CN-122025194-A - Multi-modal information-fused radioactive proctitis causal prediction method and system

CN122025194ACN 122025194 ACN122025194 ACN 122025194ACN-122025194-A

Abstract

The application discloses a method and a system for predicting cause and effect of radiation proctitis by fusing multi-modal information, wherein the method comprises the steps of acquiring and preprocessing multi-modal medical data of a patient, extracting multi-modal high-dimensional joint features through a deep network, constructing a deep generation decoupling model, decoupling the joint features into independent internal pathogenic mechanism characterization and external mixed interference characterization, blocking the influence of the mixed characterization on prediction by applying intervention based on a cause and effect inference theory, and calculating a cause and effect predicted value of the radiation proctitis by only utilizing the internal pathogenic mechanism characterization. The application effectively eliminates the mixed bias caused by heterogeneous data through causal characterization decoupling and causal intervention prediction, avoids shortcut learning of the model, and improves generalization capability, robustness and technical credibility of the model in complex clinical environments such as cross-center, multi-equipment and the like.

Inventors

LIU JUN
YANG QINJIE
CHU KEXIN
Shi Liwan
WU SANGANG
LI YIMIN
LI XING
Wang Dunhuang
HE YIPENG
Zhu Luchao
LIN LIMEI
Lin Yanzong
HUANG YUNXIA
ZHOU YUFEI

Assignees

厦门大学附属第一医院(厦门市第一医院、厦门市红十字医院、厦门市糖尿病研究所)

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The radioactive proctitis causal prediction method integrating the multi-mode information is characterized by comprising the following steps of: Firstly, acquiring and preprocessing multi-modal original electronic medical data of a pelvic tumor patient, wherein the multi-modal original electronic medical data comprises a magnetic resonance image, clinical electronic medical record data and a three-dimensional physical dose distribution map so as to generate an aligned multi-modal input tensor set; processing the aligned multi-modal input tensor set by using a deep multi-modal feature extraction network, extracting modal features and fusing the modal features into a multi-modal high-dimensional combined deep hidden feature representation; Constructing a depth generation decoupling inference model comprising an encoder, wherein the depth generation decoupling inference model is obtained by performing countermeasure training with a domain identification network in advance, so that the internal pathogenic mechanism characterization decoupled by the depth generation decoupling inference model does not comprise mixed information irrelevant to a prediction target; And fourthly, constructing a probability causal computing network model, cutting off a neuron activation channel corresponding to the exogenous confounding interference characterization in forward propagation by applying an intervention matrix mask, only preserving the activation value of the inherent pathogenic mechanism characterization, and mapping the activation value to be a radioactive proctitis causal prediction value.
2. The method of claim 1, wherein in step one the magnetic resonance image is an image of a patient's radiotherapy target volume space, and the three-dimensional physical dose profile consists of three-dimensional spatial voxels comprising absorbed dose values.
3. The method of claim 1, wherein the domain identification network is a conditional contrast domain identification discriminator for identifying domain attributes of data sources to which the input data belongs, wherein the domain identification network is connected to an output of the encoder through a gradient inversion layer, wherein the intrinsic pathogenic mechanism characterization is forced to not include confounding information reflecting a data acquisition environment or source during the contrast training, wherein the data source domain attributes include at least one of different medical centers, different scanning devices or different acquisition parameters, and wherein the loss function used for the contrast training includes a sample mutual information minimization penalty term.
4. The method of claim 1, wherein the deep multi-modal feature extraction network comprises a three-dimensional convolutional neural network cluster configured with residual connections for processing the magnetic resonance image and the three-dimensional physical dose profile, and a feed-forward fully-connected encoding network comprising a multi-headed self-attention mechanism for processing the clinical electronic medical record data.
5. The method of claim 1, wherein the depth-generating decoupling inference model is a variational self-encoder-based generation model, the encoder comprising two parallel inference subnetworks, a first inference subnetwork for outputting a mean and a variance of a first multidimensional gaussian distribution to which the intrinsic pathogenic mechanism characterization is subject, and a second inference subnetwork for outputting a mean and a variance of a second multidimensional gaussian distribution to which the extrinsic confounding interference characterization is subject.
6. The method of claim 1, wherein the applying an intervention matrix mask is specifically stitching the intrinsic pathogenic mechanism characterization and the extrinsic confounding interference characterization, multiplying the stitched features element by element with a binary mask matrix as the intervention matrix mask, the binary mask matrix being 1 in a position corresponding to the intrinsic pathogenic mechanism characterization dimension and 0 in a position corresponding to the extrinsic confounding interference characterization dimension.
7. The method of claim 2, further comprising the step of perturbing three-dimensional spatial voxels in the three-dimensional physical dose distribution map and calculating a change in causal predictors of radiation proctitis based on the probabilistic causal computation network model to perform a voxel-level inverse factual attribution backtracking to generate a three-dimensional causal sensitivity spatial attribution thermodynamic diagram.
8. The method of claim 7, wherein the inverse fact attribution backtracking specifically comprises locking the input values of the magnetic resonance image and the clinical electronic medical record data unchanged, determining perturbed three-dimensional spatial voxels with a preset three-dimensional mask window, modifying an absorption dose value corresponding to the perturbed three-dimensional spatial voxels to generate a modifier, and performing a three-dimensional edge smoothing projection mapping operation on the modifier through a preset ray scattering convolution filter matrix based on a dose space distribution characteristic, thereby generating a modified inverse fact dose map input.
9. The method of claim 8, wherein the specific step of generating the three-dimensional causal sensitivity spatial attribution thermodynamic diagram comprises inputting the modified anti-facts dose map to the probabilistic causal computing network model to compute a new predictor, inversely mapping a variance difference between the new predictor and a pre-disturbance predictor to a spatial coordinate location corresponding to the disturbed three-dimensional spatial voxel, and generating the three-dimensional causal sensitivity spatial attribution thermodynamic diagram by feature map rendering.
10. A radioactive proctitis causal prediction system fusing multimodal information, comprising: The data acquisition and preprocessing module is used for acquiring and preprocessing multi-mode original electronic medical data of a pelvic tumor patient, wherein the multi-mode original electronic medical data comprises a magnetic resonance image, clinical electronic medical record data and a three-dimensional physical dose distribution map so as to generate an aligned multi-mode input tensor set; The multi-modal feature extraction module is used for processing the aligned multi-modal input tensor set by utilizing a deep multi-modal feature extraction network, extracting modal features and fusing the modal features into a multi-modal high-dimensional combined deep hidden feature representation; The system comprises a causal characterization decoupling module, a multi-mode high-dimensional combined deep hidden feature representation and a multi-mode high-dimensional combined deep hidden feature representation, wherein the causal characterization decoupling module is used for constructing a depth generation decoupling inference model comprising an encoder, the depth generation decoupling inference model is obtained by performing countermeasure training with a domain identification network in advance, so that an intrinsic pathogenic mechanism representation decoupled by the depth generation decoupling inference model does not comprise clutter information irrelevant to a prediction target; And the causal intervention prediction module is used for constructing a probabilistic causal calculation network model, cutting off neuron activation channels corresponding to the exogenous confounding interference characterization in forward propagation by applying an intervention matrix mask, only preserving the activation value of the inherent pathogenic mechanism characterization, and mapping the activation value to be a radioactive rectitis causal prediction value.

Description

Multi-modal information-fused radioactive proctitis causal prediction method and system Technical Field The application relates to the technical field of combination of artificial intelligence and medical data processing, in particular to a radioactive proctitis causal prediction method and system integrating multi-mode information. Background Radiation proctitis is a common and serious complication during radiotherapy of pelvic malignancies (e.g., prostate cancer, cervical cancer, etc.). The pathological manifestations of the composition can be from early rectal mucosa congestion, edema to distant interstitial fibrosis, microvascular occlusion and other irreversible injuries, so that the patients can suffer from symptoms such as hematochezia, pain, defecation dysfunction and the like, and the life quality of the patients is seriously influenced. In order to evaluate and predict the risk of serious rectal injury of a patient before radiation therapy or in early treatment, a deep learning algorithm is often adopted in clinical research, and an intelligent risk evaluation model is constructed by utilizing multi-mode medical information. Such multimodal information typically includes high resolution multiparameter Magnetic Resonance Imaging (MRI), structured clinical feature data (e.g., medical history, tumor stage, biochemical indicators, etc.), and voxel-level three-dimensional physical dose distribution maps. However, existing artificial intelligence models face a challenge of generalization capability in processing such multimodal medical data. These models are typically predicted by mining statistical correlations between the high-dimensional features of the input and the risk probabilities of the output. In a practical clinical setting, the generation and collection of multimodal data can be disturbed by a number of complex non-clinical factors. For example, there are systematic differences in equipment model, scan parameters, calibration standards, and baseline characteristics of patient populations for different medical centers. This results in the model possibly learning some "false correlations" during the training process, i.e. the model incorrectly correlates background noise or systematic deviations specific to a certain data acquisition environment (e.g. image artifacts produced by a specific device, patient age structure of a specific region) as decisive features for the complications. That is, the model is easy to generate wrong association due to over-dependence on environmental noise, so that an immaterial statistical rule related to target task false in the data is learned. The problems are particularly prominent in special medical application scenes such as cross-center and multi-source heterogeneous equipment joint deployment. When a model trained in a single center is deployed to another center with different equipment and data distribution, the predictive performance is often significantly degraded. The model cannot discern true intrinsic pathogenic links due to the confounding bias present in the training data, resulting in unreliable and even erroneous risk assessment conclusions output on the new, unseen data distribution. The resulting heat map of the attribution analysis therefore loses clinical reference value and does not provide a reliable decision-making aid for the clinician in making or adjusting the radiation treatment plan. Therefore, how to eliminate the hybrid bias caused by heterogeneous data and improve the generalization capability and the technical reliability of the model in a complex clinical environment is a technical problem to be solved urgently in the current medical artificial intelligence field. Disclosure of Invention The invention aims to solve the technical problems that the existing radioactive proctitis causal prediction model fused with multi-mode information is easy to learn the mixed bias introduced by factors such as equipment heterogeneity, data acquisition environment difference and the like due to the dependence on statistical correlation in data, so that the model has insufficient generalization capability under complex scenes such as cross-center deployment and the like, and the reliability of a prediction result is low. In order to solve the technical problems, a first aspect of the present invention provides a causal prediction method for radiation proctitis by fusing multi-modal information, comprising the following steps: Firstly, acquiring and preprocessing multi-modal original electronic medical data of a pelvic tumor patient, wherein the multi-modal original electronic medical data comprises a magnetic resonance image, clinical electronic medical record data and a three-dimensional physical dose distribution map so as to generate an aligned multi-modal input tensor set; processing the aligned multi-modal input tensor set by using a deep multi-modal feature extraction network, extracting modal features and fusing the modal features into a mu