodontitis lesions and active suppression of recurrence through multimodal image fusion, reinforcement learning dynamic debridement navigation, microenvironment-responsive drug control, and time-series prediction closed-loop feedback.

Description

Technical Field

[0001] This invention relates to the field of dental intelligent auxiliary treatment technology, specifically to an auxiliary treatment system and method for refractory purulent periapical periodontitis. Background Technology

[0002] Reversible purulent periapical periodontitis, due to the complex anatomy of the root canal system and the drug resistance of biofilms, easily progresses to chronic, refractory infection. Traditional treatments often face problems such as incomplete lesion removal, high postoperative recurrence rates, and significant trauma from secondary surgeries. With the development of precision medicine and intelligent diagnostic and treatment technologies, there is an urgent need for a closed-loop treatment system that integrates multimodal image navigation, dynamic pathway planning, targeted drug control, and proactive recurrence early warning to achieve a paradigm shift in periapical periodontitis treatment from experience-driven to data-driven approaches.

[0003] In current technologies, the clinical diagnosis and treatment of periapical periodontitis mainly relies on cone-beam computed tomography (CBCT) guided debridement surgery, combined with manual instrumentation to remove infected tissue. Postoperatively, local drug irrigation or systemic antibiotic treatment is used to suppress infection recurrence. Some improved approaches use optical coherence tomography (OCT) to assist in detecting biofilm distribution or utilize electrochemical sensors to monitor inflammatory factor concentrations. Debridement pathway planning is mostly based on preoperative static imaging data, drug release depends on preset time or dosage strategies, and recurrence risk assessment is conducted through regular imaging follow-ups combined with retrospective analysis of clinical symptoms.

[0004] However, existing technologies still have some shortcomings. First, a single imaging modality cannot simultaneously capture the cross-scale characteristics of bone tissue and biofilm, leading to deviations in debridement planning. Second, static path planning lacks dynamic feedback on intraoperative biofilm thickness changes, resulting in damage to healthy tissue or residual lesions. In addition, empirical drug release is not linked to infection probability distribution and microenvironment parameters, leading to insufficient local drug concentration or excessive diffusion. Finally, discrete monitoring methods cannot predict recurrence risk in real time, delaying the timing of secondary intervention. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adjunctive treatment system and method for refractory suppurative periapical periodontitis, solving the problems of incomplete lesion removal and high recurrence rate caused by the isolation of data, static operation planning, and lack of feedback mechanisms in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adjunctive treatment system for refractory suppurative periapical periodontitis, comprising:

[0007] The multimodal image fusion module is used to fuse low-dose cone-beam CT images and optical coherence tomography (OCT) images to generate a three-dimensional fused image of the periapical region. It also segments the infected area through a lightweight convolutional neural network and outputs an infection probability map.

[0008] The reinforcement learning debridement navigation module uses the three-dimensional fused image and real-time biofilm thickness data detected by OCT to dynamically plan the debridement path using a depth deterministic strategy gradient algorithm, and optimizes the path according to the degree of damage to dentin caused by instrument operation.

[0009] The intelligent drug control module locates the position of the drug delivery catheter based on the infection probability map, triggers drug release by detecting periapical microenvironment parameters through pH and enzyme activity sensors, and dynamically adjusts the drug delivery pressure based on a PID control algorithm.

[0010] The biosensor monitoring module detects the concentration of inflammatory factors in real time and predicts the risk of infection recurrence through a long short-term memory network. When the predicted recurrence risk exceeds a preset threshold, the multimodal image fusion module is triggered to rescan the periapical region and update the debridement path of the reinforcement learning debridement navigation module based on the regenerated fused image.

[0011] Preferably, in the multimodal image fusion module:

[0012] The radiation dose of the low-dose cone-beam CT images is 50% of the conventional dose;

[0013] The OCT images have a scanning precision at the micrometer level and are used to detect the thickness of the biofilm on the root canal surface.

[0014] The 3D fused image is weighted using a formula. Generate, where α is the fusion weight coefficient, For the registered OCT data; V CBCT Three-dimensional voxel data for low-dose cone-beam CT; V fused This is the fused 3D image data.

[0015] Preferably, the lightweight convolutional neural network is the MobileNet-V3 model, and the formula for calculating its output infection probability map is as follows:

[0016]

[0017] Among them, DWConv k This indicates a depthwise separable convolutional layer of the k-th layer; σ(·) is the Sigmoid activation function; w k P represents the weight parameters of the k-th convolutional kernel. inf This is a map showing the infection probability in the periapical region; K is the total number of depthwise separable convolutional layers in the lightweight convolutional network.

[0018] Preferably, in the reinforcement learning debridement navigation module:

[0019] The reward function is designed using the Deep Deterministic Policy Gradient Algorithm (DDPG):

[0020] r t =β·ΔB t -γ·D t -η·||a t -a t-1 || 2 ;

[0021] Where, ΔB t D represents the change in biofilm thickness at time t; t To accumulate dentin damage; a t is the action vector; β, γ, and η are the weighting coefficients.

[0022] The reward function of the deep deterministic policy gradient algorithm is designed as follows:

[0023]

[0024] Where, Δd bio Δh represents the reduction in biofilm thickness. k The depth of dentin damage is represented by β and γ, which are weighting coefficients. h represents the initial biofilm thickness. th K1 represents the dentin safety damage threshold; K1 is the total number of dentin zones.

[0025] Policy Network μ θ Update via gradient ascent:

[0026]

[0027] θ represents the trainable parameters of the policy network, used to generate action a; φ represents the trainable parameters of the Critic network; s represents the state vector; a represents the action vector; J represents the objective function for policy optimization. Let be the gradient of the Critic network for action aa; The gradient of the policy network with respect to parameter θ;

[0028] The debridement path is updated every Δt time interval based on the latest biofilm thickness data, where Δt is a preset fixed time interval.

[0029] Preferably, the dentin damage depth Δh k The weighting coefficients are calculated using real-time scanning data from OCT images, with weighting coefficients β = 0.6 and γ = 0.4.

[0030] Preferably, in the intelligent drug control module:

[0031] The trigger threshold of the pH sensor is pH≤5.5, and the trigger threshold of the enzyme activity sensor is MMP-8≥50U / L;

[0032] The pressure regulation formula of the PID control algorithm is:

[0033]

[0034] Where u(t) is the adjusted injection pressure used to control the drug delivery rate; e(t) = r(t) - y(t) represents the drug release error, r(t) is the target drug dose; y(t) is the real-time release amount; K p K is the proportionality coefficient. i K is the integral coefficient; d τ is the differential coefficient; τ is the integral variable, representing any moment within the time interval [0, t]. Let e(t) be the first derivative of the error signal with respect to time, representing the rate of change of the error.

[0035] Preferably, the location of the drug delivery catheter is determined by the centroid coordinates of an infection probability map:

[0036]

[0037] Where x, y, z represent the voxel coordinates in the 3D fused image coordinate system; P inf (x,y,z) represents the probability that the voxel point with coordinates (x,y,z) belongs to the infected region, and its value ranges from [0,1]; Σ represents the summation over all voxel points in the 3D fused image; x c The coordinates of the centroid of the infected area along the x-axis are used to locate the target position of the drug delivery catheter.

[0038] Preferably, in the biosensor monitoring module:

[0039] The inflammatory factors include IL-6 and CRP, and their concentrations are detected in real time using a flexible biosensor.

[0040] The input to the Long Short-Term Memory (LSTM) network is time-series data {C}. II-6 (t),C CRP (t),d bio (t)}, where C II-6 (t) represents the concentration of interleukin-6 detected at time t, in pg / mL; C CRP (t) represents the C-reactive protein concentration detected at time t, in mg / L; d bio (t) represents the thickness of the root canal surface biofilm measured in real time at time t using optical coherence tomography (OCT), in μm, and outputs the recurrence probability P. recThe calculation formula is:

[0041] P rec =σ(W p ·h t +b p );

[0042] Among them, h t For LSTM hidden states; σ(·) is the Sigmoid function; W p b is the weight matrix of the fully connected layer; p This is the bias vector for the fully connected layer.

[0043] Preferably, the preset threshold is P. rec ≥0.8, and the interval between updating the reinforcement learning debridement path in the re-scanned fused images is consistent with the initial scan interval.

[0044] This invention also provides an adjunctive treatment method for refractory suppurative periapical periodontitis, comprising the following steps:

[0045] Periapical imaging data were acquired using low-dose cone-beam CT and optical coherence tomography, and three-dimensional fused images were generated and the infected area was segmented.

[0046] Based on fused imaging and real-time biofilm thickness data, a deep reinforcement learning algorithm is used to dynamically plan the debridement path and optimize the damage to dentin caused by instrument operation.

[0047] The drug delivery catheter is located based on the infection probability map, and drug release is triggered by the detection of microenvironment parameters by sensors, and the delivery pressure is dynamically adjusted.

[0048] Real-time monitoring of inflammatory factor concentrations predicts recurrence risk; when the risk exceeds a threshold, images are re-scanned and the debridement path is updated.

[0049] This invention provides an adjunctive treatment system and method for refractory suppurative periapical periodontitis. It has the following beneficial effects:

[0050] 1. By fusing cross-modal data from low-dose cone-beam computed tomography (CBCT) and optical coherence tomography (OCT), and combining this with a lightweight convolutional neural network, precise segmentation and three-dimensional localization of periapical infection areas are achieved. This invention overcomes the limitations of single-image modalities, effectively distinguishing between minute lesions and healthy tissue, and providing high-precision spatial navigation for debridement pathway planning and targeted drug delivery.

[0051] 2. Based on a deep reinforcement learning algorithm, this invention integrates real-time biofilm thickness data with dentin damage feedback to dynamically adjust the movement trajectory and operating parameters of the debridement instruments. This invention addresses the rigidity of traditional static planning, significantly reducing the risk of accidental dentin damage while improving debridement efficiency.

[0052] 3. By using an infection probability map to drive drug delivery, combined with a dual triggering mechanism of pH and enzyme activity sensors, on-demand release of antibacterial and anti-inflammatory drugs is achieved. This invention overcomes the blindness of traditional empirical drug administration, ensuring precise drug coverage of the core lesion area, inhibiting drug resistance, and promoting tissue repair.

[0053] 4. Based on the dynamic analysis of inflammatory factors using long short-term memory networks, a predictive model for infection recurrence probability is constructed. This invention achieves early warning of recurrence risk by mining the temporal correlation of multi-dimensional biomarkers, providing a time window for proactive intervention and avoiding the trauma of secondary surgery.

[0054] 5. Through real-time data interaction between imaging, debridement, medication, and monitoring modules, a closed-loop system of "diagnosis-treatment-feedback" is formed. This invention breaks through the fragmentation of traditional segmented treatment, enabling dynamic iterative optimization of periapical periodontitis treatment strategies and improving the treatment adaptability and clinical operability of complex cases. Attached Figure Description

[0055] Figure 1 This is a diagram of the module architecture of the present invention;

[0056] Figure 2 This is a flowchart of the method of the present invention;

[0057] Figure 3 This is a schematic diagram of the device structure of the present invention;

[0058] Figure 4 This is a schematic diagram of the structure of the catheter of the present invention.

[0059] Among them, 2. External pipeline; 3. Equipment body; 5. Pus suction tube; 6. Catheter; 8. First switch; 9. Second switch; 10. Third switch; 301. Cloud system. Detailed Implementation

[0060] 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.

[0061] Please see the appendix Figure 1 This invention provides an adjunctive treatment system for refractory suppurative periapical periodontitis, comprising:

[0062] The multimodal image fusion module is used to fuse low-dose cone-beam CT images and optical coherence tomography (OCT) images to generate a three-dimensional fused image of the periapical region. It also segments the infected area using a lightweight convolutional neural network and outputs an infection probability map.

[0063] In this embodiment, the multimodal image fusion module is used to integrate complementary information from low-dose cone-beam CT images and optical coherence tomography (OCT) images to generate high-precision three-dimensional fused images. It also uses a lightweight convolutional neural network to achieve automated segmentation of the infected area, providing spatial positioning basis for subsequent debridement path planning and drug control.

[0064] Low-dose cone-beam computed tomography (CBCT) imaging is designed with reducing radiation dose as its core principle. Radiation dose control is achieved by optimizing tube voltage, tube current, and exposure time, minimizing patient radiation exposure risk while preserving bone tissue resolution. Optical coherence tomography (OCT) imaging, based on near-infrared interferometry, provides micron-level high-resolution imaging of root canal surface biofilms and soft tissues, compensating for the insufficient contrast of CT images in soft tissue structures. Spatial and resolution differences between the two modalities are addressed through a cross-scale registration algorithm, ensuring effective fusion of multi-source data.

[0065] To achieve accurate fusion of CT and OCT images, a feature-point-based registration method is employed. Specifically, bony landmarks (such as the apical foramen and alveolar ridge crest) in CT images and biomembrane-dentin interface feature points in OCT images are extracted. Feature matching is performed using the Scale Invariant Feature Transform (SIFT) algorithm, and the OCT data is mapped to the CT coordinate system using a rigid body transformation model. Preferably, the registration error is further optimized using the Iterative Closest Point (ICP) algorithm to ensure that the three-dimensional spatial alignment accuracy meets clinical requirements.

[0066] The fusion process uses an adaptive weighting formula. Where α is the fusion weight coefficient. For the registered OCT data; V CBCT Three-dimensional voxel data for low-dose cone-beam CT; V fused The data consists of fused 3D image data. Preferably, the weighting coefficients are optimized using cross-validation to accommodate the differences in imaging characteristics of periapical lesions among different patients.

[0067] The lightweight convolutional neural network is based on the MobileNet-V3 architecture and is trained end-to-end on 3D fused images. Its output infection probability map is calculated using the following formula:

[0068]

[0069] Among them, DWConv kThis indicates a depthwise separable convolutional layer of the k-th layer; σ(·) is the Sigmoid activation function; w k P represents the weight parameters of the k-th convolutional kernel. inf This is a map showing the infection probability in the periapical region; K is the total number of depthwise separable convolutional layers in the lightweight convolutional network.

[0070] The network input is a sequence of 2D slices reconstructed from multiple planes. Local texture features are extracted through depthwise separable convolutional layers, and multi-scale contextual information is preserved by combining skip connections. The output layer is activated by the Sigmoid function to generate a pixel-wise infection probability map P. inf Its numerical range is mapped to [0,1], representing the confidence level of each voxel point belonging to the infected region. Preferably, the network training uses a weighted cross-entropy loss function to solve the class imbalance problem between infected and uninfected regions.

[0071] The infection probability map, as the core output, provides catheter positioning information for the intelligent drug control module. Specifically, the coordinates of the drug delivery target point are determined by centroid calculation for high-confidence infection areas in the probability map. When the biosensor monitoring module detects that the recurrence risk exceeds the threshold, it triggers this module to re-execute the image acquisition, registration, and segmentation process to ensure the timeliness of the fused images and the integrity of lesion coverage.

[0072] By combining low-dose CT and OCT modal complementarity, this module reduces radiation hazards while improving the detection sensitivity of micro-infection foci; the closed-loop feedback mechanism effectively addresses changes in lesion morphology during treatment by dynamically updating fused images, laying a data foundation for adaptive control of the entire system.

[0073] The reinforcement learning debridement navigation module uses biofilm thickness data from 3D fusion imaging and real-time OCT detection to dynamically plan the debridement path using a depth deterministic gradient algorithm, and optimizes the path based on the degree of damage to dentin caused by instrument operation.

[0074] In this embodiment, the reinforcement learning debridement navigation module is based on a deep reinforcement learning algorithm. It combines real-time image data and biofilm thickness to dynamically plan the debridement path and optimize the risk of damage to dentin caused by instrument operation, thereby achieving precise removal of periapical lesions.

[0075] The state space input includes a 3D fused image generated by the multimodal image fusion module, real-time biofilm thickness data (measured via optical coherence tomography, OCT), and the current position coordinates of the debridement instrument. The 3D fused image provides anatomical information of the periapical region, the real-time biofilm thickness data characterizes the progress of lesion removal, and the instrument coordinates are fed back in real time via a spatial encoder. The motion space output consists of the motion direction of the debridement instrument (including pitch and yaw angles) and cutting force parameters, controlling the instrument's trajectory and operational intensity. Preferably, the motion space uses continuous value encoding to adapt to the output characteristics of the depth-deterministic gradient algorithm.

[0076] The reward function of the deep deterministic policy gradient algorithm is designed as follows:

[0077]

[0078] Where, Δd bio This indicates the amount of reduction in biofilm thickness at the current moment, calculated using real-time OCT monitoring data, reflecting the effectiveness of the debridement procedure; The initial biofilm thickness is used to normalize the reward value and ensure comparability between different cases; Δh k The cutting depth of the debridement instrument on the k-th dentin region was calculated using real-time OCT image data. Its value exceeded the safety threshold h. th The penalty is triggered at certain times; β and γ are the weight coefficients of the reward and penalty items, respectively, which are determined through clinical experience and cross-validation to balance the needs of debridement efficiency and safety; K1 is the total number of dentin zones.

[0079] The debridement path is updated every preset time interval Δt, based on the latest biofilm thickness data, instrument placement, and dentin damage risk assessment results. Preferably, the time interval Δt is dynamically adjusted according to the lesion complexity to ensure real-time path planning and computational efficiency. During the path update process, the reinforcement learning agent regenerates the action sequence based on the current state and verifies the feasibility of the path in a simulation environment. If the updated path significantly increases the risk of dentin damage, such as... This triggers the path rollback mechanism, re-invoking the historical optimal path for execution.

[0080] The deep deterministic policy gradient algorithm is pre-trained in a simulation environment built based on real case data, including multimodal imaging, biofilm thickness variation curves, and dentin biomechanical response models. During training, the policy network and value network employ a double-delay update (TD3) mechanism to mitigate the Q-value overestimation problem.

[0081] Preferably, the model adapts to the anatomical variations of different patients through transfer learning. Specifically, the pre-trained network's convolutional layers are frozen, and only the parameters of the fully connected layers are fine-tuned.

[0082] The endpoint coordinates of the debridement path serve as a trigger signal for the intelligent drug control module. When the device reaches the endpoint of the path, the drug delivery process is automatically initiated.

[0083] Preferred, based on appendix Figure 3 and attached Figure 4 The device body 3 is connected to the cloud system 301 by wires. Under the control of the cloud system 301, it can control the first switch 8, the second switch 9 and the third switch 10 on the surface of the device body 3. The first switch 8 is used to control the pus suction mechanism built into the device body 3. The pus suction structure is existing technology and will not be described in detail here. When the pus suction structure is working, the pus suction tube 5 in the external pipe 2 fixed on the outside can remove the pus along the path and at the end point. The second switch 9 is used to control the catheter 6 in the external pipe 2 to deliver the drug through the intelligent drug control module. The third switch 10 is used to control the opening and closing of the device body 3.

[0084] The intelligent drug control module locates the position of the drug delivery catheter based on the infection probability map, triggers drug release by detecting periapical microenvironment parameters through pH and enzyme activity sensors, and dynamically adjusts the drug delivery pressure based on a PID control algorithm.

[0085] The intelligent drug control module dynamically triggers drug release and adjusts delivery pressure based on the probability distribution of the infected area and the periapical microenvironment parameters, thereby achieving precise delivery and dosage control of drugs to the lesion site to inhibit infection recurrence and promote tissue repair.

[0086] The location of the drug delivery catheter is determined using the centroid coordinates of the infection probability map, calculated using the following formula:

[0087]

[0088] Where x, y, z represent the voxel coordinates in the 3D fused image coordinate system; P inf (x,y,z) represents the probability that the voxel point with coordinates (x,y,z) belongs to the infected region, and its value ranges from [0,1]; Σ represents the summation over all voxel points in the 3D fused image; x c The coordinates of the centroid of the infected area along the x-axis are used to locate the target position of the drug delivery catheter.

[0089] The catheter tip is equipped with a microneedle array, which is moved with millimeter-level precision through a piezoelectric drive mechanism to ensure that the drug release target coincides with the core infection area.

[0090] The triggering conditions for drug release are based on real-time monitoring data of the periapical microenvironment, including pH and enzyme activity indicators:

[0091] pH sensor: Real-time monitoring of the pH of periapical tissue. When the detected value is pH≤5.5, it is determined to be an acidic microenvironment caused by infection, triggering the release of antibacterial drugs;

[0092] Enzyme activity sensor: detects the concentration of matrix metalloproteinase-8 (MMP-8). When MMP-8 ≥ 50 U / L, it is determined to be an active period of tissue destruction, triggering the release of anti-inflammatory drugs.

[0093] Preferably, the sensor adopts an integrated design of flexible electrode and microfluidic chip, and transmits the signal to the control unit wirelessly to avoid interfering with the operation of debridement instruments.

[0094] The release trigger condition is:

[0095] pH≤τ pH AND Enzyme≥τ enzyme ;

[0096] pH represents the real-time acidity or alkalinity value of the periapical region, measured by a miniature pH sensor, reflecting the acidity of the infection microenvironment; τ pH The pH threshold for drug release is indicated; values ​​below this threshold suggest the beginning of dentin demineralization. Enzyme represents the collagenase activity in the periapical region, measured using an enzyme activity biosensor, reflecting the metabolic activity of the biomembrane. τ enzyme The enzyme activity threshold that triggers drug release; values ​​above this threshold indicate that the infection is in an acute progressive phase; AND is the logical AND operator.

[0097] The drug delivery pressure is dynamically adjusted using a proportional-integral-derivative (PID) control algorithm to match the error variation between the target drug dose and the real-time release rate. The PID control formula is:

[0098]

[0099] Where u(t) is the adjusted injection pressure used to control the drug delivery rate; e(t) = r(t) - y(t) represents the drug release error, r(t) is the target drug dose; y(t) is the real-time release amount; K p K is the proportionality coefficient. i K is the integral coefficient; d τ is the differential coefficient; τ is the integral variable, representing any moment within the time interval [0, t]. Let e(t) be the first derivative of the error signal with respect to time, representing the rate of change of the error.

[0100] The PID controller employs an anti-integral saturation design to prevent pressure overshoot caused by long-term error accumulation.

[0101] The biosensor monitoring module detects the concentration of inflammatory factors in real time and predicts the risk of infection recurrence through a long short-term memory network. When the predicted risk of recurrence exceeds a preset threshold, the multimodal image fusion module is triggered to rescan the periapical region and update the debridement path of the reinforcement learning debridement navigation module based on the newly generated fused images.

[0102] In this embodiment, the biosensor monitoring module detects changes in the concentration of inflammatory factors and the thickness of the biofilm in real time, combines this with a time-series prediction model to assess the risk of infection recurrence, and drives a closed-loop feedback mechanism to update the treatment strategy, thereby achieving dynamic monitoring and adaptive adjustment of the periapical periodontitis treatment process.

[0103] The module integrates a flexible biosensor array for continuous monitoring of specific inflammatory markers in the periapical microenvironment. Specifically, the detection of interleukin-6 (IL-6) concentration is based on electrochemical impedance spectroscopy, utilizing the impedance change generated by the antigen-antibody specific binding reaction, which is then converted into a concentration reading by a signal amplification circuit. The detection of C-reactive protein (CRP) concentration employs surface plasmon resonance (SPR) technology, utilizing the resonance angle shift caused by the binding of CRP molecules to the sensor surface modified with gold nanoparticles. The shift angle is quantified by an optical system to retrieve the concentration value. Preferably, the sensor array is encapsulated in a biocompatible material, and the data is uploaded to the processing unit in real time via a wireless transmission module.

[0104] The input to a Long Short-Term Memory (LSTM) network is multi-dimensional time-series data, including IL-6 concentration C. II-6 (t), CRP concentration C CRP (t) and real-time biofilm thickness d bio (t). IL-6 and CRP concentration data are sampled at minute-level frequencies, while biofilm thickness data is updated at preset time intervals via optical coherence tomography (OCT). Preferably, the input data is standardized to eliminate dimensional differences, and a time series is constructed using a sliding time window (e.g., a 6-hour window) to capture the dynamic trends of inflammatory factors.

[0105] The output of the LSTM network is the infection recurrence probability P. rec The calculation formula is:

[0106] P rec =σ(W p ·h t +b p );

[0107] Among them, h t For LSTM hidden states; σ(·) is the Sigmoid function; W p b is the weight matrix of the fully connected layer; p This is the bias vector for the fully connected layer.

[0108] Preferably, the LSTM network contains a two-layer hidden structure, the number of hidden units is adaptively adjusted according to the input dimension, and Dropout regularization is used during training to alleviate overfitting.

[0109] When the predicted recurrence probability P rec When the threshold is exceeded (e.g., ≥0.8), the following closed-loop control process is triggered:

[0110] Image rescan command: Send a signal to the multimodal image fusion module to start a combined scan of low-dose cone-beam CT and OCT to update the three-dimensional fused image of the periapical region;

[0111] Path update command: Input the updated image data into the reinforcement learning debridement navigation module to replan the debridement path and focus on removing new or residual infected lesions;

[0112] Drug parameter reset: Notify the intelligent drug control module to adjust the PID coefficients to adapt to the drug delivery needs of newly infected areas.

[0113] LSTM cell state update formula:

[0114]

[0115] Among them, W f W i W C W o These are the weight matrices for the forget gate, input gate, cell state, and output gate, respectively; b f b i b C b o These are the bias terms for the corresponding gates, initialized to zero vectors; h t-1 The previous hidden state encodes the historical trend of inflammatory factor changes; x t The input feature vector is σ, the sigmoid activation function is σ, the hyperbolic tangent activation function is tanh, and ⊙ represents element-wise multiplication; f t The output of the forget gate determines how many historical cell states C are retained. t-1 i t The input gate output controls the new candidate state. The update ratio; For candidate cell states, store newly added information at the current moment; C t The updated cell state integrates historical and current information; t As the output gate, it determines the current hidden state h. t The output ratio; h t Hide the current state.

[0116] The prediction loss function uses weighted cross-entropy:

[0117]

[0118] Among them, y i Let i be the true label of the i-th sample; For the first The predicted recurrence probability of each sample is output by the LSTM model; For the first The class weight coefficients of each sample are used to alleviate the class imbalance problem; Σ represents the summation over all training samples.

[0119] By fusing multimodal data on inflammatory factors and biofilm thickness, the robustness of recurrence risk assessment can be improved; the temporal modeling capability of LSTM networks can effectively capture the dynamic patterns of inflammation progression and compensate for the lag of traditional threshold alarm mechanisms; the closed-loop feedback mechanism connects the data flow of "monitoring-diagnosis-treatment" to ensure the system's response speed and the timeliness of treatment strategies.

[0120] Once the reinforcement learning-based debridement navigation module completes the debridement path execution, the endpoint coordinates of the debridement path serve as a trigger signal to initiate the drug release process of the intelligent drug control module. Simultaneously, drug delivery pressure data is fed back to the biosensor monitoring module in real time. If uneven drug distribution is detected (such as pressure fluctuations exceeding a threshold), the multimodal image fusion module is triggered to rescan, update the infection probability map, and recalculate the centroid coordinates, forming a closed-loop control of "debridement-drug administration-monitoring".

[0121] The adjunctive treatment method for refractory purulent periapical periodontitis described below can be referred to in conjunction with the adjunctive treatment system for refractory purulent periapical periodontitis described above.

[0122] Please see the appendix Figure 2 The present invention also provides an adjunctive treatment method for refractory suppurative periapical periodontitis, comprising the following steps:

[0123] Periapical imaging data were acquired using low-dose cone-beam CT and optical coherence tomography, and three-dimensional fused images were generated and the infected area was segmented.

[0124] Based on fused imaging and real-time biofilm thickness data, a deep reinforcement learning algorithm is used to dynamically plan the debridement path and optimize the damage to dentin caused by instrument operation.

[0125] The drug delivery catheter is located based on the infection probability map, and drug release is triggered by the detection of microenvironment parameters by sensors, and the delivery pressure is dynamically adjusted.

[0126] Real-time monitoring of inflammatory factor concentrations predicts recurrence risk; when the risk exceeds a threshold, images are re-scanned and the debridement path is updated.

[0127] The method in this embodiment can be used to execute the above system embodiment, and its principle and technical effect are similar, so it will not be described again here.

[0128] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adjunctive treatment system for refractory suppurative periapical periodontitis, characterized in that, include: The multimodal image fusion module is used to fuse low-dose cone-beam CT images and optical coherence tomography (OCT) images to generate a three-dimensional fused image of the periapical region. It also segments the infected area through a lightweight convolutional neural network and outputs an infection probability map. The reinforcement learning debridement navigation module uses the three-dimensional fused image and real-time biofilm thickness data detected by OCT to dynamically plan the debridement path using a depth deterministic strategy gradient algorithm, and optimizes the path according to the degree of damage to dentin caused by instrument operation. The intelligent drug control module locates the position of the drug delivery catheter based on the infection probability map, triggers drug release by detecting periapical microenvironment parameters through pH and enzyme activity sensors, and dynamically adjusts the drug delivery pressure based on a PID control algorithm. The biosensor monitoring module detects the concentration of inflammatory factors in real time and predicts the risk of infection recurrence through a long short-term memory network. When the predicted recurrence risk exceeds a preset threshold, the multimodal image fusion module is triggered to rescan the periapical region and update the debridement path of the reinforcement learning debridement navigation module based on the regenerated fused image. In the multimodal image fusion module: The radiation dose of the low-dose cone-beam CT images is 50% of the conventional dose; The OCT images have a scanning precision at the micrometer level and are used to detect the thickness of the biofilm on the root canal surface. The 3D fused image is weighted using a formula. Generate, where To integrate the weighting coefficients, The registered OCT data; Three-dimensional voxel data for low-dose cone-beam CT; The resulting 3D image data; The lightweight convolutional neural network is the MobileNet-V3 model, and its output infection probability map is calculated using the following formula: ； in, Indicates the first Layer depth can be separated into convolutional layers; Use the Sigmoid activation function; For the first The weight parameters of the convolutional kernel; This is a graph showing the infection probability in the periapical region. The total number of depthwise separable convolutional layers in a lightweight convolutional network; In the reinforcement learning debridement navigation module: The reward function of the deep deterministic policy gradient algorithm is designed as follows: ； in, This represents the reduction in biofilm thickness. Depth of dentin damage; and These are the weighting coefficients; This represents the initial biofilm thickness. The threshold for safe damage to dentin; The total number of dentin zones; The debridement path is spaced at intervals. The time is updated based on the latest biofilm thickness data, among which, This is a preset fixed time interval.

2. The adjunctive treatment system for refractory suppurative periapical periodontitis according to claim 1, characterized in that, The depth of dentin damage The weighting coefficients are calculated using real-time scanning data from OCT images. =0.6、 =0.

4.

3. The adjunctive treatment system for refractory suppurative periapical periodontitis according to claim 1, characterized in that, In the intelligent drug control module: The trigger threshold of the pH sensor is pH≤5.5, and the trigger threshold of the enzyme activity sensor is MMP-8≥50U / L; The pressure regulation formula of the PID control algorithm is: ； in, The adjusted injection pressure is used to control the drug delivery rate; This indicates the error in drug release. The target drug dose; This refers to the real-time release amount; This is the proportionality coefficient; The integral coefficient; These are the differential coefficients; Let be the integral variable, representing any moment within the time interval [0, t]. Error signal The first derivative with respect to time characterizes the rate of change of the error.

4. The adjunctive treatment system for refractory suppurative periapical periodontitis according to claim 3, characterized in that, The location of the drug delivery catheter is determined by the centroid coordinates of the infection probability map: ； in, Represents the coordinates of a voxel point in the coordinate system of a 3D fused image; Indicates coordinates as The probability value of a voxel point belonging to the infected region, with a value range of [0,1]; This represents the summation of all voxel points in the 3D fused image; The coordinates of the centroid of the infected area along the x-axis are used to locate the target position of the drug delivery catheter.

5. The adjunctive treatment system for refractory suppurative periapical periodontitis according to claim 1, characterized in that, In the biosensor monitoring module: The inflammatory factors include IL-6 and CRP, and their concentrations are detected in real time using a flexible biosensor. The input to the Long Short-Term Memory (LSTM) network is time-series data. ,in, for Interleukins detected at all times Concentration, in pg / mL; for The concentration of C-reactive protein detected at any given time, in mg / L; for The thickness of the root canal surface biofilm, measured in μm, is measured in real time using optical coherence tomography (OCT), and the recurrence probability is output. The calculation formula is: ； in, This is the hidden state of the LSTM; For the Sigmoid function; This is the weight matrix of the fully connected layer; This is the bias vector for the fully connected layer.

6. The adjunctive treatment system for refractory suppurative periapical periodontitis according to claim 5, characterized in that, The preset threshold is ≥0.8, and the interval between updating the reinforcement learning debridement path in the re-scanned fused images is consistent with the initial scan interval.

Images