CN-122025127-A - Deep learning-based nerve injury risk prediction system caused by radiotherapy

CN122025127ACN 122025127 ACN122025127 ACN 122025127ACN-122025127-A

Abstract

The invention discloses a deep learning-based nerve injury risk prediction system caused by radiotherapy, which belongs to the technical field of medical artificial intelligence and radiotherapy, and comprises a nerve structure automatic segmentation module, a dose-nerve response modeling module, an individuation risk assessment module, a radiotherapy plan optimization module and a closed-loop feedback regulation unit, wherein the system adopts an encoder-decoder depth network to segment nerve tissues with high precision, the method comprises the steps of establishing a dose-neural response quantitative model by fusing dose histology and image histology characteristics, realizing individual risk assessment based on a multi-factor risk quantitative network, generating an optimization scheme for balancing tumor control and neuroprotection by pareto optimal search, and realizing cooperative optimization of each module by closed loop feedback.

Inventors

WANG GUIJUAN
LIU YANTAO
ZHANG XIAO
LI PENG
WANG HUILI
ZHANG XINHUA
LI HUI
XU HUICHAO

Assignees

济宁医学院附属医院

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (10)

1. Deep learning-based radiotherapy-induced nerve injury risk prediction system, characterized by comprising: the nerve structure automatic segmentation module is used for receiving multi-mode medical image data of a patient, performing voxel level segmentation on nerve tissues by adopting a deep nerve network of an encoder-decoder structure, and outputting a nerve structure segmentation graph and a corresponding segmentation confidence level graph, wherein the encoder extracts multi-scale nerve anatomical features through multi-layer convolution and downsampling operation, and the decoder restores spatial resolution through upsampling and jump connection; The dose-nerve response modeling module is used for receiving the radiotherapy planning parameters and the nerve structure segmentation map, constructing a quantitative relation model between the radiation dose and the nerve tissue response by fusing the dose histology characteristics and the image histology characteristics, and obtaining a nerve sensitivity map based on radiotherapy planning and follow-up data training of historical cases, wherein the nerve sensitivity map represents the sensitivity degree difference of different nerve regions to the radiation dose; The individualized risk assessment module is used for receiving the neural structure segmentation map, the segmentation confidence map, the neural sensitivity map and specific anatomical structure parameters and treatment parameters of a patient, and calculating individualized neural injury risk scores and injury probability spatial distribution through a multi-factor risk quantification network, wherein the multi-factor risk quantification network carries out weighted fusion on anatomical factors, dose factors and individual sensitivity factors; The radiotherapy plan optimizing module is used for receiving the personalized nerve injury risk score, the injury probability spatial distribution and the nerve sensitivity map, performing pareto optimal search between a tumor control rate and a nerve protection rate through a multi-objective optimizing algorithm, and generating an optimized radiotherapy scheme; And the closed-loop feedback adjustment unit is used for calculating the expected nerve injury risk deviation according to the simulation execution result of the optimized radiotherapy scheme, and transmitting feedback adjustment parameters to the nerve structure automatic segmentation module, the dose-nerve response modeling module and the individuation risk assessment module so as to update the processing parameters of each module.
2. The system of claim 1, wherein the neural structure automatic segmentation module triggers the multi-scale feature enhancement process when a segmentation confidence level is below a preset segmentation confidence level threshold, wherein the segmentation confidence level threshold has a value ranging from 0.75 to 0.90.
3. The system of claim 1, wherein in the dose-neural response modeling module, the dose-sensitivity coefficients are differentially configured according to neural tissue type, wherein the cranial nerves have a dose-sensitivity coefficient value ranging from 1.2 to 1.8, the spinal nerves have a dose-sensitivity coefficient value ranging from 1.5 to 2.2, and the peripheral nerves have a dose-sensitivity coefficient value ranging from 0.8 to 1.3.
4. The system of claim 1, wherein the risk level output by the personalized risk assessment module is divided into four levels of low risk, medium risk, high risk, and extremely high risk according to a preset risk level division threshold, wherein the low risk corresponds to a risk score of less than 0.25, the medium risk corresponds to a risk score of greater than or equal to 0.25 and less than 0.50, the high risk corresponds to a risk score of greater than or equal to 0.50 and less than 0.75, and the extremely high risk corresponds to a risk score of greater than or equal to 0.75.
5. The system of claim 1, wherein the encoder-decoder architecture of the neural architecture automatic segmentation module comprises four encoding stages and four decoding stages, each encoding stage comprising two convolutional layers and one max-pooling layer, each decoding stage comprising one transpose convolutional layer and two convolutional layers, the encoding stages and the corresponding decoding stages delivering multi-scale features via a skip connection.
6. The system of claim 1, wherein in the dose-neural response modeling module, the dose-histology features comprise a dose volume histogram feature, a dose gradient feature, and an isodose line morphology feature, the image-histology features comprise a gray level co-occurrence matrix feature, a gray level run-length matrix feature, and a shape feature, and the quantitative relationship model adaptively weighted fuses the dose-histology features and the image-histology features through a attentional mechanism.
7. The system of claim 1, wherein in the radiotherapy plan optimization module, a multi-objective optimization algorithm takes tumor control probability and normal tissue complication probability as optimization objectives, generates a set of non-dominant solution sets through pareto front search strategy, and selects a balance solution from the non-dominant solution sets as a final optimization scheme according to preset optimization weight factors.
8. The system of claim 1, wherein the multi-factor risk quantification network of the personalized risk assessment module comprises a feature encoding sub-network, a cross-modal fusion sub-network and a risk prediction sub-network, the feature encoding sub-network performs feature extraction on anatomical factors, dose factors and individual sensitivity factors respectively, the cross-modal fusion sub-network realizes deep fusion of multi-modal features through a cross-attention mechanism, and the risk prediction sub-network outputs a risk score and a damage probability distribution.
9. The system according to claim 1, wherein the expected nerve injury risk deviation calculated by the closed-loop feedback adjustment unit comprises a segmentation accuracy deviation, a response prediction deviation and a risk assessment deviation, and the feedback adjustment parameters are generated in a weighted manner according to the contribution degree of the various deviations and respectively act on the adjustable parameters of the corresponding modules.
10. The system of claim 1, wherein the neural sensitivity map is constructed based on radiotherapy planning data and corresponding follow-up MRI data for not less than 500 historical cases, the neural sensitivity map storing the radiosensitivity coefficient and cumulative damage probability for each location in voxels for not less than 12 months.

Description

Deep learning-based nerve injury risk prediction system caused by radiotherapy Technical Field The invention belongs to the technical field of medical artificial intelligence and radiotherapy, and particularly relates to a deep learning-based nerve injury risk prediction system caused by radiotherapy. Background Radiation therapy plays a vital role in clinical oncology practice as one of the three main posts in the treatment of malignant tumors. It is counted that more than about seventy percent of malignant patients worldwide need to receive varying degrees of radiation therapy during the course of treatment. Along with the rapid development and wide application of intensity-modulated radiation therapy technology, volume-modulated arc therapy technology, image-guided radiation therapy technology and stereotactic radiation therapy technology, the accuracy and effectiveness of modern radiation therapy are remarkably improved, and the local control rate of tumors and the survival time of patients are remarkably improved. However, radiation therapy inevitably causes a degree of damage to surrounding normal tissue while killing tumor cells. Neural tissue is highly sensitive to ionizing radiation due to its unique biological properties. Neural cells belong to terminally differentiated cells, which have extremely limited proliferation capacity, and once they are damaged by radiation, they have far from the ability to repair and regenerate other types of tissue cells. The clinical manifestations of radiation nerve injury are diverse and can involve various parts of the central and peripheral nervous systems. In the aspect of the central nervous system, the radioactive brain injury can be represented by neuropsychological abnormalities such as cognitive dysfunction, hypomnesis, attention loss, impaired executive function and the like, and the severe can be caused by radioactive brain necrosis, resulting in permanent neurological impairment. The radioactive spinal cord lesions may cause symptoms such as paresthesia, reduced muscle strength, sphincter dysfunction, and even paraplegia. In the peripheral nervous system, the brachial plexus, the lumbosacral plexus, and various cranial nerves can all be subjected to radiation injury, manifested as sensory disorders, hypokinesia, or autonomic dysfunction in the corresponding innervation areas. The occurrence of radiation nerve damage is closely related to a variety of factors. From a radiophysics perspective, total dose, fraction dose, irradiation volume, dose rate, and radiation type are all important factors that affect the risk of nerve damage. From the perspective of radiobiology, factors such as inherent radiosensitivity, cell proliferation kinetics, damage repair capability, and microvascular density of nerve tissue determine the tolerance of the nerve tissue to radiation. From an individual factor perspective, patient age, basal neurological status, complications, past history of radiation therapy, and use of contemporaneous chemotherapeutic agents, etc., can all affect the probability and severity of occurrence of radiation nerve damage. The existing radiotherapy planning quality control technology mainly focuses on the physical accuracy verification level of dose distribution. Chinese patent No. 113516233A discloses a neural network prediction method for VMAT radiotherapy plan, which realizes the prediction evaluation of the execution precision of the radiotherapy plan by respectively learning the mapping relation among the radiotherapy plan features, the accelerator features and the gamma passing rate by constructing a multi-branch neural network model. The technology integrates multidimensional information by adopting a feature fusion strategy, and achieves a certain effect in the aspect of prediction accuracy. However, the technology still has the following limitations that the method mainly focuses on the physical execution precision prediction of the radiotherapy plan, lacks the evaluation capability of biological effects caused by radiotherapy, particularly the damage risk of a nervous system, adopts characteristic engineering relatively traditional, fails to fully utilize the deep learning technology to excavate sensitivity difference information of nerve tissues to radiation dose, lacks a personalized analysis module aiming at the individual anatomical structure characteristics and clinical characteristics of patients, is difficult to realize precise personalized risk prediction, belongs to an open-loop prediction system, lacks an optimization feedback mechanism, and cannot guide adjustment and optimization of the radiotherapy plan according to a prediction result. In recent years, although researchers have attempted to apply deep learning techniques to the field of radiotherapy toxicity prediction, these studies are mostly limited to prediction of specific types of normal tissue complications, such as radiation pneumonitis, radiation esophagitis, and the