Eye socket MRI image analysis method and system based on deep learning

Abstract

The invention discloses an orbit MRI image analysis method and system based on deep learning, and belongs to the field of image analys.The method comprises the steps that MRI image data of an orbit area of a patient are obtained, and the MRI image data comprise a conventional T1WI sequence, a T2WI plain scanning sequence, a conventional enhancement sequence and an extraocular muscle fibrillation enhancement sequence; performing preprocessing on the MRI image to obtain a standardized image; inputting the standardized image into a deep learning segmentation model, and outputting segmentation masks of extraocular muscle, lacrimal gland and intraorbital fat; calculating quantitative indexes of a target structure based on the segmentation mask, wherein the target structure comprises extraocular muscle, lacrimal gland and intraorbital fat; and generating a diagnosis report containing the quantitative index. Human experience dependence is avoided, and objective and uniform anatomical structure segmentation results are ensured.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image analysis, and particularly relates to an orbital MRI image analysis method and system based on deep learning. BACKGROUND

[0002] Diagnosis and disease assessment of thyroid-associated ophthalmopathy (TAO) rely on accurate quantification of morphological and functional changes of orbital structures (including extraocular muscles, lacrimal glands and intraorbital fat). At present, multi-modal orbital magnetic resonance imaging (MRI) technology is used in clinical practice to obtain conventional T1WI sequence, T2WI plain sequence, conventional enhancement sequence and extraocular muscle fibrosis enhancement sequence images of the orbital region of patients. The identification and segmentation of target structures in the above MRI images and the measurement of related parameters (such as extraocular muscle volume, intraorbital fat signal ratio, and lacrimal gland enhancement ratio) are all completed by manual operation of doctors, including manually outlining anatomical boundaries, delineating regions of interest (ROI) and calculating quantitative indicators.

[0003] The existing manual processing method has the following significant defects:

[0004] Strong subjectivity: different doctors have different experiences, and the boundaries of extraocular muscles, lacrimal glands and intraorbital fat are not consistent, resulting in deviations in segmentation results and parameter measurement;

[0005] Low efficiency: artificial processing of multi-sequence MRI images of a single patient takes too long and cannot meet the rapid analysis needs of large-scale clinical image data;

[0006] Limited accuracy: under complex anatomical background (such as adhesion of extraocular muscles and fat tissue), it is difficult for artificial segmentation to accurately segment target structures, affecting the reliability of volume calculation and signal intensity ratio measurement. SUMMARY

[0007] To solve the above technical problems, the application provides an orbital MRI image analysis method and system based on deep learning to solve the problems existing in the prior art.

[0008] To achieve the above purpose, in a first aspect, the application provides an orbital MRI image analysis method based on deep learning, comprising:

[0009] Obtaining MRI image data of the orbital region of a patient, the MRI image data including conventional T1WI sequence, T2WI plain sequence, conventional enhancement sequence and extraocular muscle fibrosis enhancement sequence;

[0010] Performing preprocessing on the MRI image to obtain a standardized image;

[0011] The standardized image is input into a deep learning segmentation model, which outputs a segmentation mask for the extraocular muscles, lacrimal glands, and orbital fat.

[0012] The quantitative indicators of the target structure are calculated based on the segmentation mask. The target structure includes extraocular muscles, lacrimal glands, and orbital fat.

[0013] Generate a diagnostic report that includes the aforementioned quantitative indicators.

[0014] Preferably, the process of obtaining the extraocular muscle fibrosis enhancement sequence includes:

[0015] Imaging with a rapid small-angle excitation sequence using inversion recovery, combined with intravenous gadolinium contrast agent injection, and scanning with a scrambled gradient echo-phase-sensitive inversion recovery sequence some time after the injection, was used to obtain an enhanced sequence of extraocular muscle fibrosis.

[0016] Preferably, the process of performing preprocessing on the MRI image includes:

[0017] The resolution of the MRI images is unified through isotropic resampling;

[0018] Based on a deep learning localization network, the anatomical boundaries of the orbit are determined, and the region of interest, which includes the extraocular muscles, lacrimal glands and orbital fat, is cropped.

[0019] Histogram equalization and Z-score normalization are performed on the cropped image.

[0020] Preferably, the deep learning segmentation model is an improved 3D U-Net architecture, comprising:

[0021] The encoder section uses a ResNet-34 backbone network to extract multi-scale features;

[0022] The decoder section integrates a spatial-channel hybrid attention module;

[0023] The output layer is configured with a multi-branch structure to generate binary masks for extraocular muscles, lacrimal glands, and orbital fat, respectively.

[0024] Preferably, the process of the spatial-channel hybrid attention module processing the feature map includes:

[0025] Calculate the spatial weights of the feature map in the spatial dimension;

[0026] Select the important feature channels in the channel dimension of the feature map to obtain the channel weights;

[0027] The spatial weights and channel weights are combined to perform weighted optimization on the feature map.

[0028] Preferably, the quantitative indicators include:

[0029] volume of extraocular muscle, signal ratio of extraocular muscle to specified tissue, extraocular muscle enhancement ratio, extraocular muscle fibrosis rate;

[0030] volume of intraorbital fat, signal ratio of intraorbital fat to specified tissue;

[0031] volume of lacrimal gland, signal ratio of lacrimal gland to specified tissue, lacrimal gland enhancement ratio.

[0032] Preferably, the calculation of the signal ratio comprises:

[0033] defining a standard reference region in the MRI image, and obtaining a baseline signal intensity;

[0034] determining the signal ratio according to the ratio of the signal mean value of the target structure to the mean value of the baseline signal intensity.

[0035] Preferably, the method further comprises training the deep learning segmentation model, and the training step comprises:

[0036] constructing a data set containing expert-labeled extraocular muscle, lacrimal gland and intraorbital fat masks;

[0037] using a weighted combination of Dice loss and cross-entropy loss as the loss function;

[0038] adjusting the parameters of the deep learning segmentation model through an optimization algorithm.

[0039] Preferably, the optimization algorithm uses a TinyKAMMamba 3D segmentation network, and the TinyKAMMamba 3D segmentation network comprises:

[0040] a lightweight Kolmogorov-Arnold mapping module and a state space modeling module;

[0041] introducing a residual block, a deep supervision mechanism and a multi-scale fusion module to the encoder and decoder of the UMamba framework.

[0042] In a second aspect, the present application provides an eye orbit MRI image analysis system based on deep learning, which is used to implement the method of the first aspect, and the system comprises:

[0043] an image acquisition module, which is used to acquire MRI image data of the eye orbit region of a patient, and the MRI image data comprises a conventional T1WI sequence, a T2WI plain scan sequence, a conventional enhancement sequence and an extraocular muscle fibrosis enhancement sequence;

[0044] a preprocessing module, which is used to perform preprocessing on the MRI image to obtain a standardized image;

[0045] an intelligent analysis module configured to input the standardized image into a deep learning segmentation model, and output segmentation masks of the extraocular muscles, lacrimal glands and intraorbital fat;

[0046] a quantitative analysis module configured to calculate quantitative indexes of target structures based on the segmentation masks, the target structures including the extraocular muscles, lacrimal glands and intraorbital fat;

[0047] a report generation module configured to generate a diagnostic report containing the quantitative indexes.

[0048] Compared with the prior art, the present application has the following advantages and technical effects:

[0049] The present application provides an MRI image analysis method for the orbit based on deep learning, which comprises the following steps: first, obtaining MRI image data of the orbit region of a patient, the MRI image data including a conventional T1WI sequence, a T2WI plain scan sequence, a conventional enhancement sequence and an extraocular muscle fibrosis enhancement sequence; second, performing pre-processing on the MRI image to obtain a standardized image; third, inputting the standardized image into a deep learning segmentation model to output segmentation masks of the extraocular muscles, lacrimal glands and intraorbital fat; fourth, calculating quantitative indexes of target structures based on the segmentation masks, the target structures including the extraocular muscles, lacrimal glands and intraorbital fat; and fifth, generating a diagnostic report containing the quantitative indexes.

[0050] To solve the technical problem of the prior art that "manual segmentation leads to inconsistent boundary determination due to differences in doctor's experience", the present application automatically outputs masks of the extraocular muscles, lacrimal glands and intraorbital fat through a deep learning segmentation model, completely avoids reliance on human experience, ensures objective and uniform segmentation results of anatomical structures, and eliminates subjective bias.

[0051] To solve the technical problem of the prior art that "manual processing of multi-sequence MRI images is time-consuming and difficult to meet the rapid needs of clinical practice", the present application automatically processes images to a unified standard and crops ROI, realizes parallel segmentation of multiple target structures through a deep learning model, automatically calculates quantitative indexes instead of manual measurement, significantly shortens the analysis time of single images, and improves the processing efficiency.

[0052] To solve the technical problem of the prior art that "manual segmentation has insufficient precision in complex anatomical backgrounds (such as adhesion of extraocular muscles and fat)", the present application uses an improved 3D U-Net architecture, integrates multi-scale feature extraction and attention mechanism, enhances the perception ability of small targets and edges, processes images based on a unified standard to reduce noise interference, improves the segmentation precision of target structures, and ensures the reliability of quantitative index calculation such as volume and signal ratio.

[0053] The present application automatically generates a structured report to provide reproducible and objective diagnostic evidence for clinical practice, and assists in the determination of TAO active period and the development of treatment plans. Attached Figure Description

[0054] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0055] Figure 1 This is a schematic diagram of a method according to an embodiment of the present invention;

[0056] Figure 2 This is a diagram of a deep learning structure model according to an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of a system according to an embodiment of the present invention. Detailed Implementation

[0058] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0059] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment provides a deep learning-based method for orbital MRI image analysis, including:

[0062] S1. Acquire MRI image data of the patient's orbital region, including conventional T1WI sequence, T2WI plain scan sequence, conventional enhanced sequence, and extraocular muscle fibrosis enhanced sequence;

[0063] Furthermore, the process of obtaining enhanced sequences of extraocular muscle fibrosis includes:

[0064] Imaging with a rapid small-angle excitation sequence using inversion recovery, combined with intravenous gadolinium contrast agent injection, and scanning with a scrambled gradient echo-phase-sensitive inversion recovery sequence some time after the injection, was used to obtain an enhanced sequence of extraocular muscle fibrosis.

[0065] Specifically, this embodiment employs a rapid imaging technique using a fast small-angle excitation sequence with inversion recovery, combined with intravenous gadolinium contrast agent injection. A delayed scan is performed 10-15 minutes after the injection to obtain a delayed-enhanced image. This technique utilizes a coronal scrambled gradient echo-phase-sensitive inversion recovery (FLASH-PSIR) sequence.

[0066] S2, performing preprocessing on the MRI image to obtain a standardized image;

[0067] Further, the process of performing preprocessing on the MRI image comprises:

[0068] unifying the resolution of the MRI image through isotropic resampling;

[0069] determining the anatomical boundaries of the orbit based on a deep learning positioning network, and cropping a region of interest containing the extraocular muscles, lacrimal glands and intraorbital fat;

[0070] performing histogram equalization and Z-score standardization on the cropped image.

[0071] Specifically, the embodiment unifies the image resolution to 1mm 3 .

[0072] S3, inputting the standardized image into a deep learning segmentation model to output a segmentation mask of the extraocular muscles, lacrimal glands and intraorbital fat;

[0073] Further, as shown in the figure, the deep learning segmentation model is a 3D U-Net improved architecture, comprising: Figure 2

[0074] The encoder part adopts a ResNet-34 backbone network to extract multi-scale features;

[0075] The decoder part integrates a space-channel hybrid attention module;

[0076] The output layer is configured with a multi-branch structure to generate binary masks of the extraocular muscles, lacrimal glands and intraorbital fat respectively.

[0077] Specifically, the process of the space-channel hybrid attention module processing the feature map comprises: calculating the spatial weight of the feature map in the spatial dimension; screening important feature channels of the feature map in the channel dimension to obtain a channel weight; and fusing the spatial weight and the channel weight to optimize the feature map.

[0078] S4, calculating a quantitative index of the target structure based on the segmentation mask, the target structure comprising the extraocular muscles, lacrimal glands and intraorbital fat;

[0079] Further, the quantitative index comprises:

[0080] the volume of the extraocular muscles, the signal ratio of the extraocular muscles to a specified tissue, the extraocular muscle reinforcement ratio, and the extraocular muscle fibrosis rate;

[0081] the volume of the intraorbital fat, and the signal ratio of the intraorbital fat to a specified tissue;

[0082] ​The lacrimal gland volume, the signal ratio of the lacrimal gland to the specified tissue, and the lacrimal gland enhancement ratio.

[0083] The calculation of the signal ratio in this embodiment includes: defining a standard reference region in the MRI image to obtain a reference signal intensity; and determining the signal ratio according to the ratio of the signal mean value of the target structure to the mean value of the reference signal intensity.

[0084] Specifically, a standard reference region is specified in the orbital MR image, and a ROI is defined to obtain a reference signal intensity; and the signal intensity of each target structure (extraocular muscle, fat, and lacrimal gland) is standardized, and is represented by a signal ratio.

[0085] The quantitative indicators include the volume and signal intensity ratio (signal ratio = target structure signal mean value / reference signal intensity mean value) of each target structure (extraocular muscle, fat, and lacrimal gland) in the orbit; based on the above quantitative indicators, the clinician can determine whether TAO is in an active period and the severity of the lesion, which is beneficial for the clinician to objectively judge the condition and help develop a treatment strategy according to the condition.

[0086] The experimental data of this embodiment are shown in Table 1.

[0087] Table 1

[0088]

[0089] The embodiment also includes training the deep learning segmentation model, and the training steps include:

[0090] A data set containing expert-labeled extraocular muscle, lacrimal gland, and intraorbital fat masks is constructed;

[0091] A weighted combination of Dice loss and cross-entropy loss is used as the loss function;

[0092] The deep learning segmentation model parameters are adjusted by an optimization algorithm.

[0093] As an innovative implementation, the optimization algorithm uses a TinyKAM Mamba 3D segmentation network, which includes a lightweight Kolmogorov-Arnold mapping module (TinyKAM) and a state space modeling module (Mamba); it is used to improve the representation ability and computational efficiency of the model in high-resolution medical images; the encoder and decoder of the UMamba framework are structurally optimized, residual blocks, deep supervision, and multi-scale fusion are introduced, which significantly improves the perception ability of the model for small targets and edge structures.

[0094] S5, generating a diagnostic report containing the quantitative indicators.

[0095] Embodiment Two

[0096] As Figure 3 shown, the embodiment provides a deep learning-based eye orbit MRI image analysis system for implementing the method of embodiment one, the system comprising:

[0097] an image acquisition module for acquiring MRI image data of the patient's eye orbit region, the MRI image data including conventional T1WI sequence, T2WI plain sequence, conventional enhancement sequence and extraocular muscle fibrosis enhancement sequence;

[0098] a preprocessing module for performing preprocessing on the MRI image to obtain a standardized image;

[0099] an intelligent analysis module for inputting the standardized image into a deep learning segmentation model to output segmentation masks of extraocular muscles, lacrimal glands and intraorbital fat;

[0100] a quantitative analysis module for calculating quantitative indicators of target structures based on the segmentation masks, the target structures including extraocular muscles, lacrimal glands and intraorbital fat;

[0101] a report generation module for generating a diagnostic report containing the quantitative indicators.

[0102] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A deep learning-based method for orbital MRI image analysis, characterized in that, Includes the following steps: Acquire MRI image data of the patient's orbital region, including conventional T1WI sequence, T2WI plain scan sequence, conventional enhanced sequence, and extraocular muscle fibrosis enhanced sequence; The MRI images are preprocessed to obtain standardized images; The standardized image is input into a deep learning segmentation model, which outputs a segmentation mask for the extraocular muscles, lacrimal glands, and orbital fat. The quantitative indicators of the target structure are calculated based on the segmentation mask. The target structure includes extraocular muscles, lacrimal glands, and orbital fat. Generate a diagnostic report that includes the aforementioned quantitative indicators.

2. The method according to claim 1, characterized in that, The process of obtaining enhanced sequences of extraocular muscle fibrosis includes: Imaging with a rapid small-angle excitation sequence using inversion recovery, combined with intravenous gadolinium contrast agent injection, and scanning with a scrambled gradient echo-phase-sensitive inversion recovery sequence some time after the injection, was used to obtain an enhanced sequence of extraocular muscle fibrosis.

3. The method according to claim 1, characterized in that, The process of preprocessing the MRI images includes: The resolution of the MRI images is unified through isotropic resampling; Based on a deep learning localization network, the anatomical boundaries of the orbit are determined, and the region of interest, which includes the extraocular muscles, lacrimal glands and orbital fat, is cropped. Histogram equalization and Z-score normalization are performed on the cropped image.

4. The method according to claim 1, characterized in that, The deep learning segmentation model is an improved 3D U-Net architecture, including: The encoder section uses a ResNet-34 backbone network to extract multi-scale features; The decoder section integrates a spatial-channel hybrid attention module; The output layer is configured with a multi-branch structure to generate binary masks for extraocular muscles, lacrimal glands, and orbital fat, respectively.

5. The method according to claim 4, characterized in that, The spatial-channel hybrid attention module processes feature maps as follows: Calculate the spatial weights of the feature map in the spatial dimension; Select the important feature channels in the channel dimension of the feature map to obtain the channel weights; The spatial weights and channel weights are combined to perform weighted optimization on the feature map.

6. The method according to claim 1, characterized in that, The quantitative indicators include: The volume of extraocular muscles, the signal ratio of extraocular muscles to specified tissues, the enhancement ratio of extraocular muscles, and the fibrosis rate of extraocular muscles; Intraorbital fat volume and signal ratio of intraorbital fat to specified tissue; Lacrimal gland volume, signal ratio of lacrimal gland to designated tissue, and lacrimal gland enhancement ratio.

7. The method according to claim 6, characterized in that, The calculation of the signal ratio includes: Delineate a standard reference region in the MRI image to obtain the baseline signal intensity; The signal ratio is determined by the ratio of the mean signal strength of the target structure to the mean signal strength of the reference structure.

8. The method according to claim 1, characterized in that, The method further includes training the deep learning segmentation model, the training steps of which include: Construct a dataset containing expert-annotated extraocular muscles, lacrimal glands, and orbital fat masks; A weighted combination of Dice loss and cross-entropy loss is used as the loss function; The parameters of the deep learning segmentation model are adjusted by optimizing the algorithm.

9. The method according to claim 8, characterized in that, The optimization algorithm employs the TinyKAMMamba 3D segmentation network, which includes: Lightweight Kolmogorov-Arnold mapping module and state-space modeling module; Residual blocks, deep supervision mechanisms, and multi-scale fusion modules are introduced into the encoder and decoder of the UMamba framework.

10. A deep learning-based orbital MRI image analysis system, characterized in that, The system for implementing the method according to any one of claims 1-9 comprises: The image acquisition module is used to acquire MRI image data of the patient's orbital region, including conventional T1WI sequence, T2WI plain scan sequence, conventional enhanced sequence, and extraocular muscle fibrosis enhanced sequence. The preprocessing module is used to perform preprocessing on the MRI images to obtain standardized images; The intelligent analysis module is used to input the standardized image into the deep learning segmentation model and output a segmentation mask for the extraocular muscles, lacrimal glands and orbital fat. The quantitative analysis module is used to calculate the quantitative indicators of the target structure based on the segmentation mask, the target structure including extraocular muscles, lacrimal glands and orbital fat; The report generation module is used to generate a diagnostic report that includes the quantitative indicators.

Images