CN-121983301-A - Liver and gall postoperative lesion quantitative evaluation method and system based on DLPE algorithm

CN121983301ACN 121983301 ACN121983301 ACN 121983301ACN-121983301-A

Abstract

The invention relates to the technical field of biomedical engineering and artificial intelligence intersection, in particular to a method and a system for quantitatively evaluating liver and gall postoperative lesions based on a DLPE algorithm, which comprise the steps of obtaining multi-time-point multi-mode medical images and generating a four-dimensional data set with space alignment through deep learning registration; the method comprises the steps of carrying out multi-mode space-time feature fusion by using a parallel 3D-CNN and a cross-mode attention mechanism, inputting a fusion feature sequence into a GRU network modeling lesion dynamic evolution process, outputting a global evolution feature vector, generating a voxel level pathology evolution probability map through a three-dimensional decoder, and extracting quantitative indexes such as volume, morphology, signal intensity and the like of each pathology state. Through the technical scheme, the invention can realize accurate differentiation and dynamic quantification of tumor recurrence, inflammatory edema and scar tissues, and improves objectivity, accuracy and clinical decision support capability of liver and gall postoperative efficacy evaluation.

Inventors

MAO XIANHAI
DUAN XIAOHUI
LUO JIA
SHI LEI
XIE WEI
WANG XIAOHUI
CHEN PAN
XIAO HUA
LONG YING

Assignees

湖南省肿瘤医院

Dates

Publication Date: 20260505
Application Date: 20251202

Claims (10)

1. The method for quantitatively evaluating the liver and gall postoperative lesions based on the DLPE algorithm is executed on the basis of a computing device and is characterized by comprising the following steps of: The method comprises the steps of A, data acquisition and time sequence alignment processing, namely acquiring a medical image sequence of a patient at a plurality of discrete time nodes after operation, wherein the medical image sequence comprises at least two modes, and the modes at least comprise a Computed Tomography (CT) image and a Magnetic Resonance Imaging (MRI) image; Step B, a multi-mode space-time feature depth fusion step, namely, independently extracting depth features of each mode image data of each time node in the multi-mode four-dimensional image data set aligned in space by utilizing a parallel three-dimensional convolutional neural network (3D-CNN) encoder group to generate a feature map corresponding to each mode; Step C, a lesion dynamic evolution track modeling step, namely inputting the fused space-time characteristic sequence into a gating circulation unit (GRU) network in time sequence, capturing and encoding a dynamic change rule of a lesion region in a time dimension by the GRU network, and outputting a global evolution characteristic vector containing complete lesion evolution history information at the last time node; And D, a pathological evolution probability map generation and quantization analysis step, namely inputting the global evolution feature vector into a three-dimensional convolution decoder network, carrying out up-sampling and feature decoding on the global evolution feature vector to generate a three-dimensional voxel-level pathological evolution probability map, wherein each voxel of the pathological evolution probability map comprises probability values of a plurality of channels, each channel corresponds to a preset postoperative pathological state, and the pathological state at least comprises tumor tissue recurrence, active inflammation or edema and stable scar tissue, and calculating and outputting a quantization evaluation index based on the pathological evolution probability map.
2. The quantitative evaluation method of liver and gall postoperative lesions based on DLPE algorithm according to claim 1, wherein the anatomical-structure-aware deformable registration procedure executed in the step a specifically comprises: An unsupervised deep learning network based on VoxelMorph architecture is adopted as a registration network, a pair of three-dimensional images, namely a floating image to be registered and a fixed image as a target are taken as input to output a dense three-dimensional deformation field, a training process of the registration network is optimized by adopting a composite loss function, the composite loss function comprises an image similarity loss term, a deformation field smoothness regularization term and an anatomical structure consistency loss term, the image similarity loss term is used for measuring similarity between the floating image and the fixed image after being transformed by the three-dimensional deformation field, the deformation field smoothness regularization term is used for punishing gradient of the three-dimensional deformation field so as to ensure topological continuity of deformation, and the anatomical structure consistency loss term is used for forcing the registration network to give higher weight when aligning the key anatomical structures, wherein the key anatomical structures comprise abdominal aorta, inferior vena cava, portal vein trunk and branch and residual liver outline.
3. The quantitative evaluation method of liver and gall postoperative lesions based on DLPE algorithm according to claim 1, wherein the step of multi-modal space-time feature depth fusion in the step B specifically adopts a hierarchical progressive cross-modal attention fusion mechanism, and the method comprises the following steps: The parallel three-dimensional convolutional neural network encoder group is composed of a plurality of three-dimensional residual error networks (3D-ResNet) with the same structure, the 3D-ResNet outputs multi-scale feature maps at a plurality of feature levels with different depths, a cross-mode attention fusion unit is arranged at least one of the feature levels, the fusion unit takes the feature map of a certain mode as a Query (Query) and takes the combination of the feature maps of all modes at the same level as a Key (Key) and a Value (Value), attention weights are generated by calculating the dot product similarity between the Query and the Key, the attention weights are applied to the Value and subjected to weighted summation, so that new feature maps which are enhanced by other mode information and correspond to the certain mode are generated, and the weighted summation process is carried out once on each mode to obtain a group of feature maps which are subjected to depth interaction fusion at the current level and serve as input of a next network layer.
4. The quantitative evaluation method for the post-hepatobiliary-operation lesions based on the DLPE algorithm according to claim 1, wherein the step of modeling the dynamic evolution trajectory of the lesions in the step C specifically comprises: before the fused space-time feature sequence is input into the GRU network, a global average pooling layer and a full connection layer are firstly applied to the fused feature pattern of each time node in the sequence, the fused feature pattern is compressed into a feature vector with a fixed length, so that a feature vector sequence is formed, a multi-layer stacked bidirectional gating circulation unit (BidirectionalGRU) network is adopted to process the feature vector sequence, forward GRU in the bidirectional GRU network is used for capturing historical evolution information from the past to the present, backward GRU is used for capturing context information based on future trends, and the forward hidden state and the backward hidden state of the bidirectional GRU network at the last time node are spliced to form the global evolution feature vector.
5. The method according to claim 1, wherein in the step D, the three-dimensional convolutional decoder network adopts a three-dimensional U-Net decoder, and the decoder receives and fuses the fusion feature map output by the encoder level corresponding to the current decoding level in the step B of multi-mode space-time feature depth fusion through jump connection (SkipConnection) in the process of gradually restoring spatial resolution, and the quantized evaluation index calculated and output based on the pathology evolution probability map further comprises a time evolution curve of at least one index calculated according to each time node, wherein the time evolution curve comprises a morphological index comprising a volume, a volume change rate, a maximum three-dimensional diameter, a surface area and sphericity of each pathology state region, and a signal strength index comprising a signal strength average value, a standard deviation, a bit number and a percentage of evolution on a CT image, a T1 weighted imaging, a T2 weighted imaging and A Diffusion Coefficient (ADC) map calculated after mapping the mask of each pathology state region back to an original multi-mode image.
6. Liver and gall postoperative pathological change quantitative evaluation system based on DLPE algorithm, its characterized in that, the system includes: A data interface and time series image registration unit configured to acquire a multi-modality medical image sequence including at least Computed Tomography (CT) images and Magnetic Resonance Imaging (MRI) images of a patient at a plurality of discrete time nodes after surgery and perform a depth learning based anatomical-aware deformable registration procedure to generate a set of spatially aligned multi-modality four-dimensional image datasets; The multi-modal spatial-temporal feature fusion module is electrically connected with the data interface and the time sequence image registration unit and is configured to utilize a parallel three-dimensional convolutional neural network (3D-CNN) encoder group to conduct depth feature extraction on the four-dimensional image data set and conduct feature weighted fusion through a cross-modal attention fusion mechanism so as to generate a fused spatial-temporal feature sequence; The lesion dynamic evolution modeling module is electrically connected with the multi-mode space-time feature fusion module, integrates a gating and circulating unit (GRU) network inside, and is configured to receive and process the fused space-time feature sequence so as to output a global evolution feature vector containing complete lesion evolution history information; The pathological evolution probability map generation and quantitative analysis module is electrically connected with the pathological change dynamic evolution modeling module, comprises a three-dimensional convolution decoder network, is configured to decode the global evolution feature vector into a pathological evolution probability map which comprises a three-dimensional voxel level and a plurality of preset pathological state channels, and calculates quantitative evaluation indexes based on the probability map.
7. The quantitative evaluation system of post-hepatobiliary lesions based on the DLPE algorithm of claim 6, wherein the data interface and sequential image registration unit is internally cured with an unsupervised deep learning registration network model based on VoxelMorph architecture trained to output a three-dimensional deformation field driving floating images to align to fixed images, the optimization objective of the model being defined by a composite loss function comprising an image similarity loss term based on normalized cross-correlation (NCC) coefficients, a smoothness regularization term for ensuring physical rationality of deformation, and a dess (Dice) loss term based on critical anatomy masks to preferentially ensure precise alignment of the critical anatomy during registration.
8. The quantitative evaluation system for liver and gall postoperative lesions based on DLPE algorithm according to claim 6, wherein the multi-modal spatio-temporal feature fusion module is characterized in that the core of the multi-modal spatio-temporal feature fusion module is a neural network structure comprising a plurality of parallel feature extraction branches and a hierarchical fusion unit, each feature extraction branch is a three-dimensional residual network (3D-ResNet) and is respectively responsible for extracting deep space feature maps of different levels from aligned different modal images, the hierarchical fusion unit is composed of a plurality of cross-modal attention modules arranged at different feature levels of the 3D-ResNet, each attention module is configured to take feature maps of a certain mode as a query, combine feature maps of all modes as keys and values, and calculate and apply cross-modal attention weights so as to achieve information enhancement and effective aggregation between the feature maps.
9. The quantitative evaluation system of post-hepatobiliary-operation lesions based on the DLPE algorithm of claim 6, wherein the lesion dynamic evolution modeling module is characterized by a multi-layered stacked network of bi-directional gating cyclic units (BidirectionalGRU) for learning the evolution pattern of lesions from both forward and reverse time sequences simultaneously, and wherein the three-dimensional convolutional decoder network in the pathological evolution probability map generation and quantization analysis module has a three-dimensional U-Net decoder structure comprising a plurality of jump connection paths that directly transmit the fusion feature maps of the encoder levels in the multi-modal spatiotemporal feature fusion module to corresponding up-sampling levels in the decoder network to assist in accurate reconstruction of high resolution spatial detail.
10. The DLPE algorithm-based liver and gall postoperative lesion quantitative assessment system according to claim 6, further comprising: The result visualization and report generation module is electrically connected with the pathology evolution probability map generation and quantification analysis module and is configured to perform at least one of the functions of superposing and displaying the pathology evolution probability map on an original CT or MRI image in a pseudo-color mode, providing an interactive three-dimensional visualization interface for a user to observe three-dimensional space distribution and internal components of a lesion from any angle through a volume drawing technology, automatically integrating the quantification evaluation index, a time evolution trend map generated according to the index and a key image snapshot into a structured electronic evaluation report, and generating text description of the lesion evolution trend according to preset rules.

Description

Liver and gall postoperative lesion quantitative evaluation method and system based on DLPE algorithm Technical Field The invention relates to the technical field of biomedical engineering and artificial intelligence intersection, in particular to a liver and gall postoperative lesion quantitative evaluation method and system based on a DLPE algorithm. Background Along with the rapid development of medical imaging technology and artificial intelligence algorithms, the application of the lesion quantitative evaluation method in diagnosis and treatment after liver and gall surgery gradually becomes a research hot spot. Through the accurate quantitative analysis of the postoperative lesion area, a more accurate diagnosis basis can be provided for doctors, so that the treatment scheme is optimized and the prognosis of patients is improved. However, the existing lesion quantification evaluation method still has a plurality of defects when processing complex postoperative lesions, and particularly has to be improved in the aspects of multi-mode data fusion, lesion region segmentation precision and dynamic quantification capability. Through retrieval, a quantitative analysis method of pneumonia fibrosis based on DLPE algorithm is disclosed, wherein the patent discloses a method for quantitatively analyzing the pneumonia fibrosis based on the DLPE algorithm, the method comprises the steps of collecting lung CT data sets, carrying out normalization treatment, designing a segmentation model to segment lung, respiratory tract, blood vessel and visible lesions, and establishing a stronger segmentation and quantization model by combining a 2.5D segmentation algorithm with a human-in-the-loop program. However, the technical scheme is mainly aimed at quantitative analysis of pneumonia fibrosis, and has obvious difference between application scenes and lesions after liver and gall operation. Specifically, the method does not fully consider the specificity of the pathological change area after the liver and gall operation, for example, complex conditions such as postoperative tissue adhesion, scar formation, local inflammatory reaction and the like, and the adaptability and the accuracy of the segmentation model in treating the pathological change after the liver and gall operation are possibly insufficient. In addition, the scheme lacks modeling capability for the dynamic change rule of the lesion area, and is difficult to meet the accurate quantification requirement for lesion progress in long-term follow-up after operation. On the other hand, a liver biopsy data analysis method and system with the publication number of CN119919312B is disclosed, and the patent is used for enhancing the dynamic quantification capability of pathological features by extracting geometric center points of cell nuclei, blood vessels and fibrosis areas in liver biopsy slice images, calculating spatial distances and connection angles and generating a tissue topology network association parameter set. However, the technical scheme mainly depends on biopsy slice images, has single data source, cannot fully utilize rich information provided by multi-mode image data (such as CT, MRI and the like), and limits the comprehensiveness and accuracy of lesion quantitative analysis. In addition, the spatial distribution characteristics of the lesion area are quantified by the scheme, but the time evolution characteristics of postoperative lesions are not fully considered, so that dynamic change rules of the lesion area at different stages after the operation can be difficult to capture, and the clinical practicability of the evaluation result is affected. The problems show that the existing lesion quantitative evaluation method still has certain defects in the aspects of multi-mode data fusion, complex lesion region segmentation precision, dynamic quantification capability and the like when treating the lesions after the liver and gall operation. Therefore, the invention provides a quantitative evaluation method and a quantitative evaluation system for liver and gall postoperative lesions based on a DLPE algorithm, which aim to realize accurate quantitative evaluation of liver and gall postoperative lesions by integrating multi-mode image data, optimizing a lesion region segmentation model and introducing time evolution characteristic modeling, thereby meeting the clinical requirements of a high-efficiency and intelligent evaluation method. . Disclosure of Invention The invention aims at solving the technical problems that the existing liver and gall postoperative lesion assessment technology has analysis paradigm limitation when facing compound type operations such as postoperative anatomical structure change, multi-pathological factor mixing, disease course dynamic evolution and the like, namely, the existing liver and gall postoperative lesion assessment technology mainly depends on a single image modality and lacks explicit modeling capability on lesion