Search

CN-121983328-A - Ophthalmic prognosis vision simulation system and training method thereof

CN121983328ACN 121983328 ACN121983328 ACN 121983328ACN-121983328-A

Abstract

The invention discloses an ophthalmology prognosis vision simulation system and a training method thereof, wherein the system comprises a feature extraction module, an evolution analysis module and a forward deduction module, wherein the feature extraction module extracts biomarkers based on OCT images and historical diagnosis and treatment records of a user, the evolution analysis module carries out course trend analysis based on the biomarkers and provides a processing scheme, and the forward deduction module carries out forward deduction of counter facts based on the processing scheme to obtain a predicted anatomical image of the fundus of the user at a future moment. The system can provide intuitive long-range diagnosis and treatment planning basis for doctors and patients, further ensure the safety of clinical decisions, and simultaneously provide a low-cost simulated diagnosis and treatment error-testing environment for the doctors to assist in optimizing a clinical test scheme.

Inventors

  • Song Diping
  • JIA HUIXUN
  • QU YANLIN
  • YAN JIANGTAO
  • SUN XIAODONG

Assignees

  • 上海人工智能创新中心
  • 上海市第一人民医院

Dates

Publication Date
20260505
Application Date
20260409

Claims (10)

  1. 1. An ophthalmic prognostic vision simulation system, comprising: a feature extraction module configured to extract biomarkers based on OCT images and historic diagnosis and treatment records of a user; an evolution analysis module configured to perform course trend analysis based on the biomarkers and propose a treatment regimen; A forward deduction module configured to perform a forward deduction of the counterfactual based on the processing scheme, resulting in a predicted anatomical image of the fundus of the user at a future time.
  2. 2. The ophthalmic prognostic vision simulation system according to claim 1, wherein the biomarker includes intraretinal fluid accumulation, subretinal fluid accumulation of the user.
  3. 3. The ophthalmic prognostic vision simulation system according to claim 1, wherein the evolution analysis module is further configured to convert the treatment plan into natural language instructions and send to the forward deduction module.
  4. 4. The ophthalmic prognostic vision simulation system according to claim 1, wherein the evolution analysis module is further configured to: and determining the risk level of the treatment scheme based on the predicted anatomical image deduced by the forward deduction module, and selecting and/or correcting the treatment scheme based on the risk level.
  5. 5. The ophthalmic prognostic vision simulation system according to claim 1, wherein the feature extraction module includes a multi-modal capable large model, or a reinforcement learning decision network; the evolution analysis module includes a large model with multi-modal capability, or a reinforcement learning decision network.
  6. 6. The ophthalmic prognostic vision simulation system according to claim 1, wherein the forward deduction module includes a diffusion model, or a generation countermeasure network, or a variational self-encoder, or a video generation network.
  7. 7. A method of training an ophthalmic prognostic vision simulation system according to any one of claims 1 to 6, comprising: Based on a pre-constructed training data set, the evolution analysis module internalizes the reasoning logic through a thinking chain technology; Adopting GRPO algorithm, introducing multidimensional rewarding function, training the feature extraction module and the evolution analysis module to strengthen multi-granularity logic alignment; based on the instruction-guided diffusion model, the forward deduction module is trained to learn the pharmacodynamic response.
  8. 8. The training method of claim 7, wherein the multidimensional rewards function comprises a format specification, a marker recall, and a logical consistency.
  9. 9. The training method of claim 7, wherein the constructing of the training data set comprises: scanning the key slice of the OCT through a multi-modal large language model based on hierarchical semantic synthesis, and primarily extracting a candidate biomarker set; Filtering the model illusion by clinical logic of the candidate biomarker set; And carrying out track modeling on the filtered candidate biomarker set to generate dense trend labels.
  10. 10. The training method of claim 9, wherein filtering model illusions by clinical logic of the candidate biomarker set comprises: reversely pushing active focus necessarily existing in the OCT image based on actual treatment behaviors to reject false negative samples; and eliminating abnormal samples with complete mutation or rapid quantity change of the feature set in a short time by using a sequential rationality gate and utilizing a continuity principle of pathological evolution.

Description

Ophthalmic prognosis vision simulation system and training method thereof Technical Field The invention relates to the technical field of artificial intelligence, in particular to an ophthalmic prognosis vision simulation system and a training method thereof. Background Neovascular eye diseases, such as wet age-related macular degeneration, have irreversible damage to blindness, and if not intervened in time, blood and exudates damage retinal photoreceptor cells, leading to rapid decrease in vision in a short time, and often irreversible central vision loss. Even if leakage is controlled for a short period, repetitive motion, structural damage accumulation, such as fibrosis, etc., may occur for a long period. In addition, the number of groups of the aged and diabetics is large, and long-term follow-up and repeated treatment are needed, so that huge social and economic burdens are brought. Anti-vascular endothelial growth factor (anti-VEGF) drugs play an important role in the treatment of various neovascular eye diseases. However, anti-VEGF treatment has a large individual response variability, and long-range management is a Partially Observable Markov Decision Process (POMDP), and there is a discrepancy between long-term, frequent frequency of treatment and compliance, and long-term treatment has potential side effects, so doctors need to trade-off between "timely intervention" and "avoidance of excessive treatment. However, the existing clinical treatment schemes often depend on the experience of doctors, which results in different routes and uneven fates. Aiming at the problem, AI technology assisted treatment can be introduced, but the existing ophthalmic AI technology mainly stays at two separate layers, namely a traditional discriminant model, such as a classification network, which can only perform binary judgment on static images and lacks timing memory and generating capability, and a general medical multi-mode large language model, which can generate text suggestions but belongs to an open-loop reasoning and lacks a physical verification mechanism of a visual layer. Both of these two technical paths cannot construct a closed loop of "hypothesis-simulation-correction", resulting in a lack of intuitive basis and safety assessment for long-range diagnosis and treatment decisions. In particular, existing ophthalmic AI decision systems lack anatomical verification and are low in safety. The traditional medical visual model only predicts whether injection is needed according to OCT images, historical treatment information cannot be effectively utilized for modeling, and the training test process is a black box test and is difficult to explain. The existing multi-mode large language model adopts open-loop text reasoning, and whether treatment advice accords with pathological evolution rules cannot be verified through visual simulation, so that the problem of high risk of hallucination and clinical application is easily generated. Furthermore, there is currently a lack of specialized multi-modal large language model systems built for anti-VEGF diagnostic decisions. Secondly, the traditional visual model and the existing multi-mode large language model often cannot simulate the specific influence of 'treatment or not' on the future fundus anatomical structure at the pixel level, so that the long-range diagnosis and treatment planning lacks visual physical basis. Furthermore, clinical history data often only have sparse "treatment outcome" records, such as whether to inject and follow-up time intervals, lack Fine-grained pathology labeling (Fine-grained Annotations), and make it difficult to train a high-precision time-series pathology model. Disclosure of Invention In order to solve the technical problems of insufficient modeling method, scarce data and the like of the existing medical visual model, the first aspect of the invention provides an ophthalmic prognosis visual simulation system, which comprises: The feature extraction module is used for extracting biomarkers based on time sequence optical coherence tomography (Optical coherence tomography, OCT) images and historical diagnosis and treatment records of a user; an evolution analysis module for performing course trend analysis based on the biomarker and proposing a treatment regimen; And the forward deduction module is used for performing forward deduction of the counterfactual based on the processing scheme to obtain a predicted anatomical image of the fundus of the user at the future moment. Further, the biomarkers include, but are not limited to, intraretinal fluid accumulation, subretinal fluid accumulation in a user. Further, after the processing scheme is converted into a natural language instruction, the natural language instruction is sent to the forward deduction module. Further, the evolution analysis module is further configured to: and determining the risk level of the treatment scheme based on the predicted anatomical image deduced by t