Intestinal tract dynamic MRI image enhancement method and system based on deep learning

Abstract

The invention relates to the technical field of image enhancement, and discloses an intestinal dynamic MRI image enhancement method and system based on deep learning, through the image enhancement method based on deep learning, the quality of a small intestine cine-MRI image is improved, the problems of low resolution and poor contrast of a traditional image are improved, and the identifiability of a small intestine intestinal wall structure is enhanced. On the basis, pixel-level registration processing of continuous frame images is combined, the intestinal segment movement track is accurately extracted, and movement parameters such as contraction frequency, lumen diameter variation amplitude, pixel displacement mean value and variance are calculated. Furthermore, a pre-training evaluation model is utilized to intelligently output a small intestine dynamic state judgment result based on motion parameters, and standardized auxiliary recognition of small intestine dynamic disorder diseases and severity of the small intestine dynamic disorder diseases is achieved. According to the method, the image processing quality, the motion feature extraction accuracy and the evaluation intelligence level of small intestine dynamic evaluation are integrally improved, and the method has important clinical application value and popularization significance.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to a method and system for enhancing intestinal dynamic MRI images based on deep learning. Background Technology

[0002] In recent years, magnetic resonance imaging (MRI) has gradually become an important tool for assessing intestinal motility due to its radiation-free nature and high imaging resolution. In particular, cine-MRI technology, through multi-frame imaging at continuous time points, can dynamically observe changes in intestinal peristalsis, providing new possibilities for early disease screening and evaluation of treatment effectiveness.

[0003] However, existing intestinal cine-MRI imaging and analysis methods still have many shortcomings. Due to the inherent complexity and variability of intestinal motility, coupled with signal noise and motion artifacts during MRI scanning, the quality of the acquired raw images is inconsistent, with limited spatial resolution and poor inter-frame continuity, affecting the accuracy of subsequent motility analysis. Furthermore, current intestinal motility analysis largely relies on manual visual assessment, lacking automated and standardized parameter extraction methods and computational processes. This hinders the automatic assessment of small intestinal motility disorders and leaves a lack of effective support systems for clinical decision-making. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for enhancing intestinal dynamic MRI images based on deep learning, so as to solve the technical problem that the current intestinal cine-MRI imaging and analysis process cannot automatically assess small intestinal motility disorders and cannot effectively assist clinical decision-making.

[0005] The first aspect of this invention discloses a deep learning-based method for enhancing dynamic MRI images of the intestine, the method comprising the following steps:

[0006] S1. Acquire cine-MRI images of the subject based on unified imaging parameters; the cine-MRI images are multi-layer cross-sectional images of the small intestine at continuous time points;

[0007] S2. Construct a region of interest (ROI) for the intestinal segment based on the cine-MRI image for motion analysis; the ROI is used to locate and encompass the target intestinal segment.

[0008] S3. Input the cine-MRI image into a pre-trained image enhancement model to perform image enhancement, and output the enhanced image;

[0009] S4. Perform pixel-level registration on the enhanced image to extract the displacement change information of ROI between consecutive frames;

[0010] S5. Calculate the motion parameters of the target intestinal segment based on the displacement change information;

[0011] S6. Input the motion parameters into the pre-trained evaluation model, and output the judgment result of the small intestinal motility status through the evaluation model; the judgment result is used to assist medical staff in analyzing small intestinal motility disorders.

[0012] Furthermore, the image enhancement includes improving image spatial resolution, enhancing contrast, and optimizing motion continuity.

[0013] Furthermore, the construction and training process of the image enhancement model includes the following steps:

[0014] S31. Construct the network structure of the image enhancement model; the network structure of the image enhancement model integrates a super-resolution reconstruction module for improving image spatial resolution, an attention mechanism module for enhancing image contrast, and a three-dimensional convolution module for optimizing the motion continuity between image frames.

[0015] S32. Collect and construct a first training sample set for training the image enhancement model; the first training sample set includes intestinal cine-MRI images and corresponding enhancement feature labels;

[0016] S33. The image enhancement model is trained based on the first training sample set. During the training process, a multi-objective loss function, including image reconstruction loss, contrast preservation loss and inter-frame temporal consistency loss, is used for joint optimization.

[0017] S34. Complete model training and solidify model parameters to create the final image enhancement model.

[0018] Further, step S4 includes the following sub-steps:

[0019] S41. The pixel correspondence between consecutive frames in the enhanced image is matched by an optical flow estimation algorithm to obtain the registration result;

[0020] S42. Calculate the displacement vector of each pixel based on the registration result to form a displacement vector map; the displacement vector is used to represent the motion trajectory of the intestinal wall between consecutive frame images;

[0021] S43. Extract pixel displacement change information within the ROI from the displacement vector diagram.

[0022] Further, step S41 includes the following sub-steps:

[0023] S411. Perform inter-frame pairing processing on the enhanced image sequence to construct pairs of adjacent image frames;

[0024] S412. The constructed image pair is used as the input image pair for optical flow estimation, and the two-dimensional motion vector of each pixel is calculated using the optical flow estimation algorithm.

[0025] S413. The two-dimensional motion vectors are used to form a registration field; the registration field is used to describe the displacement matching relationship of pixels in the image pair, and serves as the basis for subsequent displacement vector maps.

[0026] Furthermore, the motion parameters include contraction frequency, lumen diameter change amplitude, and pixel displacement mean and variance.

[0027] Further, step S5 includes the following sub-steps:

[0028] S51. Based on the displacement change information of ROI between consecutive frame images, count the number of periodic displacement changes per unit time and calculate the contraction frequency of the target intestinal segment.

[0029] S52. Measure the diameter of the intestinal lumen along the direction perpendicular to the intestinal wall in the enhanced image frame, and combine the displacement change information to count the range of change of the intestinal lumen diameter in each image frame to obtain the amplitude of the change in the lumen diameter.

[0030] S53. Calculate the mean and variance of the displacement vectors of all pixels in the ROI to obtain the mean and variance of pixel displacement in the region.

[0031] S54. Use the various motion parameters as the motion feature vector of the target intestinal segment.

[0032] Further, step S6 includes the following sub-steps:

[0033] S61. Input the motion parameters calculated in step S5 as input feature vectors into the pre-trained evaluation model.

[0034] S62. Perform inference operations through the evaluation model and output the small intestinal motility status judgment result of the target intestinal segment based on the input feature vector; the judgment result includes the classification label, severity level and corresponding scoring index of small intestinal motility disorder.

[0035] Furthermore, the construction and training process of the evaluation model includes the following steps:

[0036] S71. Construct the network structure of the evaluation model; the network structure of the evaluation model is a multilayer perceptron neural network.

[0037] S72. Collect and construct a second training sample set; the second training sample set includes small bowel cine-MRI image sequences with known clinical labels and corresponding motion parameters; the clinical labels include disease type classification labels and severity grading labels;

[0038] S73. Based on the second training sample set, the evaluation model is optimized using the cross-entropy loss function to obtain the trained evaluation model.

[0039] The second aspect of this invention discloses a deep learning-based dynamic MRI image enhancement system for the intestine. This system is implemented based on the method disclosed in the first aspect and includes an acquisition module, a region construction module, an image enhancement module, a displacement extraction and calculation module, and an output module; wherein,

[0040] The acquisition module is used to acquire cine-MRI images of the subject based on uniform imaging parameters; the cine-MRI images are multi-layer cross-sectional images of the small intestine at continuous time points;

[0041] The region construction module is used to construct regions of interest (ROIs) for intestinal segments for motion analysis based on the cine-MRI images; the ROIs are used to locate and encompass the target intestinal segment.

[0042] The image enhancement module is used to input the cine-MRI image into a pre-trained image enhancement model for image enhancement and output the enhanced image;

[0043] The displacement extraction and calculation module is used to perform pixel-level registration processing on the enhanced image, extract the displacement change information of ROI between consecutive frames, and calculate the motion parameters of the target intestinal segment based on the displacement change information.

[0044] The output module is used to input the motion parameters into a pre-trained evaluation model, and output the judgment result of the small bowel motility status through the evaluation model; the judgment result is used to assist medical staff in analyzing small bowel motility disorders.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] This invention improves the quality of small intestinal cine-MRI images using a deep learning-based image enhancement method, overcoming the problems of low resolution and poor contrast in traditional images and enhancing the recognizability of the small intestinal wall structure. Building upon this, pixel-level registration processing of consecutive frames is used to accurately extract intestinal segment motion trajectories and calculate motion parameters such as contraction frequency, lumen diameter variation, and mean and variance of pixel displacement. Furthermore, a pre-trained evaluation model intelligently outputs small intestinal motility status assessment results based on these motion parameters, achieving standardized auxiliary identification of small intestinal motility disorders and their severity. This invention comprehensively improves the image processing quality, accuracy of motion feature extraction, and intelligence level of small intestinal motility assessment, overcoming the lack of standardized methods and insufficient motion quantification capabilities in existing technologies, and has significant clinical application value and promotional significance. Attached Figure Description

[0047] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a flowchart illustrating a deep learning-based method for enhancing dynamic MRI images of the intestine, as disclosed in an embodiment of the present invention. Detailed Implementation

[0049] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0050] Example 1

[0051] The first aspect of this invention discloses a deep learning-based method for dynamic MRI image enhancement of the intestine. (See also...) Figure 1 , Figure 1 This is a flowchart illustrating a deep learning-based method for enhancing dynamic MRI images of the intestine, as disclosed in an embodiment of the present invention. The method includes the following steps:

[0052] S1. Acquire cine-MRI images of the subject based on unified imaging parameters; the cine-MRI images are multi-layer cross-sectional images of the small intestine at continuous time points;

[0053] S2. Construct a region of interest (ROI) for the intestinal segment based on the cine-MRI image for motion analysis; the ROI is used to locate and encompass the target intestinal segment.

[0054] S3. Input the cine-MRI image into a pre-trained image enhancement model to perform image enhancement, and output the enhanced image;

[0055] S4. Perform pixel-level registration on the enhanced image to extract the displacement change information of ROI between consecutive frames;

[0056] S5. Calculate the motion parameters of the target intestinal segment based on the displacement change information;

[0057] S6. Input the motion parameters into the pre-trained evaluation model, and output the judgment result of the small intestinal motility status through the evaluation model; the judgment result is used to assist medical staff in analyzing small intestinal motility disorders.

[0058] In this embodiment of the invention, uniform imaging parameters refer to using preset and consistent image acquisition settings during the MRI scans of all subjects to ensure that the imaging results are consistent in terms of spatial resolution, temporal resolution and imaging conditions, so as to reduce data deviations caused by differences in imaging parameters.

[0059] Specifically, standardized imaging parameters include, but are not limited to, sequence type (e.g., using a balanced steady-state free precession sequence), temporal resolution, spatial resolution, number of scan slices, field of view, flip angle, echo time, repetition time, and respiratory control. Typically, subjects need to fast before data acquisition and drink 500-1000 ml of water as needed to fill the small intestine and improve imaging results. After drinking water, wait approximately 30 minutes for the small intestine to reach a stable state before starting data acquisition.

[0060] By setting unified imaging parameters and standardizing scanning procedures, it is possible to ensure that the cine-MRI image data collected from different subjects maintain a high degree of consistency in temporal continuity, spatial coverage, and image quality, thereby improving the accuracy, stability, and repeatability of the analysis results.

[0061] Furthermore, in this embodiment of the invention, the ROI is used to locate and cover the target small intestinal segment region from the cine-MRI images acquired in continuous time frames. The purpose is to ensure that subsequent motility analysis operations are focused on the image portion that is highly correlated with the small intestinal motility state, thereby improving processing efficiency and analysis accuracy.

[0062] Specifically, in each layer of cine-MRI image, the location and extent of the small intestine segment are initially identified by observing changes in image intensity and anatomical structural features. Segments with abundant intraluminal fluid, clearly defined intestinal walls, and continuous traceability are preferentially selected as candidate regions, excluding interference from non-target structures such as the stomach, colon, liver, and adipose tissue. Based on the initial identification, a rectangular or polygonal outline containing the target intestinal segment is drawn manually or semi-automatically as an initial ROI mask. In the semi-automatic method, edge detection algorithms (such as the Canny operator or gradient-based segmentation methods) can be applied to assist in contour extraction, improving the consistency and efficiency of ROI drawing.

[0063] Subsequently, the initial ROI mask is consistently tracked over time. This involves adjusting and replicating the ROI position in consecutive frame images to accommodate slight movements and deformations caused by intestinal peristalsis, ensuring that the ROI accurately covers the target intestinal segment throughout the entire time series. After construction, the ROI regions in all frame images are standardized to unify their size, position, and shape.

[0064] The region of interest (ROI) of the intestinal segment constructed through the above steps can effectively shield the interference of background tissue, other organs and irrelevant noise, so that the motion trajectory of the intestinal wall can be accurately extracted during pixel registration, while improving the calculation accuracy of motion parameters (such as contraction frequency, displacement amplitude, etc.).

[0065] It should be noted that, in this embodiment of the invention, the execution order of steps S2 and S3 can be adjusted according to specific implementation requirements. That is, the enhancement processing of the cine-MRI image can be performed first, and then the region of interest (ROI) of the intestinal segment can be constructed. Alternatively, the ROI can be constructed based on the original cine-MRI image first, and then the ROI image can be enhanced. They can also be performed simultaneously to adapt to different analysis accuracy requirements and computing resource conditions. The execution order is not limited by the step number.

[0066] Furthermore, image enhancement includes improving image spatial resolution, enhancing contrast, and optimizing motion continuity. The process of building and training an image enhancement model includes the following steps:

[0067] S31. Construct the network structure of the image enhancement model; the network structure of the image enhancement model integrates a super-resolution reconstruction module for improving image spatial resolution, an attention mechanism module for enhancing image contrast, and a three-dimensional convolution module for optimizing the motion continuity between image frames.

[0068] S32. Collect and construct a first training sample set for training the image enhancement model; the first training sample set includes intestinal cine-MRI images and corresponding enhancement feature labels;

[0069] S33. The image enhancement model is trained based on the first training sample set. During the training process, a multi-objective loss function, including image reconstruction loss, contrast preservation loss and inter-frame temporal consistency loss, is used for joint optimization.

[0070] S34. Complete model training and solidify model parameters to create the final image enhancement model.

[0071] In this embodiment of the invention, the image enhancement step includes improving the spatial resolution, contrast, and motion continuity of the cine-MRI image based on a deep learning model, so as to improve the quality of the original dynamic MRI image in terms of spatial detail, grayscale contrast, and temporal series coherence, and provide a more accurate and stable image basis for subsequent motion analysis.

[0072] Specifically, in the construction and training of the image enhancement model, the network structure of the image enhancement model is an end-to-end trainable deep neural network, mainly composed of three functional modules: a super-resolution reconstruction module, an attention mechanism module, and a three-dimensional convolution module. Among them, the super-resolution reconstruction module adopts a super-resolution convolutional network structure composed of residual blocks to upsample the input cine-MRI image, improve the spatial resolution of the image, and enhance the discernibility of fine structures such as intestinal wall contours and intraluminal fluid interfaces.

[0073] A spatial attention mechanism is introduced into the attention mechanism module to weight and enhance the features of the intestinal segment region. By adaptively adjusting the response of channel features or spatial features, important information related to intestinal motility is strengthened, while background noise or irrelevant regional features are suppressed, thereby improving the local contrast of the image.

[0074] In the 3D convolution module, a time-series-based 3D convolution structure is used to jointly model continuous frame image sequences, capture inter-frame motion trajectory features, optimize image continuity in the time dimension, reduce inter-frame jitter or breakage artifacts caused by breathing and creeping changes, and improve the overall coherence and smoothness of dynamic images.

[0075] The above modules are connected in series according to their functions within the deep neural network structure to form a unified end-to-end deep augmentation framework.

[0076] In addition, during model training, the first training sample set includes raw intestinal cine-MRI images as input to the model, and enhanced feature labels obtained through high-quality reconstruction, human expert adjustment, or simulation data augmentation are used as the target output. The training data covers the small intestine region under different subjects and different acquisition conditions to improve the model's generalization ability and robustness.

[0077] During training, a multi-objective loss function is used for joint optimization. The image reconstruction loss aims to guide the model output to maintain high fidelity with the enhanced target image at the pixel level; the contrast preservation loss aims to maintain the gray-scale distribution characteristics of the target region during the enhancement process; and the inter-frame temporal consistency loss aims to constrain the coherence and smoothness of the motion trajectory of corresponding pixels between consecutive frames, thereby reducing motion artifacts.

[0078] Through the above-mentioned multi-objective loss collaborative training, the image enhancement model is guaranteed to improve the overall image quality in the spatial, grayscale and temporal dimensions.

[0079] This invention significantly improves the overall quality of original cine-MRI images in terms of spatial detail capture, grayscale differentiation, and dynamic coherence by performing super-resolution reconstruction, attentional contrast enhancement, and three-dimensional temporal continuity optimization. This not only effectively improves the clarity of the interface between the small intestinal wall and its contents but also provides a reliable and standardized data foundation for subsequent pixel-level displacement registration, motion parameter extraction, and small intestinal motility assessment, thereby greatly enhancing the accuracy and automation of intestinal motility disorder analysis.

[0080] In another embodiment of the present invention, the image reconstruction loss is assumed to be:

[0081]

[0082] in, The image reconstruction loss is N; N is the number of training samples. The image of the i-th sample after model enhancement; This is the enhanced target image corresponding to the i-th sample; The mean square error (MSE) of the L2 norm; It serves as a structural similarity index; This is a weighting coefficient, preferably ranging from 0.5 to 0.8, used to balance the contributions of MSE and SSIM to the total reconstruction loss.

[0083] Image reconstruction loss minimizes the pixel-level difference between the enhanced output image and the high-quality target image, and combines structural similarity index (SSIM) constraints to ensure that the enhanced image achieves high standards in both detail fidelity and overall structural restoration. This effectively improves the clarity of minute structures such as intestinal wall boundaries and cavity interfaces, providing accurate spatial references for subsequent motion trajectory extraction.

[0084] The contrast retention loss is:

[0085]

[0086] in, To preserve contrast loss; The local gradient of the i-th enhanced image; Let be the local gradient of the i-th enhanced target image; It is the L2 norm, used to measure the difference in local gradients.

[0087] Contrast preservation loss enhances the gray-scale contrast features of the intestinal segment region by minimizing the difference in local gradient changes between the enhanced image and the target image, suppresses background noise and irrelevant tissue signals, improves the distinguishability between intestinal wall motion features and surrounding tissues, ensures the visual recognizability of key anatomical features during enhancement, and contributes to the stability and accuracy of motion parameter extraction.

[0088] The inter-frame temporal consistency loss is:

[0089]

[0090] in, The loss is the inter-frame temporal consistency loss; M is the number of image frames in the training sequence. An enhanced image at time step t; Enhanced image at time step t+1: The optical flow field between time t and t+1; The image obtained after registration based on the optical flow field; This represents the local gradient of the optical flow field, used to measure the smoothness of changes in optical flow. The value is the optical flow regularization weight coefficient, preferably 0.1 to 0.5, used to balance inter-frame consistency and optical flow smoothness.

[0091] By calculating the registration difference of consecutive frames based on optical flow guidance and introducing optical flow field smoothing regularization, the temporal jitter and sequence breakage caused by motion artifacts, acquisition errors or enhancement processes are effectively suppressed, significantly improving the coherence and smoothness of the cine-MRI image sequence in the temporal dimension, laying a reliable foundation for subsequent pixel-level registration and dynamic motion analysis.

[0092] Finally, the image reconstruction loss, contrast preservation loss, and inter-frame temporal consistency loss are weighted and summed to obtain the total loss. Through the joint optimization of the above three types of sub-losses, this invention can improve the local contrast of intestinal segments without sacrificing spatial detail clarity, while maintaining the natural coherence of motion trajectories in consecutive frames, thus forming high-quality, highly consistent standardized cine-MRI image data.

[0093] Further, step S4 includes the following sub-steps:

[0094] S41. The pixel correspondence between consecutive frames in the enhanced image is matched by an optical flow estimation algorithm to obtain the registration result;

[0095] S42. Calculate the displacement vector of each pixel based on the registration result to form a displacement vector map; the displacement vector is used to represent the motion trajectory of the intestinal wall between consecutive frame images;

[0096] S43. Extract pixel displacement change information within the ROI from the displacement vector diagram.

[0097] Further, step S41 includes the following sub-steps:

[0098] S411. Perform inter-frame pairing processing on the enhanced image sequence to construct pairs of adjacent image frames;

[0099] S412. The constructed image pair is used as the input image pair for optical flow estimation, and the two-dimensional motion vector of each pixel is calculated using the optical flow estimation algorithm.

[0100] S413. The two-dimensional motion vectors are used to form a registration field; the registration field is used to describe the displacement matching relationship of pixels in the image pair, and serves as the basis for subsequent displacement vector maps.

[0101] In this embodiment of the invention, in order to accurately extract the continuous motion trajectory of the small intestinal wall in the cine-MRI image sequence, pixel-level registration processing is performed on the enhanced image sequence in step S4, which specifically includes steps S41 to S43. Specifically, in step S41, the pixel correspondence between consecutive frames in the enhanced image sequence is matched using an optical flow estimation algorithm to generate registration results at consecutive time points.

[0102] First, inter-frame pairing is performed on the enhanced image sequence. For each frame in the time series, its next adjacent frame is selected as the pairing target, constructing a set of image pairs. To improve the continuity of pairing and the integrity of intestinal segment movement trajectories, a region of interest (ROI) continuity check is introduced during the pairing process. This ensures the spatial continuity of the intestinal segment ROI in each image pair, avoiding ROI mismatch issues caused by excessive movement amplitude or image truncation. The paired image pairs will serve as the input basis for optical flow estimation, ensuring that the registration process closely revolves around the small intestinal segment region.

[0103] Subsequently, the constructed image pairs are input into the optical flow estimation algorithm to perform pixel-level motion estimation. To adapt to the weak texture, local occlusion, and small-amplitude non-rigid motion characteristics present in small intestinal cine-MRI images, this invention employs multiple strategies in the optical flow estimation process to improve estimation accuracy.

[0104] Specifically, at the input end, structural feature maps output from the intermediate feature layer of the image enhancement model are extracted and combined with the enhanced original grayscale image to form the input features for optical flow estimation. Grayscale information primarily preserves overall brightness consistency, while the feature maps enhance local structural details such as intestinal wall boundaries and lumen contours. Dual-channel feature fusion effectively improves pixel matching stability in weakly textured regions. During optical flow energy optimization, a ROI mask constraint is introduced. Optical flow estimation only performs pixel displacement matching within the ROI mask area; high-weight smoothing regularization is applied to areas outside the mask, or computation is directly ignored, suppressing the interference of background tissue motion on the registration results. Simultaneously, an adaptive regularization mechanism based on local texture awareness is incorporated into the displacement field smoothing term design. Specifically, the optical flow regularization weights are dynamically adjusted based on the texture intensity distribution within each pair of input image local regions: in texture-rich regions (such as intestinal wall edges), smoothing intensity is reduced to preserve details; in texture-weakened regions (such as inside cavities or signal attenuation areas), smoothing regularization is enhanced to prevent noise amplification or spurious motion estimation. Through the above multiple strategies, the generated optical flow field can accurately depict the real movement of intestinal segments and has good spatial continuity and noise suppression capabilities.

[0105] After optical flow estimation is completed, in step S413, the calculated two-dimensional pixel motion vectors of each image pair are organized into a registration field. The registration field adopts the form of a dense displacement vector field, where each pixel position corresponds to a two-dimensional displacement vector (Δx, Δy), describing the displacement change from the current frame to the next frame. To ensure the temporal continuity between registration fields, smooth connections are made based on the differences in flow fields between frames when organizing the registration fields, avoiding breakage of motion trajectories caused by local abnormal motion or intensity changes. The generated continuous frame registration fields will serve as direct input data for subsequent displacement vector calculation (step S42), laying the foundation for accurately recovering the dynamic motion pattern of the small intestinal wall.

[0106] In step S42, the displacement vectors of each pixel in the consecutive frame images are calculated based on the generated registration field to form a displacement vector map. Specifically, the continuous motion trajectory features of the small intestine segment during the observation period are quantified by accumulating pixel displacement changes frame by frame. To ensure the specificity and accuracy of the extraction results, ROI masking is applied during the displacement vector map generation process, retaining only the pixel displacement data within the region of interest and removing background and irrelevant motion information to ensure that subsequent motion feature analysis focuses on the target intestinal segment region.

[0107] Finally, in step S43, pixel displacement change information within the ROI is further extracted from the displacement vector diagram. By combining ROI masking for localization filtering, the displacement amplitude and direction changes of pixels within the target intestinal segment are recorded frame by frame. The extracted pixel motion trajectory data within the ROI will directly serve as the input basis for subsequent motion parameter calculations, providing reliable support for the auxiliary identification and severity grading of small intestinal motility disorders.

[0108] Through the above processing, this invention fully integrates innovative strategies such as depth enhancement feature assistance, ROI mask region limitation, and local texture perception adaptive regularization in the pixel-level motion extraction stage, which significantly improves the accuracy and stability of small intestinal segment motion extraction and improves the overall accuracy and clinical practical value of small intestinal motility assessment.

[0109] It should be noted that, in this invention, the optical flow estimation used in the inter-frame temporal consistency loss introduced during the image enhancement model training process is an auxiliary loss design in the training phase. Its main function is to guide the image enhancement model to maintain the consistency and smoothness of image content across frames in the temporal dimension by calculating the optical flow field between consecutive frames, thereby improving the coherence and visual stability of dynamic image sequences. The optical flow estimation is only used as a latent variable for temporary calculation during training, participating in the calculation of the loss function and model weight optimization; it is not used as the final output or saved as data after training is completed.

[0110] In the inference phase following image enhancement, pixel-level registration is performed on the enhanced cine-MRI image sequence. This pixel registration is based on an optical flow estimation algorithm, aiming to accurately extract displacement changes in intestinal wall pixels between consecutive frames, providing data support for subsequent motion parameter calculations. Therefore, optical flow estimation in the image enhancement phase and optical flow estimation in the pixel registration phase are different applications of the training and inference phases, respectively. They differ in their functional objectives, calculation results, and data usage, and do not conflict with each other.

[0111] Furthermore, the motion parameters include contraction frequency, lumen diameter variation amplitude, and mean and variance of pixel displacement.

[0112] Step S5 includes the following sub-steps:

[0113] S51. Based on the displacement change information of ROI between consecutive frame images, count the number of periodic displacement changes per unit time and calculate the contraction frequency of the target intestinal segment.

[0114] S52. Measure the diameter of the intestinal lumen along the direction perpendicular to the intestinal wall in the enhanced image frame, and combine the displacement change information to count the range of change of the intestinal lumen diameter in each image frame to obtain the amplitude of the change in the lumen diameter.

[0115] S53. Calculate the mean and variance of the displacement vectors of all pixels in the ROI to obtain the mean and variance of pixel displacement in the region.

[0116] S54. Use the various motion parameters as the motion feature vector of the target intestinal segment.

[0117] In this embodiment of the invention, in order to quantitatively evaluate the motion characteristics of the small intestinal segment, based on the pixel displacement change information within the continuous inter-frame ROI extracted in step S4, step S5 is performed to calculate and analyze the motion parameters.

[0118] Specifically, firstly, a time series analysis is performed on the overall average displacement of pixels within the ROI region at continuous time points to identify periodic peak changes in the displacement curve. Each contraction corresponds to a peak or trough change in the average displacement. Based on a set amplitude threshold and time interval standard, the number of effective cycles occurring within a certain time window is counted. The counted number of effective cycles is divided by the corresponding time length to obtain the contraction frequency per unit time, which serves as a key dynamic indicator reflecting the peristaltic rhythm of the small intestine.

[0119] In step S52, multiple straight measurement paths perpendicular to the local intestinal wall are extracted based on the ROI region in each frame image. Pixel intensity changes are sampled along the path direction, and the positions of the inner and outer boundaries of the intestinal lumen are detected by intensity gradient detection. After initially identifying the intestinal lumen boundary positions based on intensity gradient detection, the positional offset of the boundary pixels over time is calibrated using the displacement trajectory of pixels within the ROI, further improving the accuracy of diameter measurement. The maximum and minimum diameters after calibration in consecutive frames are combined to statistically analyze the range of intestinal lumen diameter changes, serving as an indicator reflecting the amplitude of intestinal wall contraction and relaxation.

[0120] In step S53, the pixel displacement mean is calculated by averaging the cumulative displacement vectors of all pixels within the ROI over a continuous time period, reflecting the average displacement trend of the overall movement of the intestinal segment. The pixel displacement variance, on the other hand, is calculated by determining the degree of dispersion of pixel displacements from the mean, quantitatively reflecting the consistency and spatial differences in movement within the intestinal segment. A lower displacement variance indicates good overall coordination of intestinal wall movement, while a higher displacement variance may suggest local functional abnormalities or inconsistencies in movement.

[0121] Finally, the calculated motion parameters—contraction frequency, lumen diameter change amplitude, mean pixel displacement, and variance of pixel displacement—are used to form the motion feature vector of the target intestinal segment. This motion feature vector serves as the input feature for subsequent small intestinal motility assessment models, comprehensively reflecting the frequency characteristics, deformation amplitude, overall motility volume, and motility consistency of the intestinal segment. It can provide multi-dimensional, quantitative, and standardized data support for the auxiliary identification and severity grading of small intestinal motility disorders.

[0122] Through the above operations, this invention not only achieves accurate quantification of the continuous dynamic motion characteristics of small intestinal segments, but also effectively integrates multiple key dimensions in intestinal dynamics analysis, laying a solid data foundation for subsequent intelligent assessment of small intestinal motility based on deep learning evaluation models.

[0123] Further, step S6 includes the following sub-steps:

[0124] S61. Input the motion parameters calculated in step S5 as input feature vectors into the pre-trained evaluation model.

[0125] S62. Perform inference operations through the evaluation model and output the small intestinal motility status judgment result of the target intestinal segment based on the input feature vector; the judgment result includes the classification label, severity level and corresponding scoring index of small intestinal motility disorder.

[0126] Preferably, step S62 includes the following sub-steps:

[0127] S621. Input the motion parameter vector into the feature encoding layer of the evaluation model. The feature encoding layer is used to perform nonlinear transformation on the input features and extract discriminative features.

[0128] S622. The output of the feature encoding layer is input to the multi-task inference module, the multi-task inference module comprising:

[0129] The disease classification submodule is used to output classification labels for small intestinal motility disorders. It outputs the probability distribution of each small intestinal motility disorder category through a Softmax classifier and selects the category with the highest probability as the classification label.

[0130] The severity grading submodule is used to output the severity level of small bowel motility disorder. It outputs the probability distribution of each severity level through an independent Softmax grader and selects the level corresponding to the highest probability as the severity grading result.

[0131] The scoring index generation submodule is used to calculate the scoring values ​​corresponding to the classification labels and severity levels. Based on the classification labels, severity levels and their corresponding prediction confidence levels output by the evaluation model, a comprehensive scoring index is calculated. The comprehensive scoring index is used to quantify the degree of abnormal intestinal motility.

[0132] S623. Output the classification labels, severity levels, and scoring indicators as the judgment results of small intestinal motility status, and generate structured reasoning output to assist medical staff in identification and classification decisions.

[0133] Through the above steps, combined with a multi-task reasoning process involving feature encoding, disease classification, severity grading, and scoring, this invention can efficiently and accurately identify the types of small intestinal motility disorders, quantify the degree of motility abnormalities, and assist in the standardized and intelligent clinical assessment of small intestinal motility status.

[0134] Understandably, in order to further quantify the degree of abnormality in small intestinal motility and provide a more granular intelligent assessment index than disease classification labels and severity levels, this invention sets up a scoring index generation submodule in the multi-task reasoning module.

[0135] The scoring indicators are expressed as continuous scores, usually set in a standardized range of 0 to 100. Higher scores indicate better small bowel motility, while lower scores suggest more severe motility dysfunction.

[0136] Specifically, the scoring indicator generation submodule takes into account the following three aspects:

[0137] First, based on the Softmax predicted probabilities of each disease category output by the disease classification submodule, the maximum category probability is extracted to measure the confidence level of the disease classification judgment. Second, based on the Softmax predicted probabilities of each severity level output by the severity grading submodule, the maximum level probability is extracted to measure the confidence level of the severity judgment. Furthermore, based on the latent feature vector output by the feature encoding layer, its similarity score relative to the distribution center of normal small intestinal motility features is calculated to quantify the degree of closeness between the feature level and the normal state. These multi-source information are then combined and weighted for summation. In this way, the scoring index can comprehensively reflect the three aspects of classification reliability, grading accuracy, and feature similarity, forming a continuous and quantitative assessment of the degree of small intestinal motility dysfunction.

[0138] During the reasoning process, the scoring indicator generation submodule automatically calculates the comprehensive score and combines it with the category label and severity level to form a structured reasoning output. The output includes disease category, severity level, and comprehensive score fields, and supports standard formats such as JSON and CSV, facilitating automatic reading, display, and subsequent decision support by healthcare systems.

[0139] By introducing scoring indicators, this invention not only achieves standardized classification and grading of small intestinal motility status, but also realizes fine-grained risk quantification based on continuous indicators, providing clinicians with richer and more accurate intelligent auxiliary information, and improving the intelligence level and clinical applicability of the small intestinal motility disorder assessment system.

[0140] Furthermore, the evaluation model construction and training process includes the following steps:

[0141] S71. Construct the network structure of the evaluation model; the network structure of the evaluation model is a multilayer perceptron neural network.

[0142] S72. Collect and construct a second training sample set; the second training sample set includes small bowel cine-MRI image sequences with known clinical labels and corresponding motion parameters; the clinical labels include disease type classification labels and severity grading labels;

[0143] S73. Based on the second training sample set, the evaluation model is optimized using the cross-entropy loss function to obtain the trained evaluation model.

[0144] Specifically, in this embodiment of the invention, in order to achieve intelligent assessment and severity grading of small intestinal motility, an assessment model needs to be constructed and trained based on a dataset with known clinical labels.

[0145] In step S71, the evaluation model employs a multi-layer perceptron (MLP) structure, specifically comprising several fully connected layers and activation units. Layers are connected via weight matrices, and a nonlinear activation function, such as ReLU (Rectified Linear Unit) or LeakyReLU, is applied after each layer to introduce nonlinear mapping capabilities into the feature space. The model's input receives motion parameter feature vectors, which undergo high-dimensional nonlinear transformation through a feature encoding layer to extract more discriminative feature representations. At the output, a multi-task inference module performs disease classification, severity grading, and comprehensive score calculation. This approach balances feature extraction capabilities with inference output requirements across multiple tasks, ensuring the model's generalization ability and inference accuracy in complex clinical scenarios.

[0146] In step S72, the second training sample set consists of small bowel cine-MRI image sequences with known clinical labels and corresponding motion parameter data. Specifically, by collecting small bowel cine-MRI image sequences from clinical patients, and combining image enhancement, pixel-level registration, displacement vector extraction, and motion parameter calculation processes, standardized motion parameter feature vectors are generated. Simultaneously, based on the patients' clinical diagnoses, corresponding disease type classification labels (e.g., chronic pseudo-obstruction) and severity grading labels (e.g., mild, moderate, severe) are labeled. Through systematic data collection and labeling, the second training sample set ensures rich disease heterogeneity and diverse motor functions, meeting the needs of supervised training of the evaluation model and improving the model's adaptability to real clinical samples.

[0147] During training, the cross-entropy loss function is used as the primary optimization objective to measure the difference between the model's output class distribution and the true labels. Specifically, the disease classification submodule and the severity grading submodule independently calculate the cross-entropy loss. The parameters are updated based on the cross-entropy loss using the backpropagation algorithm, and the network weights of the evaluation model are iteratively optimized to finally obtain a well-trained evaluation model with good generalization performance on the small intestine motility state recognition and grading tasks.

[0148] Through the above operations, this invention can automatically classify and identify the small bowel motility status, determine the severity level, and quantify the comprehensive score based on standardized motility parameter feature vectors. This model not only improves the objectivity and standardization of small bowel motility disorder assessment but also significantly reduces reliance on manual analysis and optimizes clinical workflow.

[0149] Example 2

[0150] A second aspect of this invention discloses a deep learning-based dynamic MRI image enhancement system for the intestine. This system includes an acquisition module, a region construction module, an image enhancement module, a displacement extraction and calculation module, and an output module; wherein,

[0151] The acquisition module is used to acquire cine-MRI images of the subject based on uniform imaging parameters; the cine-MRI images are multi-layer cross-sectional images of the small intestine at continuous time points;

[0152] The region construction module is used to construct regions of interest (ROIs) for intestinal segments for motion analysis based on the cine-MRI images; the ROIs are used to locate and encompass the target intestinal segment.

[0153] The image enhancement module is used to input the cine-MRI image into a pre-trained image enhancement model for image enhancement and output the enhanced image;

[0154] The displacement extraction and calculation module is used to perform pixel-level registration processing on the enhanced image, extract the displacement change information of ROI between consecutive frames, and calculate the motion parameters of the target intestinal segment based on the displacement change information.

[0155] The output module is used to input the motion parameters into a pre-trained evaluation model, and output the judgment result of the small bowel motility status through the evaluation model; the judgment result is used to assist medical staff in analyzing small bowel motility disorders.

[0156] It should be noted that the specific implementation process of Example 2 is similar to that of Example 1, and will not be repeated in Example 2.

[0157] Finally, it should be noted that the above-described embodiments include multiple parallel implementations of the present invention. Deleting or otherwise adjusting one or more of these implementations will not affect the implementation of the solution. Furthermore, the deep learning-based intestinal dynamic MRI image enhancement method and system disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for enhancing dynamic MRI images of the intestine, characterized in that, The method includes the following steps: S1. Acquire cine-MRI images of the subject based on unified imaging parameters; the cine-MRI images are multi-layer cross-sectional images of the small intestine at continuous time points; S2. Construct a region of interest (ROI) for the intestinal segment based on the cine-MRI image for motion analysis; the ROI is used to locate and encompass the target intestinal segment. S3. Input the cine-MRI image into a pre-trained image enhancement model to perform image enhancement, and output the enhanced image; S4. Perform pixel-level registration on the enhanced image to extract the displacement change information of ROI between consecutive frames; S5. Calculate the motion parameters of the target intestinal segment based on the displacement change information; S6. Input the motion parameters into the pre-trained evaluation model, and output the judgment result of the small intestinal motility status through the evaluation model; the judgment result is used to assist medical staff in analyzing small intestinal motility disorders.

2. The method for enhancing intestinal dynamic MRI images based on deep learning according to claim 1, characterized in that, The image enhancement includes improving image spatial resolution, enhancing contrast, and optimizing motion continuity.

3. The deep learning-based method for enhancing dynamic MRI images of the intestine according to claim 2, characterized in that, The process of constructing and training the image enhancement model includes the following steps: S31. Construct the network structure of the image enhancement model; the network structure of the image enhancement model integrates a super-resolution reconstruction module for improving image spatial resolution, an attention mechanism module for enhancing image contrast, and a three-dimensional convolution module for optimizing the motion continuity between image frames. S32. Collect and construct a first training sample set for training the image enhancement model; the first training sample set includes intestinal cine-MRI images and corresponding enhancement feature labels; S33. The image enhancement model is trained based on the first training sample set. During the training process, a multi-objective loss function, including image reconstruction loss, contrast preservation loss and inter-frame temporal consistency loss, is used for joint optimization. S34. Complete model training and solidify model parameters to create the final image enhancement model.

4. The method for enhancing intestinal dynamic MRI images based on deep learning according to claim 1, characterized in that, Step S4 includes the following sub-steps: S41. The pixel correspondence between consecutive frames in the enhanced image is matched by an optical flow estimation algorithm to obtain the registration result; S42. Calculate the displacement vector of each pixel based on the registration result to form a displacement vector map; the displacement vector is used to represent the motion trajectory of the intestinal wall between consecutive frame images; S43. Extract pixel displacement change information within the ROI from the displacement vector diagram.

5. The deep learning-based method for enhancing dynamic MRI images of the intestine according to claim 4, characterized in that, Step S41 includes the following sub-steps: S411. Perform inter-frame pairing processing on the enhanced image sequence to construct pairs of adjacent image frames; S412. The constructed image pair is used as the input image pair for optical flow estimation, and the two-dimensional motion vector of each pixel is calculated using the optical flow estimation algorithm. S413. The two-dimensional motion vectors are used to form a registration field; the registration field is used to describe the displacement matching relationship of pixels in the image pair, and serves as the basis for subsequent displacement vector maps.

6. The method for enhancing intestinal dynamic MRI images based on deep learning according to claim 1, characterized in that, The motion parameters include contraction frequency, lumen diameter change amplitude, and pixel displacement mean and variance.

7. The deep learning-based method for enhancing dynamic MRI images of the intestine according to claim 6, characterized in that, Step S5 includes the following sub-steps: S51. Based on the displacement change information of ROI between consecutive frame images, count the number of periodic displacement changes per unit time and calculate the contraction frequency of the target intestinal segment. S52. Measure the diameter of the intestinal lumen along the direction perpendicular to the intestinal wall in the enhanced image frame, and combine the displacement change information to count the range of change of the intestinal lumen diameter in each image frame to obtain the amplitude of the change in the lumen diameter. S53. Calculate the mean and variance of the displacement vectors of all pixels in the ROI to obtain the mean and variance of pixel displacement in the region. S54. Use the various motion parameters as the motion feature vector of the target intestinal segment.

8. The method for enhancing intestinal dynamic MRI images based on deep learning according to claim 1, characterized in that, Step S6 includes the following sub-steps: S61. Input the motion parameters calculated in step S5 as input feature vectors into the pre-trained evaluation model. S62. Perform inference operations through the evaluation model and output the small intestinal motility status judgment result of the target intestinal segment based on the input feature vector; the judgment result includes the classification label, severity level and corresponding scoring index of small intestinal motility disorder.

9. The deep learning-based method for enhancing dynamic MRI images of the intestine according to claim 8, characterized in that, The process of constructing and training the evaluation model includes the following steps: S71. Construct the network structure of the evaluation model; the network structure of the evaluation model is a multilayer perceptron neural network. S72. Collect and construct a second training sample set; the second training sample set includes small bowel cine-MRI image sequences with known clinical labels and corresponding motion parameters; the clinical labels include disease type classification labels and severity grading labels; S73. Based on the second training sample set, the evaluation model is optimized using the cross-entropy loss function to obtain the trained evaluation model.

10. A deep learning-based intestinal dynamic MRI image enhancement system, said system being implemented based on the method described in any one of claims 1-9, characterized in that, The system includes an acquisition module, a region construction module, an image enhancement module, a displacement extraction and calculation module, and an output module; wherein, The acquisition module is used to acquire cine-MRI images of the subject based on uniform imaging parameters; the cine-MRI images are multi-layer cross-sectional images of the small intestine at continuous time points; The region construction module is used to construct regions of interest (ROIs) for intestinal segments for motion analysis based on the cine-MRI images; the ROIs are used to locate and encompass the target intestinal segment. The image enhancement module is used to input the cine-MRI image into a pre-trained image enhancement model for image enhancement and output the enhanced image; The displacement extraction and calculation module is used to perform pixel-level registration processing on the enhanced image, extract the displacement change information of ROI between consecutive frames, and calculate the motion parameters of the target intestinal segment based on the displacement change information. The output module is used to input the motion parameters into a pre-trained evaluation model, and output the judgment result of the small bowel motility status through the evaluation model; the judgment result is used to assist medical staff in analyzing small bowel motility disorders.

Images