Internet hospital-based astigmatism artificial intelligence assisted diagnosis and treatment system

Abstract

This invention relates to the field of medical artificial intelligence software, and discloses an AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital. The system comprises five modules: a patient information management module, an intelligent diagnostic assistance module, a remote monitoring module, a disease progression risk warning and prognosis prediction module, and an online consultation and eye health education module. The image segmentation model uses wavelet transform operators to decompose and extract multi-scale frequency domain features from the original 3D-OCT volumetric data of the eye. These multi-scale frequency domain features are then merged and concatenated with the original 3D-OCT volumetric data of the eye in the channel dimension to generate a multi-channel feature cube. This multi-channel feature cube is input into the Unet algorithm network unit, and after encoder, decoder, and hierarchical skip connection operations, a multi-channel probability map is generated. Finally, the multi-channel probability map is converted into a binary lesion segmentation mask for output. The intelligent image analysis module of this invention can improve the accuracy and efficiency of image screening for early identification of pathological myopia.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence software, and more particularly to an AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital. Background Technology

[0002] Refractive errors (such as myopia, hyperopia, and astigmatism) are highly prevalent visual disorders worldwide. In particular, with the continuous increase in the myopic population, pathological myopia caused by high myopia is often accompanied by serious organic fundus lesions (such as posterior staphyloma, macular schisis, and paradiscal atrophy), which is an important cause of irreversible visual impairment.

[0003] Fundus lesions in pathological myopia are often structurally complex and have extremely low contrast. Existing deep learning image segmentation algorithms (such as traditional convolutional neural networks (CNNs) or conventional U-Nets) typically employ conventional pooling layer downsampling operations when processing volume data to expand the receptive field. This operation irreversibly destroys and loses a large amount of high-frequency spatial details (such as minute retinal tears, edges of micro-schisis cavities, etc.), resulting in blurred lesion segmentation. Furthermore, most existing technologies remain at the level of qualitative lesion identification in a two-dimensional plane, unable to automatically extract and accurately calculate key three-dimensional spatial indicators such as the maximum depth and total volume of posterior staphyloma and the multi-level spatial volume of macular schisis. This heavily relies on doctors manually drawing each layer, which is time-consuming, labor-intensive, and subject to significant subjective errors. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital, the system comprising: The patient information management module is configured to collect and store patients' ocular 3D-OCT body data and identity and medical history characteristics. The intelligent diagnostic assistance module includes an intelligent image analysis unit; The image intelligent analysis unit is configured to call the lightweight automated image segmentation model TL-TransUNet to perform three-dimensional structural analysis on the 3D-OCT volume data of the eye. The image segmentation model decomposes the original 3D-OCT volume data of the eye using wavelet transform operators to extract multi-scale frequency domain features, and merges and concatenates the multi-scale frequency domain features with the original 3D-OCT volume data of the eye in the channel dimension to generate a multi-channel feature cube. The multi-channel feature cube is input into the Unet algorithm network unit, and after encoder, decoder and hierarchical skip connection operations, a multi-channel probability map is generated. Finally, the multi-channel probability map is converted into a binary lesion segmentation mask output by the Softmax function.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a comprehensive system encompassing patient information collection, intelligent image analysis, remote monitoring, disease progression risk warning and prognosis prediction, online consultation and eye health education, etc., to systematically and one-stop manage the entire process of patient information from pre-treatment to long-term follow-up after treatment, which helps to improve the management of chronic diseases and the level of personalized medical services for high myopia and pathological myopia.

[0006] The intelligent image analysis module of this invention uses artificial intelligence technology to assist doctors in image analysis and diagnosis, reducing human error and improving the accuracy and efficiency of image screening for early identification of pathological myopia. Specifically, the TL-TransUNet ocular 3D OCT image analysis algorithm involved in this invention has the following main advantages over traditional Unet and Transformer-based image segmentation algorithms: ① Excellent local detail preservation ability: This method uses wavelet downsampling to preserve high-frequency information (preserving fine structures through high-frequency subbands), avoiding the loss of fine structures such as scleral boundaries and split gaps caused by traditional pooling, while Transformer-based segmentation algorithms use self-attention to capture the global picture but may lose local details; ② Strong noise resistance: This method can maintain segmentation accuracy under high noise conditions, which is particularly important for OCT image analysis, while traditional Unet image segmentation algorithms are sensitive to noise and do not require large-scale pre-training like Transformer-based image segmentation algorithms; ③ Higher computational efficiency: The learnable filter parameters involved in this method The number of data points is relatively small, while the traditional Unet image segmentation algorithm involves a large amount of 3D convolution computation, and the attention computation complexity of the Transformer-based image segmentation method increases with the square of the number of voxels; ④ Better morphological robustness: The multi-scale decomposition used in this method is naturally adapted to the morphological variations of high myopia and pathological myopia, and is suitable for handling irregular shapes such as posterior staphyloma, while the traditional Unet algorithm has poor generalization to this type of morphological variation, and the Transformer-based image segmentation method requires a large amount of labeled data; ⑤ Clear interpretability and learnability: The wavelet coefficients involved in this method have clear frequency domain meaning, and the adjustable wavelet filter allows the model to optimize the frequency decomposition for specific pathological features (possessing sustainable learning ability), while the traditional Unet algorithm and the Transformer-based image segmentation method are similar to black boxes.

[0007] The remote monitoring module of this invention can realize remote real-time monitoring and personalized management of patient medication / vision training / glasses insertion and removal, improve patient compliance, and reduce the risk of disease recurrence.

[0008] The disease progression risk warning and prognosis prediction module of this invention helps to identify patients at high risk of progression of pathological myopia-related posterior staphyloma, patients at high risk of progression of retinal schisis to retinal detachment, and patients with poor long-term visual prognosis after receiving anti-VEGF treatment for high myopia mCNV in advance, so as to carry out timely prevention and intervention, adjust treatment strategies to delay disease progression, and realize personalized and precision medicine.

[0009] The online consultation and eye health education module of this invention facilitates patient access to medical care, enhances patients' understanding and management of diseases, and improves patient satisfaction. The internet hospital platform breaks down geographical limitations, enabling high-quality medical resources to reach a wider range of patients and improving the efficiency of medical resource utilization. Attached Figure Description

[0010] Figure 1 This is a system diagram of the AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital, as described in this invention.

[0011] Figure 2 This is a flowchart of the image intelligent analysis unit in the intelligent auxiliary diagnosis module of the present invention. Detailed Implementation

[0012] The terms first, second, third, fourth, etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms include or have any variations thereof, intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] like Figure 1 An AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital, the system comprising: The patient information management module is configured to collect and store patients' 3D-OCT body data and identity and medical history information. Specifically, this includes the patient's detailed basic personal information, current medical history, past medical history (especially optometry and glasses prescription history, ophthalmology and optometry treatment history, surgical history), family history (whether parents have high myopia / high hyperopia / astigmatism, albinism, congenital melanocytosis, RP, color vision abnormalities, etc.), and daily eye habits (primarily long-distance / near-distance eye use, whether driving, fine eye use, etc.). The system interfaces with the information systems of cooperating medical institutions to automatically obtain the patient's past imaging reports (including whole-body CT, MRI, ocular optical biometry, OCT and OCTA, ultra-wide-angle fundus photography, anterior segment photography, anterior segment OCT, corneal topography, ocular surface analysis, visual field, etc.), laboratory reports (blood reports, etc.), gene testing reports, diagnosis and treatment records, etc. The system will structure the information obtained. The BioBERT or PubMedBERT model is used to extract structured feature vectors (specific feature descriptions, such as "posterior staphyloma" and "macular schisis") from structured medical records of high myopia or pathological myopia. Through cross-modal attention mechanism, cross-modal fusion is performed with the corresponding ophthalmic image (such as OCT and ultra-wide-angle fundus photography) features to intelligently learn text-image association and train the model to improve the accuracy of pathological myopia screening. The intelligent diagnostic assistance module includes an intelligent image analysis unit; The image intelligent analysis unit, such as Figure 2 The configuration calls the lightweight automated image segmentation model TL-TransUNet to perform three-dimensional structural analysis on the 3D-OCT volumetric data of the eye. The image segmentation model decomposes the original 3D-OCT volumetric data of the eye using wavelet transform operators to extract multi-scale frequency domain features, and merges and concatenates the multi-scale frequency domain features with the original 3D-OCT volumetric data of the eye in the channel dimension to generate a multi-channel feature cube. The multi-channel feature cube is input into the Unet algorithm network unit, and after encoder, decoder and hierarchical skip connection operations, a multi-channel probability map is generated. Finally, the multi-channel probability map is converted into a binary lesion segmentation mask output by the Softmax function.

[0014] The wavelet transform operator in the image intelligent analysis unit uses a filter bank to perform wavelet transform, including a trainable low-pass filter h0, a high-pass filter h1, and synthetic filters f0 and f1; the mathematical mapping relationship between the high-pass filter h1 and the low-pass filter h0 is as follows: ; The mathematical mapping relationship between the synthesis filters f0 and f1 and the low-pass filter h0 is as follows: f0(n) = h0*(N-1-n) f1(n) = -(-1)n * h0(n); Where N represents the spatial length parameter of the filter, and n represents the sequence index position parameter of the filter; the multi-scale frequency domain features extracted by decomposition include a low-frequency sub-band, three unidirectional high-frequency sub-bands, three bidirectional high-frequency sub-bands, and one omnidirectional high-frequency sub-band.

[0015] The hierarchical skip connections in the Unet algorithm network unit and the logical steps of feature fusion and reconstruction performed by the decoder include: The hierarchical skip connection directly extracts and isolates wavelet subband feature parameters of the corresponding scale from the digital signal of the original eye 3D-OCT volume data. The wavelet subband feature parameters are combined and spliced ​​with the original high-resolution feature parameters from the corresponding network layer of the encoder without downsampling processing to generate a skip connection fusion feature cluster. The decoder performs an upsampling resolution recovery operator operation on the high-level semantic feature parameters extracted internally to generate an initial coarse spatial mapping feature layer. After the coarse spatial mapping feature layer and the skip connection fusion feature cluster are concatenated across channels, they are passed to the next level convolutional kernel sequence to reconstruct and output the multi-channel probability map.

[0016] The image intelligent analysis unit is configured to extract and quantify the feature parameters of scleral staphyloma based on the lesion segmentation mask, and the calculation logic it executes is as follows: Within the two-dimensional projection coordinate set Ω of the enface of the posterior staphyloma region, calculate the actual Z-axis coordinate parameter Z of the scleral surface corresponding to each projection coordinate point. surface (x,y) and the Z-axis coordinate parameter Z of the reference sphere fitted based on the normal scleral region around the optic disc. sphere The absolute vertical difference between (x, y) is extracted by the sorting network layer, and the maximum floating-point number in the set of all absolute vertical differences is assigned to the maximum indentation depth parameter variable. Within the two-dimensional projection coordinate set Ω, for each projection coordinate point, calculate the product of the corresponding absolute vertical difference, the plane differential area element dxdy, and the voxel spacing constant δz in the Z-axis spatial direction. Perform a three-dimensional spatial accumulation integration operation on all product terms and assign the value to the total volume parameter variable of the posterior staphyloma.

[0017] The image intelligent analysis unit is configured to extract and quantify macular schisis feature parameters based on the lesion segmentation mask, and the calculation logic it executes is as follows: A three-dimensional spatial coordinate set corresponding to all voxel points in the lesion segmentation mask is constructed. The individual volume constants associated with all voxels in this set are iterated and summed, and the output is assigned to the total split volume parameter variable. The local split cavity volume values ​​contained in the three-dimensional spatial coordinate set within the eight independently divided single-layer retinal structures are calculated according to the depth traversal sequence. For each single-layer retinal structure assigned an identifier index parameter i, the ratio of the local split cavity volume value to the total volume value of the single-layer retinal structure itself is calculated. An extreme value comparator is run to retrieve the corresponding identifier index parameter i when this ratio becomes extremely large, and this is output as the main split layer location identifier field.

[0018] The image intelligent analysis unit is configured to extract and quantify the paradiscal atrophy arc area parameters based on the lesion segmentation mask, and the calculation logic executed is as follows: Extract the first set of projected closed coordinates indicating the Bruch membrane missing region from the corresponding two-dimensional projection mapping surface. Retrieve and run the accumulator to perform area integration calculation on the preset area micro-elements corresponding to all pixels in the first set of projected closed coordinates, and output the area parameter value of the γ region. Extract the second set of projected closed coordinates indicating the choroid atrophy region from the two-dimensional projection mapping surface. Retrieve and run the accumulator to perform area integration calculation on the preset area micro-elements corresponding to all pixels in the second set of projected closed coordinates, and output the area parameter value of the δ region.

[0019] Specifically: The lightweight, automated image segmentation model TL-TransUNet was used to analyze the raw 3D OCT images of the eye. Specifically, the 3D-OCT data was first augmented using an adaptive wavelet transform (Equations 1-3) to obtain a multi-channel feature cube (Equation 4). This multi-scale frequency domain feature was then used as input by the Unet algorithm (containing encoder, decoder, and skip connections, Equations 5-7) to interpret and translate it, outputting a multi-channel probability map of the same size as the input 3D-OCT. The latter was then converted into the final segmentation probability map using the Softmax function (Equation 8). The clinical reliability of the model segmentation was evaluated by measuring the consistency between the model's segmentation results and expert manual annotations using the Dice coefficient (Equation 9).

[0020] Formula 1: f0(n)=h0 * (N-1-n) Formula 2: h1(n) = (-1) n * h0(N-1-n) Formula 3: f1(n)=-(-1) n * h0(n) Wherein, h0 is a trainable low-pass filter that can adaptively capture the overall concavity morphology of scleral staphyloma or the overall curvature of the macular region after high myopia; h1 is a high-pass filter responsible for capturing high-frequency information such as minute separations between retinal layers (splitting gaps) and the termination point of the RPE layer (atrophic arc boundary); f0 and f1 are synthetic filters that ensure that the features after learnable wavelet downsampling can be perfectly reconstructed back to the original resolution, achieving pixel-level precise positioning; N represents the filter length; n represents the filter index position.

[0021] Formula 4: X input = Concat[ f(x,y,z), cLLL, cLLH, cLHL, cHLL, cLHH, cHLH,cHHL, cHHH] Concat represents the feature channel splicing operation, which merges the original OCT data with the eight sub-bands decomposed by 3D wavelet transform into an enhanced multi-channel data cube in the channel dimension, serving as the input for the subsequent U-Net; f(x,y,z) represents the original 3D-OCT volume data, containing all the 3D anatomical information from the internal limiting membrane (ILM) to the sclera in the 3D-OCT image of the examined eye; cLLL is the low-frequency sub-band, representing low-frequency information such as the overall 3D morphology of the sclera and retina; cLLH, cLHL, and cHLL are unidirectional high-frequency sub-bands, capturing rapid changes along a specific direction; cLHH, cHLH, and cHHL capture the features of 3D-OCT volume data undergoing high-frequency changes simultaneously in two specific directions; cHHH is the omnidirectional high-frequency sub-band, capturing isotropic microstructures such as photoreceptor chimeras and microcystic edema.

[0022] Formula 5: Here, Down represents downsampling, which is usually replaced by a learnable wavelet downsampling module instead of traditional pooling; Conv represents a three-dimensional convolutional layer that automatically extracts high-order semantic features from the input fused features through learnable convolutional kernels. This represents preliminary, coarse spatial information, extracting local features such as subtle undulations and minute split cavities in the ILM (Intense Least Motion) model. Representing high-level semantic features, it is... The abstraction and generalization of these concepts allow for the analysis of global semantics, such as the concept of depression in posterior staphyloma.

[0023] Formula 6: Here, Up represents upsampling, which gradually restores the size of the feature map to the original input size; Concat represents feature concatenation, which combines the features from the previous layer of the decoder. Skip connection features from the encoder The layers are stitched together to ensure that when reconstructing the segmentation mask, both the semantic information of deep features and the position and boundary information of shallow features can be utilized. The preliminary, coarse spatial information represents the deep semantic features. It was obtained by upsampling based on the data. This represents skip connection features, i.e., high-resolution spatial information directly transmitted from the corresponding layer of the encoder, preserving... The edges, textures, and other details in the image, without the abstraction loss caused by downsampling, can be fed back the original boundaries organized at a precise location when the decoder restores the resolution; This represents a precise segmentation image that incorporates details, and is... Skip connection features from the encoder The result is obtained by fusion and then convolution.

[0024] Formula 7: in, It is the output from the l-th layer of the encoder, without downsampling, representing the high-resolution features that the model automatically learns from the data; It is a wavelet subband (such as cHLH) feature that is independent of model training, directly extracted from the original OCT signal and directly injected into the skip connection. It can provide data-independent, general directional edge information, which is especially crucial for the segmentation of irregular lesions (such as the irregular boundary of atrophic arc). As and The fusion and stitching results can simultaneously provide lesion-specific details and general anatomical constraints, guiding the decoder to perform accurate and robust segmentation.

[0025] Formula 8: in, Representing the segmentation mask, for each voxel, the model outputs the probability that it belongs to different categories such as "background," "posterior staphyloma," "retinal layer," or "splitting cavity." Binarization (taking the category with the highest probability) can evaluate specific fundus changes in pathological myopia, including posterior staphyloma, peripapillary atrophy arc, and macular schisis.

[0026] Formula 9: In this model, TP represents a true positive, indicating that the model correctly segmented the lesion voxels; FP represents a false positive, indicating that the model incorrectly labeled normal voxels as lesions; and FN represents a false negative, indicating that the model missed lesion voxels. The closer the Dice coefficient is to 1 (or 100%), the more consistent the model's automated segmentation results are with the judgment of ophthalmologists.

[0027] Based on the automatic analysis of the above-mentioned 3D-OCT images, the system automatically screens for the presence of important characteristic changes in pathological myopia and quantifies their degree, providing important data support for the development of personalized treatment plans and dynamic tracking of the condition.

[0028] (1) Posterior staphyloma ① The maximum depth of posterior staphyloma depression and its annualized growth rate and annualized growth rate Formula 10-13: Among them, D max Ω represents the maximum vertical indentation distance of the posterior staphyloma (unit: mm); Ω represents the set of two-dimensional projected coordinates of the posterior staphyloma region. The Z-coordinate (depth) of the actual scleral surface at coordinates (x, y) represents the actual scleral position measured by OCT. Z is the Z-coordinate of the reference sphere fitted at the same coordinate point, representing the theoretical normal sclera position (an ideal sphere fitted based on the center of the optic disc); R is the radius of the sphere fitted based on the normal sclera region around the optic disc (unit: mm). The annualized growth rate of the maximum indentation depth of posterior staphyloma (unit: mm / year). The annualized growth rate of the maximum indentation depth of posterior staphyloma; Represents the maximum vertical indentation distance of the posterior staphyloma calculated in the current image (unit: mm); Δt represents the maximum vertical indentation distance of the posterior staphyloma calculated from the baseline image (unit: mm); Δt represents the time interval between the current examination and the baseline (unit: year).

[0029] ② The volume of posterior staphyloma and its annualized growth rate and annualized growth rate Formula 14-16: Among them, V staph Total volume of posterior staphyloma (unit: mm) 3 ); Ω represents the set of two-dimensional projected coordinates of the posterior staphyloma region; dx represents the perpendicular distance between the reference sphere and the actual concave surface at point (x,y); dx and dy represent the differential area elements with respect to the X and Y axes, with each element corresponding to a pixel in the OCT image; δz represents the voxel spacing in the Z-axis direction (unit: mm / voxel). Annualized growth rate of posterior staphyloma volume (unit: mm) 3 / Year); This represents the annualized growth rate of posterior staphyloma volume; This represents the posterior staphyloma volume calculated from the current image (unit: mm). 3) The volume of posterior staphyloma calculated from baseline images (unit: mm) 3 ); Δt represents the time interval (in years) between the current check and the baseline.

[0030] ③ Posterior staphyloma asymmetry index Formula 17-18: AI stands for Posterior Staphyloma Asymmetry Index, which measures the dispersion of the steepness of the staphyloma surface. It represents the maximum rate of change of the scleral surface in the horizontal (X) and vertical (Y) directions, i.e., the slope or steepness of that point; This represents the maximum value of the gradient within the staphyloma region. This represents the minimum gradient within the staphyloma region. This represents the average gradient within the staphyloma region.

[0031] (2) Para-optic disc atrophy arc ① Area of ​​the γ region and its annualized growth rate and annualized growth rate Formula 19-21: Among them, A γ Represents the area of ​​the γ region (unit: mm) 2 ), that is, the total surface area of ​​the Bruch membrane missing region; Ω γ The enface projection region represents the γ region; dA represents the area micro-element, corresponding to the actual area represented by one pixel (unit: mm). 2 ); ΔA γ Annualized growth rate of the γ region area (unit: mm) 2 / year); ΔA γ % represents the annualized growth rate of the γ region area; A γ, t This represents the calculated area of ​​the γ region in the current image (unit: mm). 2 A γ, t0 The area of ​​the γ region calculated from the baseline image (unit: mm) 2 ); Δt represents the time interval (in years) between the current check and the baseline.

[0032] ②The area of ​​the δ region and its annualized growth rate and annualized growth rate Formula 22-24: Among them, A δ Represents the area of ​​the δ region (unit: mm) 2 ), that is, the total surface area of ​​the choroidal atrophy area; Ω δ The enface projection region represents the delta region; dA represents the area micro-element, corresponding to the actual area represented by one pixel (unit: mm).2 ); ΔA δ The annualized growth rate of the area representing the δ region (unit: mm) 2 / year); ΔA δ % represents the annualized growth rate of the δ region area; A δ, t This represents the calculated area of ​​the δ region in the current image (unit: mm). 2 A δ, t0 The area of ​​the δ region calculated from the baseline image (unit: mm) 2 ); Δt represents the time interval (in years) between the current check and the baseline.

[0033] ③ Shrinkage arc index Formula 25: PPAI stands for the atrophy arc index, which measures the ratio of the atrophy arc to the optic disc size; A γ Represents the area of ​​the γ region (unit: mm) 2 ), that is, the total surface area of ​​the Bruch membrane missing region; A δ Represents the area of ​​the δ region (unit: mm) 2 ), that is, the total surface area of ​​the choroidal atrophy area; A disc Represents the total surface area of ​​the display screen (unit: mm) 2 ).

[0034] (3) Yellow spot splitting ①Location of the split layer Formula 26: Where Layer represents the primary retinal layer where the schistomeric cavity is located, and the output value is: outer layer schistomeric, inner layer schistomeric, full-layer schistomeric, etc.; arg max i This represents the index i that maximizes the expression, and returns the level number that maximizes the ratio within the parentheses; i represents the retinal level index, i=1: ILM-NFL gap, i=2: NFL-GCL, i=3: GCL-IPL, i=4: IPL-INL, i=5: INL-OPL, i=6: OPL-ONL, i=7: ONL-ELM, i=8: ELM-RPE; V cavity This represents the total volume of the split cavity, the sum of the volumes of all split cavity voxels divided from the model (unit: mm). 3 );V layer,i This represents the volume of the i-th retinal layer (unit: mm). 3 ).

[0035] ② Total split volume and its annualized growth rate and annualized growth rate Formula 27-29: Among them, Vschisis The total volume of the split is the sum of the three-dimensional volumetric elements of all split cavities (unit: mm). 3 ); Ω cavity dv represents the three-dimensional spatial region occupied by the split cavity, and is the set of all split cavity voxels (x, y, z) segmented from the model; dv is a volume element, corresponding to the actual volume represented by one voxel (unit: mm). 3 ); ΔV schisis Annualized growth rate of total split volume (unit: mm) 3 / year); ΔV schisis % represents the annualized growth rate of the total split volume; V schisis,t Represents the total split volume calculated from the current image (unit: mm). 3 ); V schisis,t0 Represents the total split volume calculated from the baseline image (unit: mm). 3 ); Δt represents the time interval (in years) between the current check and the baseline.

[0036] The remote monitoring module includes a smart device data sharing unit, a remote medical order execution monitoring unit, and a follow-up reminder unit; The intelligent device data sharing unit is characterized in that: the system has an open Internet of Things access function, supports patient authorization to connect to a variety of intelligent devices (such as smartphone apps, smart bracelets, etc.) and home health management devices (such as smart blood glucose meters, blood pressure monitors, etc.), and collects data such as home self-test vision and refractive error, blood glucose, blood pressure, exercise and rest in real time or at regular intervals, and uploads them to the system after automatic encryption; The remote medical order execution monitoring unit is characterized by the following: the system sends daily reminders (including medication / vision training / glasses removal / wearing) to the patient's mobile phone and other bound smart devices (such as smart bracelets) according to pre-given personalized medical orders, along with precautions (such as some medications need to be refrigerated, avoid contact between the bottle tip and the eyes when applying eye drops, different eye drops need to be applied 5 minutes apart, some eye drops need to be gradually reduced in dosage, avoid using eye drops beyond their expiration date, and strictly follow the doctor's orders). The patient completes the feedback loop for the corresponding medical order (medication / training / glasses removal / wearing) by clicking confirm. The system continuously monitors the patient's daily medical order completion during the follow-up period, recording a non-executed medical order record for each missing feedback message for the corresponding medical order execution. The aforementioned follow-up reminder unit is characterized by: the system continuously monitoring whether short-term fluctuations in the patient's visual acuity and refractive status, measured by the patient using a smart device at home, exceed set thresholds, and whether the number of times medical orders have not been followed exceeds an acceptable threshold. If abnormal fluctuations in visual acuity or refractive status are detected in a short period, or if there is serious non-compliance with medical orders for medication / training / glasses removal / wearing, or if the agreed follow-up date has been exceeded, the system automatically triggers an alert, reminding the doctor to conduct an online assessment, and pushing a warning message based on the assessment results regarding whether an immediate offline follow-up examination is necessary. Simultaneously, updated personalized medical orders are pushed again, and guidance and demonstrations on correct medication use, vision training, and glasses removal / wearing are provided to the patient through videos, animations, and other means.

[0037] Disease progression risk warning and prognosis prediction module Risk and protective factors analysis: The system aggregates patients' demographic, sociological, biological, and imaging data: age, sex, genetic information, past medical history, detailed eye habits, living environment, biomarkers, and longitudinal multimodal ophthalmic imaging data.

[0038] Principal component analysis (PCA) was used to calculate the weights of multiple influencing factors in the development of high myopia and pathological myopia. The weights of variables such as outdoor activity time, electronic product usage time, near-field reading and work time, and sleep duration in the "environmental factors" were calculated. The relative contributions of "genetic factors" and "behavioral factors" to the development of the disease were quantified. PCA analysis was performed on the changes in each factor in the longitudinal follow-up data to discover the typical trajectory of disease progression.

[0039] The Apriori and FP-Growth algorithms were used to identify combinations of risk factors that jointly promote the progression to high myopia or pathological myopia within 3 or 5 years, and to discover combinations of protective factors that jointly reduce the risk of progression to high myopia or pathological myopia within 3 or 5 years.

[0040] Risk and prognosis prediction model construction: Based on a massive amount of de-identified cohort data of patients with and without pathological myopia accumulated over time, we trained and validated predictive models for the risk of pathological myopia development, complication risk, and treatment prognosis.

[0041] 1. Decision tree models such as random forest or XGBoost: Analyze the weights of different influencing factors and high-dimensional features such as certain nonlinear relationships and threshold effects to quantify the risk probability of the occurrence and development of high myopia or pathological myopia.

[0042] 2. Temporal Transformer Model: Employing a temporal self-attention mechanism to process high-quality, multimodal follow-up data of high myopia including non-uniform follow-up data, a risk prediction model for rapid progression of posterior staphyloma in high myopia was constructed to identify patients with rapid progression early, providing a basis for determining the timing of posterior scleral reinforcement surgery. A risk prediction model for the progression of retinal schisis to retinal detachment in high myopia was also constructed to identify high-risk individuals early, providing a basis for determining the timing of vitrectomy surgery. Finally, a long-term visual prognosis grading prediction model was constructed for patients with high myopia mCNV after receiving anti-VEGF treatment, guiding individualized adjustments to dosing regimens (combinations of different drugs and number of injections, etc.).

[0043] Early warning and prevention recommendations: When the system predicts that a patient is in a medium-to-high-risk state, it automatically generates an early warning message, which is automatically pushed to the attending / contracted doctor and the patient through the system platform and smart devices. Based on the identified combinations of risk factors, personalized prevention and treatment recommendations are automatically generated and pushed to users. ① Adjust your lifestyle: such as controlling your BMI, reducing prolonged close-range, high-intensity eye use, and increasing the interval between looking into the distance after a period of close-range eye use (20-20-20 rule). ② Dietary recommendations: Avoid high-sugar and high-fat diets, increase the intake of fish oil, lutein, etc., and maintain a balanced diet; ③ Exercise recommendations: Avoid strenuous activities and extreme sports (bungee jumping, diving, trampoline, boxing, prolonged handstands, etc.), enhance protection against sports injuries (especially badminton, tennis, volleyball, basketball, football, etc.), and recommend gentle exercises (walking, swimming, jogging, Tai Chi, Baduanjin, etc.). ④ Regular follow-up: It is recommended that high-risk individuals undergo regular imaging monitoring such as OCT and fundus photography. ⑤ Recommendations for adjusting intervention methods: It is recommended that high-risk patients with poor long-term visual prognosis under the original treatment plan seek medical treatment offline in a timely manner and follow the doctor's advice to adjust the individualized treatment plan.

[0044] The online consultation and eye health education module includes an online consultation unit and an eye health education unit. The online consultation unit is characterized in that: The system establishes a multimodal instant communication mechanism with end-to-end encrypted transmission, supporting various interactive forms such as text, voice calls, and high-fidelity real-time video, to meet patients' needs for online diagnosis and treatment in the Internet environment, and to ensure privacy and communication continuity during the consultation process. Based on an intelligent fusion mechanism, the system automatically presents the patient's structured full-cycle health record on the doctor's consultation interface, and simultaneously loads the pre-analyzed imaging report from the intelligent auxiliary diagnosis and treatment module. Combining the ophthalmological disease knowledge graph and the latest clinical guidelines, it recommends a ranked list of possible diagnoses (sorted from high to low probability), assisting doctors in completing a comprehensive assessment of the patient's condition in the shortest possible time, improving the accuracy and efficiency of remote diagnosis, and thus developing more targeted treatment intervention plans.

[0045] The eye health science popularization unit is characterized by: Knowledge Base Construction: Establish a structured multimodal knowledge base, which stores knowledge about the entire course of ophthalmic diseases using node-edge relationships. The entire course includes at least seven nodes: prevention, screening, diagnosis, treatment, rehabilitation, nutrition, and psychological adjustment. Each knowledge node in the knowledge base is associated with at least one digital carrier, which includes popular science articles, short videos, animated demonstrations, or infographics. User profile creation: Collect individualized feature data of target users, including static and dynamic attributes; the static attributes include disease diagnosis tags, and the dynamic attributes include current health status parameters, rehabilitation stage tags, and historical interaction behavior data; Recommendation engine construction: Construct a hybrid recommendation model, which includes at least a collaborative filtering unit and a content similarity calculation unit; the collaborative filtering unit is used to identify homogeneous user groups with similar characteristics to the target user, and the content similarity calculation unit is used to calculate the semantic matching degree between each knowledge node in the knowledge base and the feature data of the target user; Personalized push: The hybrid recommendation model generates a recommendation score based on the output results of the collaborative filtering unit and the content similarity calculation unit; the knowledge nodes in the knowledge base are sorted in descending order according to the recommendation score, and the digital carrier corresponding to at least one of the top-ranked knowledge nodes in the sorting results is selected and actively pushed to the target user through the user terminal.

[0046] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital, characterized in that: The system includes: The patient information management module is configured to collect and store patients' ocular 3D-OCT body data and identity and medical history characteristics. The intelligent diagnostic assistance module includes an intelligent image analysis unit; The image intelligent analysis unit is configured to call the automated image segmentation model TL-TransUNet to perform three-dimensional structural analysis on the 3D-OCT volume data of the eye. The image segmentation model decomposes the original 3D-OCT volume data of the eye using wavelet transform operators to extract multi-scale frequency domain features, and merges and concatenates the multi-scale frequency domain features with the original 3D-OCT volume data of the eye in the channel dimension to generate a multi-channel feature cube. The multi-channel feature cube is input into the Unet algorithm network unit, and after encoder, decoder and hierarchical skip connection operations, a multi-channel probability map is generated. Finally, the multi-channel probability map is converted into a binary lesion segmentation mask output by the Softmax function.

2. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The wavelet transform operator in the image intelligent analysis unit uses a filter bank to perform wavelet transform, including a trainable low-pass filter h0, a high-pass filter h1, and synthetic filters f0 and f1; the mathematical mapping relationship between the high-pass filter h1 and the low-pass filter h0 is as follows: ； The mathematical mapping relationship between the synthesis filters f0 and f1 and the low-pass filter h0 is as follows: f0(n) = h0*(N-1-n) ； Where N represents the spatial length parameter of the filter, and n represents the sequence index position parameter of the filter; the multi-scale frequency domain features extracted by decomposition include a low-frequency sub-band, three unidirectional high-frequency sub-bands, three bidirectional high-frequency sub-bands, and one omnidirectional high-frequency sub-band.

3. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The hierarchical skip connections in the Unet algorithm network unit and the logical steps of feature fusion and reconstruction performed by the decoder include: The hierarchical skip connection directly extracts and isolates wavelet subband feature parameters of the corresponding scale from the digital signal of the original eye 3D-OCT volume data. The wavelet subband feature parameters are combined and spliced ​​with the original high-resolution feature parameters from the corresponding network layer of the encoder without downsampling processing to generate a skip connection fusion feature cluster. The decoder performs an upsampling resolution recovery operator operation on the high-level semantic feature parameters extracted internally to generate an initial coarse spatial mapping feature layer. After the coarse spatial mapping feature layer and the skip connection fusion feature cluster are concatenated across channels, they are passed to the next level convolutional kernel sequence to reconstruct and output the multi-channel probability map.

4. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The image intelligent analysis unit is configured to extract and quantify the feature parameters of scleral staphyloma based on the lesion segmentation mask, and the calculation logic it executes is as follows: Within the two-dimensional projection coordinate set Ω of the enface of the posterior staphyloma region, calculate the actual Z-axis coordinate parameter Z of the scleral surface corresponding to each projection coordinate point. surface (x,y) and the Z-axis coordinate parameter Z of the reference sphere fitted based on the normal scleral region around the optic disc. sphere The absolute vertical difference between (x, y) is extracted by the sorting network layer, and the maximum floating-point number in the set of all absolute vertical differences is assigned to the maximum indentation depth parameter variable. Within the two-dimensional projection coordinate set Ω, for each projection coordinate point, calculate the product of the corresponding absolute vertical difference, the plane differential area element dxdy, and the voxel spacing constant δz in the Z-axis spatial direction. Perform a three-dimensional spatial accumulation integration operation on all product terms and assign the value to the total volume parameter variable of the posterior staphyloma.

5. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The image intelligent analysis unit is configured to extract and quantify macular schisis feature parameters based on the lesion segmentation mask, and the calculation logic it executes is as follows: Construct a three-dimensional spatial coordinate set corresponding to all voxel points in the lesion segmentation mask, perform a traversal and summation operation on the individual volume constants associated with all voxels in the set, and assign the output result to the total split volume parameter variable; calculate the local split cavity volume values ​​contained in the three-dimensional spatial coordinate set in the 8 independently divided single-layer retinal structures according to the depth traversal sequence. For each of the single-layer retinal structures assigned the identifier index parameter i, calculate the ratio of the local slit cavity volume value to the total volume value of the single-layer retinal structure itself; Run the extreme value comparator to retrieve the corresponding identifier index parameter i when the ratio number reaches an extremely large amount of data, and output it as the main splitting layer location identifier field.

6. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The image intelligent analysis unit is configured to extract and quantify the paradiscal atrophy arc area parameters based on the lesion segmentation mask, and the calculation logic executed is as follows: Extract the first projected closed coordinate set indicating the Bruch film missing region on the corresponding two-dimensional projection mapping surface, call up and run the accumulator to perform area integration calculation on the preset area micro-elements corresponding to all pixels in the first projected closed coordinate set, and output the γ region area parameter value. Extract the second projection closed coordinate set indicating the choroidal atrophy area from the two-dimensional projection mapping surface, retrieve and run the accumulator to perform area integration on the preset area micro-elements corresponding to all pixels in the second projection closed coordinate set, and output the area parameter value of the δ region.

7. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The system also includes a remote monitoring module for: It accepts home-based self-test data on visual acuity and refractive error, blood sugar, blood pressure, and exercise and lifestyle uploaded from smart devices and home health management devices linked to the patient. Based on pre-given personalized medical orders, the system sends daily reminders and precautions to patients' mobile phones and other linked smart devices. These precautions include the need to refrigerate some medications, avoid contact between the bottle opening and the eyes when applying eye drops, the need to wait 5 minutes between different eye drops, the need to gradually reduce the dosage of some eye drops, the need to avoid using eye drops beyond their expiration date, and the need to strictly follow the doctor's instructions when using medications. The system continuously monitors whether short-term fluctuations in the patient's vision and refractive status, measured by the smart device at home, exceed the set threshold, and whether the number of times medical orders are not followed exceeds the acceptable threshold. When abnormal fluctuations in vision or refractive status are detected in a short period of time, or when there is serious non-compliance with medical orders for medication / training / glasses removal, or when the scheduled follow-up date is exceeded, an alert is automatically triggered and the responsible doctor is notified to conduct an online assessment to determine whether an immediate offline medical follow-up examination is necessary.

8. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The system also includes a disease progression risk early warning and prognosis prediction module, used for: This study aggregates patient demographic, sociological, biological, and imaging data, including age, sex, genetic information, medical history, detailed eye habits, living environment, biomarkers, and longitudinal multimodal ophthalmic imaging data. Principal component analysis (PCA) is used to calculate the weights of multiple influencing factors in the development of high myopia and pathological myopia, as well as typical disease progression trajectories. Apriori and FP-Growth algorithms are employed to identify combinations of risk factors and protective factors that contribute to progression to high myopia or pathological myopia within 3 or 5 years. Based on historical de-identified cohort data of patients with pathological and non-pathological myopia, predictive models for the risk of pathological myopia development, complications, and treatment prognosis were trained and validated. These models included random forest or XGBoost decision tree models and temporal Transformer models, used to quantify the probability of the occurrence and development of high myopia or pathological myopia, the probability of rapid progression of scleral staphyloma after high myopia, the probability of retinal detachment from retinal schisis in high myopia, and the probability of poor long-term visual prognosis in patients with high myopia mCNV after receiving anti-VEGF treatment. When the probability of the aforementioned risks reaches the medium-to-high risk threshold, the system automatically generates an early warning message and pushes it to the responsible doctor and the patient, and generates personalized prevention and treatment recommendations based on the main risk factors.

9. The AI-assisted diagnosis and treatment system for refractive errors based on an internet hospital as described in claim 1, characterized in that, The system also includes an online consultation and eye health education module, used for: A multimodal instant communication mechanism with end-to-end encrypted transmission is established, supporting text, voice calls, and high-fidelity real-time video online consultations. The mechanism automatically presents patients' structured full-cycle health records and pre-analyzed imaging reports from the intelligent auxiliary diagnosis and treatment module to online doctors. It also combines an ophthalmology disease knowledge graph and the latest clinical guidelines to synchronously recommend a ranked list of possible diagnoses, sorted from highest to lowest probability. Construct and maintain a structured knowledge base for ophthalmic diseases, covering popular science content related to disease prevention, screening, diagnosis, treatment, rehabilitation, nutrition and psychological adjustment. Each knowledge node is associated with at least one digital carrier, including popular science articles, short videos, animated demonstrations or infographics. User profiles are built by collecting individualized feature data of target users. The feature data includes static attributes and dynamic attributes. The static attributes include disease diagnosis tags, and the dynamic attributes include current health status parameters, rehabilitation stage tags, and historical interaction behavior data. A hybrid recommendation model is constructed, comprising at least a collaborative filtering unit and a content similarity calculation unit; the collaborative filtering unit is used to identify homogeneous user groups with similar characteristics to the target user, and the content similarity calculation unit is used to calculate the semantic matching degree between each knowledge node in the knowledge base and the target user's feature data; The hybrid recommendation model generates a recommendation score based on the outputs of the collaborative filtering unit and the content similarity calculation unit; it sorts the knowledge nodes in the knowledge base in descending order according to the recommendation score, selects at least one digital carrier corresponding to the top-ranked knowledge node in the sorting results, and actively pushes it to the target user through the user terminal.

Images