CN-121998972-A - Pancreatic cancer CT image prediction method and system based on iterative self-evolution

CN121998972ACN 121998972 ACN121998972 ACN 121998972ACN-121998972-A

Abstract

The invention belongs to the technical field of medical image analysis and artificial intelligence aided diagnosis, and particularly relates to a pancreatic cancer CT image prediction method and system based on iterative self-evolution. The method comprises the following steps of S1, constructing a segmentation data set and a fine adjustment visual tool of pancreatic cancer, S2, generating interactive image-text reasoning by utilizing a pre-trained visual-language large model, S3, constructing a reasoning data screening and training set based on result consistency, S4, carrying out model iteration self-evolution training based on the screening data, and S5, deploying and reasoning output a final diagnosis model after multiple iterations. The invention introduces FastSAM tools for fine adjustment of specific diseases and an interactive reasoning framework, actively operates to improve the focus detection rate and the interpretability, constructs a self-training mechanism to form a closed-loop framework, reduces the dependence on manual labeling, generates a visual diagnosis report after deployment, and provides an intelligent diagnosis and treatment scheme with high precision, strong interpretation and continuous learning for clinic.

Inventors

REN HE
YU HAOJUN
XIN HAIYAN
Lv Caichao
LIU YINBO
Jia Xiangran
ZHENG JIACHEN
Shi Runxin
DING HUI

Assignees

青岛大学

Dates

Publication Date: 20260508
Application Date: 20260320

Claims (10)

1. The pancreatic cancer CT image prediction method based on iterative self-evolution is characterized by comprising the following steps of: S1, constructing a segmentation data set and a fine adjustment visual tool of pancreatic cancer specific diseases; S2, generating interactive image-text reasoning by utilizing a pre-trained visual-language big model; s3, filtering reasoning data based on result consistency and constructing a training set; S4, model iteration self-evolution training based on screening data; S5, deploying and reasoning output of a final diagnosis model after multiple iterations.
2. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 1, wherein: the S1 construction of the pancreatic cancer specific disease segmentation data set and the fine adjustment of the visual tool comprises the following specific steps: s11, constructing a segmentation data set of pancreatic cancer specific diseases: Inputting CT image data Corresponding expert labeling mask , The included categories are pancreas parenchyma, tumor focus and expanded pancreatic duct, and an image segmentation data set is formed; S12, constructing a fine adjustment vision tool for pancreatic cancer specific diseases: Selecting FastSAM pre-trained as a basic vision tool model, and constructing a mixed loss function containing Dice loss and cross entropy loss Performing parameter fine tuning on FastSAM to obtain a fine tuning visual tool special for pancreatic anatomy extraction, wherein an objective function of the parameter fine tuning is expressed as: (1) Wherein: representing the optimal visual tool model parameters obtained after fine tuning; Represented as model parameter variables to be optimized, A parametric operation denoted as finding a minimization of the loss function; represented as expected calculations for all pairs of samples within the data set; A predictive output function expressed as a visual segmentation model; represented as a hybrid segmentation loss function for measuring the difference of the predictive mask from the real mask.
3. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 2, wherein: the step S2 of generating interactive graphic reasoning by utilizing the pre-trained visual-language big model comprises the following specific steps: s21, utilizing a pre-trained visual-language big model Generating multi-step reasoning for CT image data of unlabeled reasoning chain, wherein the model follows a cycle mechanism of planning, executing, observing and describing, and the path is based on historical reasoning Generating visual operation instructions Then calling the fine-tuned visual tool Executing the instruction to generate a visual evidence diagram ; S23, combine Generating text evidence of the current step Entire inference chain The generation process of (a) is expressed as maximizing the joint probability: (2) Wherein: Expressed as generating a complete inference chain given image input and model parameters Is a joint probability of (2); an unlabeled CT image represented as an input; parameters expressed as a visual-language big model; Indicated as S1, the resulting visual tool is trimmed; total number of steps expressed as reasoning; Represented as at the first A visual operation instruction generated in the step; Denoted as the first A historic inference path prior to a step comprising ; Represented as according to instructions And an image A generated visual evidence diagram; A textual description represented as generated based on visual evidence; Represented as a conditional generation probability function.
4. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 3, wherein: the S3 reasoning data screening and training set construction based on result consistency comprises the following specific steps: Defining the finally derived diagnostic predictive label of the inference chain as The true pathological label corresponding to the case is ; Defining an indication function For screening effective reasoning paths and constructing self-evolution training data set : (3) Wherein: Represented as a screened high quality dataset for subsequent self-evolutionary training; c represents an inference chain generated by the model autonomously; Pathological diagnosis labels expressed as true cases; represented as models through inference chains The obtained predictive diagnosis conclusion; expressed as an indicator function, when the condition is in brackets The prediction result is consistent with the gold standard, the value is 1 when the prediction result is established, the sample is reserved, and otherwise, the value is 0, and the prediction result is taken as noise data to be removed.
5. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 4, wherein: The model iterative self-evolution training based on the screening data of the S4 comprises the following specific steps: S41, screening out high-quality data set And optimizing and updating the vision-language big model by adopting a supervision fine tuning technology: Defining a self-evolving loss function In order to maximize the log likelihood probability of generating the correct inference chain C, the calculation formula is as follows: (4) Wherein: a target loss function expressed as self-evolution training; expressed as visual-language big model parameters to be updated; represented as a high quality training dataset constructed in S3; Represented as image-inference chain sample pairs in a dataset; The total step length expressed as an inference chain C; represented as the first in inference chain C A text evidence segment; Represented as a generator A previous historical reasoning path; Expressed as a corresponding visual evidence graph; Expressed as a logarithmic probability for generating a target text sequence; s42, updating parameters through a gradient descent algorithm, and after one round of training is completed, taking the updated model as a new base model to carry out the next round of iteration.
6. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 5, wherein: The deployment and reasoning output of the final diagnosis model after the multiple iterations of S5 comprises the following specific steps: s51, will go through The final model of the round iterative evolution is deployed at the clinical end, and in the reasoning stage, the model is input into CT image to be diagnosed Automatically planning and generating optimal inference paths Diagnostic results : (5) Wherein: a diagnostic prediction result expressed as a final output; A visual reasoning path expressed as optimal; A new patient CT image to be diagnosed, represented as clinical input; the final large model parameters are expressed after multiple rounds of iterative self-evolution training; Argmax is expressed as finding a combination of diagnostic conclusions and reasoning processes that maximizes the joint posterior probability score; s52, reasoning and outputting a diagram containing key visual evidence And corresponding text parsing The whole process visual white box diagnosis is realized.
7. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 1, wherein: The prediction method further comprises the following instructions or judgment: Instructions for operation planning call FastSAM to segment the pancreatic whole contour; an active focusing instruction, which is to amplify the head hooking area and perform density analysis; An instruction for indirect sign verification, namely measuring the diameters of a main pancreatic duct and a common bile duct; Judging whether low-density blood supply lump, double-barrelled symptom, painless jaundice and CA19-9 are obviously increased or not; Judging whether to highly prompt pancreatic cancer or not and judging whether to perform ultrasonic endoscopic puncture biopsy or not.
8. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 1, wherein: the prediction method also comprises the following specific steps of: the input data is an enhanced CTDICOM sequence containing pancreatic stage, displaying images of the head uncrushing portion low-density range and the distal main pancreatic duct expansion, and clinical information of the patient, including age, sex, complaint and laboratory examination results.
9. The pancreatic cancer CT image prediction method based on iterative self-evolution according to claim 1, wherein: The prediction method also comprises the evaluation of model performance, and comprises the following specific steps: The performance of the model is evaluated through the receiver operating characteristic ROC curve and the AUC value, and the AUC value of the improved reasoning model is obviously superior to that of the traditional non-reasoning model and the non-improved reasoning model.
10. The pancreatic cancer CT image prediction system based on iterative self-evolution adopts the pancreatic cancer CT image prediction method based on iterative self-evolution as set forth in any one of claims 1-9, and is characterized by comprising the following modules: the segmentation data set module is used for collecting the CT image data of the pancreatic cancer, inviting medical professionals to mark pancreatic parenchyma, tumor focus and expand pancreatic duct, and forming an image segmentation data set; The vision tool fine adjustment module is used for selecting a pre-training FastSAM model, constructing a Dice and cross entropy mixed loss function by utilizing a pancreatic cancer specific disease segmentation data set, and carrying out parameter fine adjustment to obtain a special pancreatic anatomy structure extraction tool; The interactive image-text reasoning module combines the pre-training visual-language big model and the fine-tuning visual tool, generates visual operation instructions based on the historical reasoning path, and executes the instructions to generate visual evidence diagrams and text descriptions so as to form a complete reasoning chain; The reasoning data screening module is used for defining a diagnosis prediction label and a real pathology label, screening an effective reasoning path through result consistency check, constructing a self-evolution training data set, reserving a prediction consistency sample and eliminating noise data; the model iterative self-evolution training module optimizes a visual-language large model by using a high-quality data set and adopting a supervision fine tuning technology, defines a self-evolution loss function to maximize the log-likelihood probability of generating a correct reasoning chain, and iteratively trains and improves the model performance by gradient descent update parameters; And the diagnosis report output module is used for deploying a final model after multiple iterations at a clinical end, inputting a CT image to be diagnosed, automatically generating an optimal reasoning path and a diagnosis result, and outputting a complete white box diagnosis report containing a key visual evidence diagram and text analysis.

Description

Pancreatic cancer CT image prediction method and system based on iterative self-evolution Technical Field The invention belongs to the technical field of medical image analysis and artificial intelligence aided diagnosis, and particularly relates to a pancreatic cancer CT image prediction method and system based on iterative self-evolution. Background Pancreatic cancer has hidden onset, rapid progression and extremely high mortality. Early pancreatic cancer lesions often only show equal density or low density nodules in the pancreas parenchyma or very fine pancreatic duct cut-off and expansion signs on CT images, are easily confused with other benign diseases, lead to difficult early diagnosis and high misdiagnosis rate. Along with the digitization process of medical imaging technology, hospitals accumulate massive abdomen CT scan data, and a large number of pathological morphological features which are not fully mined are precipitated in the data. The prior art utilizes artificial intelligence technology to deeply analyze the complex image information, realizes accurate positioning, qualitative and logical reasoning diagnosis of pancreatic cancer, and preliminarily improves early survival rate of patients. Traditional non-inferential medical image models (e.g., lesion classifiers based on dedicated CNN or ViT architectures) are typically trained using sliding window mechanisms or local clipping based on lesion bounding boxes, which concentrate the overall computational effort on local pathology textures. While the simple Qwen-VL-8B model supports high resolution dynamic inputs when receiving full panoramic abdominal CT images, it produces thousands of massive visual tokens. In a global field of view containing liver, gastrointestinal tract, spleen and large amounts of adipose tissue, a 1.5cm early occult pancreatic micro-nodule occupies only a very small number of Token. Under the global self-attention mechanism of the transducer, the attention weight of the model is very easily dispersed and diluted by a large and complex background organization. Without the intervention of active local focus, the sensitivity of the large model to massive tiny foci is naturally lower than that of a traditional classifier which is specially fitted to local pixels. Chinese patent document CN121101470a discloses a method for discriminating the degree of risk based on prostate Magnetic Resonance (MRI) scan images and PSA results using a graph rolling network (GCN) and generating a cancer growth video generated against the network (GAN). Chinese patent document CN120954689a discloses a method for segmenting and diagnosing prostate ultrasound video using a multi-modal large model (MedSAM 2 +clip). There remains a deficiency in the prior art and its representative improvements. Although the Chinese patent document CN121101470A builds a graph structure of antigen examination-magnetic resonance examination, the edge characteristics are simply defined as examination consistency degree, clinical information (such as symptoms) and image details (such as focus morphology and strengthening characteristics) are not fused deeply, the fusion dimension is single, complex reasoning cannot be supported, the generation of an anti-network generation cancer simulated growth video is relied on, the consistency of the simulated video and the real focus development is not verified, the short-term neural network model only judges the risk degree based on the video time sequence characteristics, the step reasoning logic is not combined, the clinical interpretation is lacking in quantification of the risk degree, the object to be further examined is screened only through the antigen examination risk value > a preset threshold value, the preliminary screening of an image layer is not introduced, the early cases of normal antigen examination but abnormal images are possibly missed, and the early screening of cancers is unfavorable. The Chinese patent document CN120954689A relies on a two-stage process of screening-alignment training, a gradually refined reasoning mechanism is not designed, the reasoning reliability of a key area (such as focus details) cannot be enhanced through the step logic of positioning-analysis-verification, the recognition capability of a small focus and a boundary fuzzy focus is weak, the dynamic adaptability is lacking in a mode of sampling (fixing 8 frames) and clustering selected key frames through similarity screening and compressing irrelevant information, the capturing of focus dynamic changes in an ultrasonic video is insufficient, and an iterative data optimization mechanism is not introduced, so that the data quality is difficult to continuously improve. In summary, there is a need to provide a pancreatic cancer CT image prediction method and system based on iterative self-evolution, so as to solve the problems in the prior art. Disclosure of Invention The invention aims to solve the technical problems that the traditi