MRI bone segmentation method and system based on ct domain pseudo-label active learning

Abstract

The application provides a kind of MRI bone segmentation method and system based on CT domain pseudo label active learning, belong to medical image processing technical field.The method includes: using paired CT-MRI training MRI to CT image translation network, adopts the bone pseudo label generation mechanism of double path to generate the bone pseudo label of MRI same space and constructs pseudo label pool, high-quality pseudo label is used as weakly supervised sample to participate in the initial training and subsequent iteration training of MRI bone segmentation model;The label confirmed or corrected by doctor end enters the labeled pool and is used as strongly supervised sample to participate in training;Based on pseudo label pool and labeled pool, the MRI bone segmentation model is iteratively trained using a closed-loop active learning strategy until the final MRI bone segmentation model is obtained.The application effectively solves the problems of weak MRI bone structure signal, high artificial labeling cost and low model training efficiency.The final deployed model only needs to input MRI to output accurate bone segmentation results.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence, specifically to an MRI bone segmentation method and system based on active learning of CT domain pseudo-labels. Background Technology

[0002] In modern medicine, magnetic resonance imaging (MRI) is increasingly used for procedures such as punctures, biopsies, ablation surgeries, and preoperative pathway planning. In these applications, physicians need to accurately determine the spatial relationships between soft tissues, blood vessels, and bony structures (such as ribs, sternum, spine, and pelvis) on MRI. Bone structures are not only important anatomical references for three-dimensional navigation and positioning, but they can also become physical obstacles obstructing the path of the puncture needle. Therefore, achieving rapid and precise segmentation of bone structures on MRI is of significant value.

[0003] However, unlike computed tomography (CT) images where bone appears as a high signal with well-defined boundaries, in conventional MRI sequences, the bone cortex typically appears as a low signal or lacks signal, resulting in blurred bone boundaries. This is especially true for slender ribs, complex vertebral appendages, the sternal margin, and pelvic structures, which are prone to fractures, adhesions, or complete omissions in MRI images. This presents a significant challenge to MRI-based automated bone segmentation.

[0004] Currently, obtaining a large amount of high-quality MRI bone annotation data is the main bottleneck in training high-performance segmentation models. Directly having experienced physicians delineate multiple categories of bone labels layer by layer on 3D MRI images is an extremely time-consuming, costly, and inconsistent work among annotators.

[0005] On the other hand, bone structures are clearly visible in CT images, and existing CT bone segmentation models (such as TotalSegmentator-CT) can achieve very reliable segmentation accuracy. An intuitive idea is whether the clear annotations from CT can be used to assist MRI segmentation. Current techniques attempt to map CT labels from the same patient to MRI space using image registration. However, even if the CT and MRI images are from the same patient, the same treatment cycle, or acquired at similar times, achieving pixel-level precise alignment of bone structures remains very difficult due to factors such as imaging principles, patient positioning, and organ deformation. Registration errors directly lead to label noise, severely affecting the training effect of the final model.

[0006] In addition, there are some general MRI multi-structure segmentation models, but due to the lack of large-scale, fine bone annotation data, their segmentation effect on bone structures, especially in distinguishing details and instances, is often unsatisfactory.

[0007] Therefore, how to effectively supervise the training of MRI bone segmentation models by utilizing the mature segmentation capabilities of the CT domain without relying on a large amount of manual annotation and imperfect image registration, so as to achieve high-precision and high-robust MRI bone segmentation at low annotation cost, has become an urgent technical problem to be solved. Summary of the Invention

[0008] The purpose of this invention is to provide an MRI bone segmentation method and system based on CT domain pseudo-label active learning, so as to solve the technical problems of weak MRI bone structure signals, high cost of manual annotation, and low model training efficiency in the prior art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides an MRI bone segmentation method based on CT domain pseudo-label active learning, the method comprising the following steps: S1. Obtain paired CT-MRI datasets and train an MRI-to-CT image translation network based on these datasets; S2. A dual-path bone pseudo-label generation mechanism is used to generate bone pseudo-labels in the same MRI space, wherein: For MRI samples with paired CT scans, MRI spatial bone pseudo-labels are generated directly using paired CT labels or CT bone segmentation results, after spatial registration, resampling, and quality control. For MRI samples without paired CT images, a virtual CT image in the same space is generated using the MRI-to-CT image translation network; and a pre-trained CT bone segmentation model is used to segment the virtual CT image to generate bone pseudo-labels in the same space as the MRI. A pseudo-label pool is constructed based on bone pseudo-labels generated by dual paths, and quality control and confidence assessment are performed to screen out high-confidence pseudo-labels. These high-confidence pseudo-labels are used as weakly supervised samples in the initial training and subsequent iterative training of the MRI bone segmentation model. Labels confirmed or corrected by doctors are added to the labeled pool and used as strongly supervised samples in training. S3. Based on the pseudo-label pool and the labeled pool, a closed-loop active learning strategy is used to iteratively train the MRI bone segmentation model. During training, the MRI bone segmentation model outputs predicted labels for candidate MRIs, calculates the inconsistency between the predicted labels and pseudo-labels / labeled labels, filters difficult samples and sends them to the doctor for discarding, confirmation or correction, and adds the confirmed or corrected labels to the labeled pool; iterative training continues until the preset convergence condition is met, and the final MRI bone segmentation model is obtained. S4. Use the final MRI bone segmentation model to segment the MRI image to be segmented into bone structures and output the segmentation results.

[0011] Furthermore, in step S1, the MRI-to-CT image translation network adopts a 3D ResUNet-L network, whose training loss function includes image similarity loss, discriminator loss, and geometric consistency loss.

[0012] Furthermore, in step S2, before generating a virtual CT image in the same space using the MRI-to-CT image translation network, the method further includes: performing spatial standardization and grayscale normalization preprocessing on the MRI samples to eliminate data distribution differences caused by acquisition parameters of different MRI devices.

[0013] Furthermore, in step S2, the quality control specifically includes at least one of the following: small connected domain removal, hole filling, left-right symmetry inspection, rib continuity inspection, vertebral sequence continuity inspection, and anatomical adjacency inspection.

[0014] Further, in step S2, the confidence assessment specifically includes: Extract the intermediate layer feature map or output probability map of the pre-trained CT bone segmentation model when generating the bone pseudo-label; Based on the intermediate layer feature map or output probability map, calculate the segmentation uncertainty measure at the voxel level; A confidence score is generated based on the uncertainty measure, and a confidence threshold is set. Pseudo-labels that are higher than the threshold are marked as high-confidence pseudo-labels.

[0015] Furthermore, in step S3, the predicted labels are sorted according to the prediction uncertainty, the difference with pseudo-labels / labeled labels, the quality risk of pseudo-labels and sample diversity. Difficult samples are selected and pushed to the doctor for discarding, confirmation or correction. Confirmed / corrected labels are added to the labeled MRI dataset and participate in the next round of training with preset weights. Discarded samples are excluded from the current training queue and the reason is recorded.

[0016] Secondly, the present invention also provides an MRI bone segmentation system based on CT domain pseudo-label active learning. The system includes: an MRI to CT image translation module, a CT bone segmentation pseudo-label module, a pseudo-label quality control module, an MRI bone segmentation module, and an active learning interaction module. The system applies the above-mentioned method to obtain the final MRI bone segmentation model, performs bone structure segmentation on the MRI image to be segmented, and outputs the segmentation results.

[0017] Thirdly, the present invention also provides an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the above-described method.

[0018] As can be seen from the above technical solution, the present invention provides an MRI bone segmentation scheme based on CT domain pseudo-label active learning, and its beneficial effects compared with the prior art include: 1. A training framework was constructed that combines cross-modal pseudo-label generation, doctor-end confirmation / refinement data pooling, and active learning iteration of 3D V-Net predicted labels. This significantly reduced the reliance on full manual annotation and greatly improved model training efficiency and final segmentation accuracy.

[0019] 2. High-quality supervision signals are generated using a mature CT segmentation model, and key errors are corrected through a closed-loop correction process at the doctor's end, effectively solving the problems of blurred and easily broken MRI bone boundaries, resulting in high segmentation accuracy of the final model.

[0020] 3. Although CT data and models were used during the training phase, the final deployment consisted of only a standalone MRI bone segmentation model (3DV-Net). The inference chain was simple and easy to integrate and apply in a clinical environment.

[0021] 4. The final deployed model only requires input from a single-modal MRI to output high-precision bone segmentation results, providing reliable technical support for precise clinical surgical planning, and has significant clinical application value and industrialization prospects.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0023] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0026] Figure 1 This is a schematic diagram of the MRI bone segmentation method based on CT domain pseudo-label active learning provided in an embodiment of the present invention.

[0027] Figure 2This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0029] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0030] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0031] The following is combined with Figure 1 As shown, the specific embodiments and working principles of the present invention will be described in detail below: I. Inventive Concept: See Figure 1 As shown, this invention provides an MRI bone segmentation method based on active learning of CT domain pseudo-labels. The core idea of ​​this technology is as follows: First, an MRI-to-CT image translation network is trained using paired CT-MRI; then, in the pseudo-label generation stage, for MRI cases with paired CT, the bone labels of the paired CT or the CT bone segmentation results are directly used to obtain MRI co-spatial pseudo-labels; for MRI cases without paired CT, a virtual CT is first generated through the trained image translation network, and then the CT bone segmentation model is called to obtain MRI co-spatial pseudo-labels; finally, a 3DV-Net is trained together with the pseudo-label pool and the labeled pool confirmed / refined by the doctor, and high-uncertainty predictions are continuously selected by the doctor for discarding or refining through active learning.

[0032] II. Inputs, Outputs, and Data Sources: Input Data 1: Paired CT-MR data used to train the MRI-to-CT image translation network. Paired data comes from CT and MRI scans of the same patient and the same anatomical region. Completely synchronized scans are not required. Training pairs can be formed through rigid registration, deformable registration if necessary, resampling, and anatomical region cropping.

[0033] Input data 2: MRI data for which pseudo-labels are to be generated, including MRI cases with paired CT scans and MRI cases without paired CT scans. For the former, MRI cospatial pseudo-labels can be generated directly using paired CT labels. For the latter, pseudo-labels are generated after creating a virtual CT scan through an MRI-to-CT image translation network. MRI sequences can include T1, T2, VIBE, Dixon, etc.

[0034] Input data 3: Doctor-annotated or corrected data, used to correct pseudo-labeling errors and participate in model training as a highly reliable supervisory signal. Doctors can annotate the entire case or only correct the local 3D ROIs selected by active learning.

[0035] Intermediate outputs: trained MRI to CT image translation network, virtual CT, bone pseudo-labels output by the CT bone segmentation model, pseudo-label quality score, pseudo-label pool, labeled pool after doctor confirmation / refinement, and active learning candidate sample list.

[0036] Final output: MRI bone segmentation model. During deployment, a real MRI scan is input, and the model directly outputs labeled images of bony structures, including ribs, costal cartilage, sternum, thoracic vertebrae, lumbar vertebrae, sacrum, pelvis, hip bones, etc.

[0037] III. Overall Technical Approach: The technical route is divided into three steps, corresponding to Figure 1 Steps 1, 2, and 3 are described below. Step 1 is the image translation network training phase, where paired CT-MRI images are input, and the trained MRI images are output to the CT image translation network. Step 2 is the pseudo-label generation and data pool construction phase. For cases with paired CT images, MRI cospatial pseudo-labels are directly generated from paired CT labels or CT bone segmentation results. For cases without paired CT images, a virtual CT image is first generated via the MRI-to-CT image translation network, and then MRI cospatial pseudo-labels are generated via the CT bone segmentation model. All results enter the pseudo-label pool (pseudo-annotation pool), from which doctors confirm / refine high-uncertainty or high-information-content samples to form the labeled pool. Step 3 is the 3D V-Net active learning training phase. The 3D V-Net is first trained using the pseudo-label pool and the labeled pool, and then predicts bone labels from MRI images. After uncertainty assessment, high-uncertainty samples are discarded or refined by the doctor and returned to the labeled pool.

[0038] Step 1: Training the MRI to CT image translation network: An MRI-to-CT image translation network is employed. The input is an MRI image, and the output is a synthesized CT image with CT grayscale and texture features, i.e., a virtual CT. During training, paired CT-MR data is used, enabling the network to learn the mapping relationship between MRI intensity distribution and CT intensity distribution in the same anatomical region.

[0039] The image translation network employs 3D ResUNet-L. This network consists of a 3D convolutional encoder, a bottleneck layer, a 3D upsampling decoder, residual convolutional blocks, and skip connections. The encoder extracts multi-scale anatomical features from the MRI volume data level by level, the bottleneck layer learns a high-level mapping from MRI intensity to CT grayscale distribution, and the decoder recovers the spatial resolution level by level. Skip connections transmit superficial bone boundary and spatial location information to the decoder, ensuring that the generated virtual CT has CT-like grayscale while maintaining the same voxel coordinate system as the input MRI. It is important to emphasize that the goal of image translation is not simply to generate aesthetically pleasing CT scans, but to generate virtual CT scans that are suitable for CT bone segmentation model processing and maintain the same spatial co-location as the input MRI.

[0040] The structural parameters of 3D ResUNet-L are as follows: the input is a single-channel MRI 3D patch, with a typical size of 128×128×64 or adjustable according to the video memory; the encoder includes 4 levels of downsampling, with the number of channels set to 32, 64, 128, and 256, and the number of channels in the bottleneck layer is set to 512; each level contains two 3D convolutions, a normalization layer, and LeakyReLU activation, and is stabilized through residual connections; the decoder uses 3D transposed convolutions or upsampling convolutions to restore the resolution step by step, and finally outputs a single-channel virtual CT through a 1×1×1 convolution.

[0041] The discriminator uses a 3D PatchGAN structure: the input is a real CT volume C or a virtual CT volume C. It outputs local patch-level true / false probability maps instead of a single global true / false scalar. The discriminator consists of multiple 3D convolutional layers, normalization layers, and LeakyReLU, progressively expanding the receptive field to constrain the virtual CT to have local textures close to those of real CT in the bone cortex boundary, medullary cavity, and surrounding soft tissue transition areas.

[0042] The adversarial loss is implemented using a least-squares GAN.

[0043] The discriminator loss is:

[0044] The generator adversarial loss is:

[0045] in, Represents the mathematical expectation. This represents the discriminator, and C represents the paired real CT scan. This represents a virtual CT scan generated by an image translation network.

[0046] During training, the generator-side loss includes image similarity loss, generator adversarial loss, and geometric consistency loss; the discriminator loss is used separately to update the discriminator parameters.

[0047] MRI to CT image translation training objective: Virtual CT is Generation, the core constraint is Instead of registering to MRI via post-processing, the input M is kept in the same voxel coordinate system.

[0048]

[0049] Where: M represents the input MRI volume data. This represents an MRI-to-CT translation network with parameter θ.

[0050] Weighted image similarity loss is used:

[0051]

[0052] in, Ω represents the weighted image similarity loss; v represents the voxel position, and Ω represents the set of voxels involved in the training. Voxel representation The weights; Indicates the weighting coefficient for the bone region; For indicator functions, when If the bone region is located, take 1; otherwise, take 0.

[0053] The structure preservation term is written as:

[0054] in: This represents the geometric consistency loss; Indicates the bone region extraction operator; Used to measure the overlap between virtual CT bone regions and real CT bone regions; Represents the image gradient; Indicates the gradient constraint weights at the bone edge; This represents the norm calculated in the bone margin region.

[0055] Bone region extraction operator Obtain the bone mask by calling an existing CT bone segmentation model. After outputting the binary bone region mask, Simultaneously constraining the overlap of bone regions in Dice and the gradient difference of bone boundaries; thus, even if there is a slight deviation in the overall grayscale of the virtual CT, the bone boundaries and topology most sensitive to the subsequent CT bone segmentation model can be prioritized.

[0056] The overall translation loss is as follows:

[0057] in: The total loss on the image translation generator side; This represents the multi-scale structural similarity loss. This indicates that the generator is adversarial against loss; The weights for each loss term are used to balance grayscale similarity, structural similarity, adversarial realism, and bone structure consistency.

[0058] The discriminator loss is not incorporated into the total loss on the generator side, but is optimized alternately with the generator during GAN training: when the generator is fixed, the discriminator is updated to distinguish between real CT and virtual CT, and when the discriminator is fixed, the generator is updated to improve the CT domain realism of virtual CT.

[0059] Step 2: Generate the pseudo-label pool and the labeled pool: When generating MRI cospatial bone pseudo-labels based on a trained image translation network and paired CT scans, two input paths are explicitly defined. Path A is for MRI cases with paired CT scans: first, the paired CT scans are coarsely aligned using rigid registration; then, deformable registration or local non-rigid correction is used to map them to the MRI space as needed. Using artificial bone labels from the paired CT scans or CT bone segmentation results, after resampling, registration quality control, and label quality control, MRI cospatial pseudo-labels are obtained. Path B is for MRI cases without paired CT scans: first, the MRI scans are input into the CT image translation network trained in step 1 to generate virtual CT scans; then, the virtual CT scans are input into the CT bone segmentation model to obtain bone labels, and the original MRI space is inherited as the MRI cospatial pseudo-label. Labels generated by both paths are used as pseudo-labels before physician confirmation and are uniformly entered into the pseudo-label pool.

[0060] After pseudo-labels are generated, quality control is required, including small connected component removal, hole filling, left-right consistency checking, rib continuity checking, vertebral sequence continuity checking, spatial relationship between the sternum and the anterior ribs, adjacency relationship between the sacrum and pelvis, false positive fragment removal, confidence scoring, and uncertainty assessment. MC dropout is used to estimate uncertainty; pseudo-labels with low uncertainty and high quality scores can be retained in the pseudo-label pool as weakly supervised samples; cases / local ROIs with high uncertainty, high error risk, or high information content are submitted for physician confirmation or refinement, and physician-confirmed labels are added to the labeled pool.

[0061] The pseudo-label pool stores weakly supervised labels that have not been individually verified by doctors but have passed quality control; the labeled pool stores highly reliable labels that have been verified by doctors or refined. Both types of data pools serve as input for training the 3D V-Net in step 3.

[0062] Pseudo-tag generation, quality control, and pooling conditions: CT bone segmentation model denoted as The spatially aligned images of paired CT scans are denoted as MRI is recorded as The MRI-to-CT image translation network is denoted as Then pseudo-tags can be generated from two paths:

[0063]

[0064] in: This indicates a spatial bone pseudo-label in MRI. This represents the quality control operators, including connected component filtering, hole filling, left-right symmetry checking, and anatomical topology checking. The first path is used for cases with paired CT scans, and the second path is used for cases without paired CT scans; both outputs are MRI cospatial pseudo-labels, which enter the same pseudo-label pool.

[0065] TotalSegmentator-CT is used in this scheme as an existing CT domain bone segmentation tool during the training phase. Its input can be real paired CT or virtual CT generated by MRI, and its output is multi-class bone structure labels.

[0066] The TotalSegmentator-CT output requires post-processing: small connected components with volumes smaller than a threshold are deleted, internal bone cavities are filled, volume and positional consistency of paired structures are checked, and topological checks are performed on rib continuity, vertebral sequence continuity, and sacral-pelvic adjacency relationships. Finally, MRI cospatial pseudo-labels are obtained. The connected component volume threshold is set according to the low quantile of the volume distribution of the corresponding bone category in the training set, for example, removing isolated fragments smaller than 1% of the median volume of that category; left-right symmetry can be judged comprehensively based on the left-right volume ratio, centroid distance, and principal axis direction difference; topological checks are used to detect breaks, abnormal adjacencies, or obviously missed segmentation areas.

[0067] Pseudo-label quality score: When the CT bone segmentation model repeats forward inference T times on the same virtual CT, the random prediction probability of the t-th time is obtained. :

[0068] right The average of the random predictions is used to obtain the average class probability of MC dropout. :

[0069] Prediction entropy is calculated based on average probability. And calculate the expected entropy based on each random prediction. :

[0070]

[0071] The difference between the two can be used as an approximation of cognitive uncertainty or information gain in the model:

[0072] Voxel-level mutual information is summarized within the candidate bone structure regions, and combined with the proportion of low-confidence voxels to obtain the pseudo-label uncertainty. :

[0073] in: Represents a set of bone categories. Represents a candidate region for bone structure, taking the set of voxels whose maximum bone class probability exceeds a threshold; The weighting for low confidence levels is preferably between 0.5 and 2.0. It is an indicator function. This indicates a low confidence threshold, preferably between 0.5 and 0.7; The number of repeated inferences is preferably 5 to 20 times, with 10 times being more ideal. This uncertainty is used as a penalty in the pseudo-label quality score; that is, the higher the uncertainty, the lower the pseudo-label quality score.

[0074] The information content score in step 2 is only used to determine the priority of pseudo-label review and does not reference the current prediction differences of 3D V-Net, because the final MRI segmentation model has not yet completed active learning training at this point. Three types of information are used: pseudo-label uncertainty, low-quality risk, and sample diversity.

[0075] in: The higher the value, the more worthy a doctor's review is; This indicates the inherent uncertainty of pseudo-labels; This indicates that the false label carries a low credibility risk; This indicates the difference between the sample and the labeled pool distribution; to The corresponding weights.

[0076] The pseudo-label quality score is used to determine whether the pseudo-labels generated in step 2 can be directly used as weakly supervised training samples. This score is calculated separately from the review priority score in step 2. A higher value indicates a more credible pseudo-tag. The higher the value, the more worthy the doctor is of reviewing the sample.

[0077]

[0078] in: Measuring the reliability of image or label sources: For path A, it represents the quality of paired CT images, CT-MRI registration accuracy, and reliability of CT label sources; for path B, it represents the quality of virtual CT, MRI-to-CT translation stability, and the segmentability of virtual CT bone structures. Measuring bone structure connectivity and the presence of anomalous small connected regions; Measuring the consistency of paired bone structures on both sides; Assess the anatomical topological rationality between the ribs, sternum, vertebrae, sacrum, and pelvis; The summation of the aforementioned prediction entropy, mutual information, low-confidence voxel ratio, or test-time consistency enhancement is used as a penalty for the quality score. All scores for both paths are normalized to 0 to 1 before being incorporated into the same quality score formula, thus ensuring comparability.

[0079] Weight settings: All terms are assigned a weight of 1.0, and then adjusted based on the acceptability of pseudo-labels, the missing segmentation rate, and the false positive rate on the validation set. If a certain type of error has a greater impact on subsequent 3D V-Net training, the weight of the corresponding term can be increased.

[0080] →Enter the pseudo-tag pool; or Higher → Doctor confirmation / refinement queue; Judgment rules: High-quality pseudo-labels with low uncertainty enter the pseudo-label pool and are used as weakly supervised samples weighted by quality score in step 3; low-quality, high-uncertainty samples or samples with high review priority in step 2 enter the doctor confirmation / refinement queue; samples in the middle range can be temporarily deferred or have their training weight reduced.

[0081] Parameter description: The threshold for inclusion in the high-confidence pool is... For low-confidence verification thresholds, both can be set according to the candidate pool or validation set quantiles, for example... Take the top 85 percentile of the mass fraction. Take the bottom 15% quantile. (Step 2) The difference between the 3D V-Net prediction and the pseudo-label in step 3 is used separately for active learning sampling, and the two are not used together.

[0082] Step 3: Actively learn and train the 3D V-Net MRI bone segmentation model: An active learning paradigm was used to train a 3D V-Net MRI bone segmentation model. During initial training, 3D V-Net simultaneously used the labeled pooling and pseudo-labeling pooling formed in step 2: physician-confirmed / refined labels served as strong supervision signals, while pseudo-labels served as weak supervision signals weighted by quality scores. After training, the model predicted unlabeled or candidate MRI images to obtain MRI bone structure prediction labels, and the uncertainty and error risk of these prediction labels were assessed.

[0083] Uncertainty and error risk calculation method: Calculate the prediction entropy for each voxel category probability output by the MRI bone segmentation model, and statistically analyze the proportion of voxels with high entropy or low maximum probability. Uncertainty in bone segmentation models. Active learning selects samples based on the predicted labels from MRI output using 3D V-Net. Prediction uncertainty, discrepancies between predictions and pseudo-labels / annotated labels, low reliability risk of pseudo-labels, and sample diversity are combined into a unified comprehensive score. For local 3D ROIs, the 95th percentile within the ROI can be used. The higher the value, the more unstable the current prediction of 3D V-Net is, the more likely it is to conflict with existing supervision signals, or the more it can supplement the training set distribution, and the higher the priority it should be submitted to the doctor for processing.

[0084] In the active learning cold start phase, to avoid sampling bias caused by the initial 3D V-Net being trained solely on pseudo-labels, a small number of physician-confirmed samples are used as the seed set. The seed set comes from the samples confirmed / refined by physicians in step 2, or from a small number of independently labeled samples; preferably, it is randomly drawn from different MRI sequences, different body types, and different anatomical regions. Subsequently, in each round, high-value cases or ROIs are selected based on the comprehensive active learning score.

[0085] Active learning stop condition In addition, the maximum number of rounds can be set. Conditions include the upper limit of the number of samples reviewed by doctors in a single round, the upper limit of the overall annotation budget, and the lack of significant improvement in the Dice or HD95 validation sets for two consecutive rounds. The validation set Dice can be increased by 0.001, or the process can be stopped when the performance difference relative to the fully supervised test is lower than a preset threshold.

[0086] Active learning is represented by feedback arrows in the diagram: After the 3D V-Net outputs MRI bone prediction labels, the uncertainty and error risk of these labels are calculated. High-uncertainty predictions are processed by doctors, who can either discard the data or refine the prediction labels. The refined labels are then added to the labeled pool. The labeled pool and the pseudo-label pool are used to update the 3D V-Net. This feedback loop is a key part that distinguishes this technique from one-time pseudo-label training.

[0087] The active learning interaction module consists of a pseudo-label pool, an labeled pool, a candidate sample pool, a 3D V-Net inference scorer, a doctor review / labeling end, a corrected label entry module, and an iterative training scheduler. Its working principle is as follows: 3D V-Net outputs predicted labels to MRI. The scorer ranks cases or ROIs based on prediction uncertainty, differences from pseudo-labels / labeled labels, pseudo-label quality risk, and sample diversity. Doctors discard, confirm, or refine high-scoring samples. Confirmed / refined labels enter the labeled pool with high weight for the next round of training. Discarded samples are removed from the current training queue, and the reasons are recorded. If necessary, they are returned to the candidate pool for subsequent quality control. After the model is updated, the candidate pool is scored again, forming an active learning closed loop.

[0088] In the final deployment, only the MRI bone segmentation model is retained. For clinical use, inputting the MRI image will output bony structure labels; there is no need to run an image translation network, a CT bone segmentation model, or a real CT scan in real time.

[0089] IV. System Architecture: 1. MRI to CT image translation module: This module is used to convert MRI into virtual CT. The core requirement is to maintain the spatial position of bone structures so that the labels output by the subsequent CT bone segmentation model can be directly used as pseudo-labels in the same spatial area as MRI. A 3D ResUNet-L image translation network structure is adopted, focusing on paired CT-MR training, bone region weighting, and structure preservation constraints.

[0090] 2. CT bone segmentation pseudo-label module: This module is used to generate MRI isospace bone pseudolabels for step 2. For cases with paired CT scans, the CT bone segmentation model is run on the paired CT or existing CT bone labels are used; for cases without paired CT scans, the CT bone segmentation model is first run on a virtual CT, and then the MRI isospace bone pseudolabels are obtained. The CT bone segmentation model uses TotalSegmentator-CT; this paper focuses on its use for pseudolabel generation and data pool construction, rather than as a final clinical deployment model.

[0091] 3. Counterfeit Label Quality Control Module: This module evaluates the reliability of pseudo-labels. Evaluation metrics include virtual CT quality, bone label connectivity, left-right consistency, rib / vertebral continuity, and pelvic-sacral adjacency. Virtual CT quality is calculated using MAE, PSNR, MS-SSIM, and bone boundary HD95; connectivity is calculated using the number of connected domains in bone instances, the percentage of the largest connected domain, and the break length; left-right consistency is calculated using the volume difference and spatial symmetry difference between corresponding bone categories on the left and right sides; and topological consistency is determined by rules checking rib-sternal anterior end, vertebral sequence, and sacral-pelvic adjacency relationships.

[0092] 4. MRI bone segmentation model (MRI bone segmentation module): This module represents the final deployment model. The model takes real MRI scans as input and outputs multi-class bone structure labels. The current implementation uses a V-Net segmentation network, consisting of a 3D encoder, a bottleneck layer, and a decoder. The encoder extracts bone / soft tissue features at different scales through downsampling, the decoder progressively restores spatial resolution, skip connections preserve details such as rib boundaries, vertebral appendages, and pelvic edges, and finally, multi-class bone structure probability maps are output through 1x1x1 convolutions and Softmax. Training losses include Dice loss, geometric consistency loss, and cross-entropy loss.

[0093] V-Net employs the original V-Net 3D encoder-decoder structure. In the preferred implementation, the input is an MRI 3D patch, with a size of 128×128×64, and the number of output channels equals the number of target bone categories plus the number of background categories. The encoder downsamples 4 times, and the decoder recovers fine-grained boundaries through skip connections. During inference, a sliding window approach is used to process the entire MRI sample, with a window overlap rate set to 50%. Overlapping regions are fused using probability averaging or Gaussian weighted fusion to reduce block boundary artifacts.

[0094] The final hybrid supervised training loss for the MRI bone segmentation model: The final deployment model is denoted as The predicted probability is defined as follows. A strongly supervised loss is applied to the doctor-labeled samples. Weakly supervised loss weighted by confidence level is used for pseudo-labeled samples. .

[0095]

[0096]

[0097] in Suppressing the negative impact of low-reliability pseudo-labels on training:

[0098] Parameter description: The parameter is MRI bone segmentation model; This represents the multi-class probability map output by the model; This indicates a highly reliable label confirmed by a doctor; Indicates a pseudo-tag; , and These represent Dice loss, cross-entropy loss, and anatomical geometry consistency loss, respectively. As weight.

[0099] The loss of anatomical geometric consistency can be defined as:

[0100] in: Punishment for bone structure fractures and abnormally small connected regions. Constraints should ensure consistency between the gradient or distance transformation of the predicted bone boundary and the labeled bone boundary. Violations of anatomical topological rules, such as rib-sternal adjacency, vertebral body sequence continuity, and sacral-pelvic adjacency, are penalized. This loss is used to reduce fragmentation and undersegmentation caused by low signal intensity in bone on MRI.

[0101] Overall training objectives: A consistency regularity term can be defined as:

[0102] in ` represents data augmentation methods such as random rotation, flipping, intensity perturbation, and cropping. This requirement mandates that the predictions of the same MRI remain consistent before and after enhancement, and can also be extended to a prediction consistency constraint between adjacent active learning iterative models.

[0103] The optimal values ​​for the segmentation loss weights can be given: Take 1.0, Take 1.0, Take 0.3, Take 0.5. The doctor confirms the label corresponds to... It is consistently regarded as a highly reliable and strongly supervised item, with a higher weight than the pseudo-label, weakly supervised item.

[0104] The final total loss during segmentation training is:

[0105] in: This represents the total training loss for the final segmentation. These represent the number of doctor-labeled samples and the number of pseudo-labeled samples, respectively. This indicates weakly supervised weights for pseudo-labels; Indicates the weight of the consistency regularization term; Used to constrain the consistency of predictions for the same case across augmented, different views, or adjacent iterations.

[0106] 5. Active learning interaction module: This module focuses physician annotations on the most valuable samples. Steps 2 and 3 share the same labeled pool, but the objects processed by the physicians differ: in Step 2, physicians deal with pseudo-labels generated from paired CT paths or virtual CT paths, primarily performing confirmation or refinement to convert credible pseudo-labels into highly credible labels; in Step 3, physicians deal with the predicted labels from the MRI output by 3D V-Net, primarily performing discarding, confirmation, or refinement to correct high uncertainty predictions in the current model. Physicians can annotate entire cases or correct only local ROIs, such as segmenting fractured rib areas, missing segmentation of vertebral appendages, or pelvic boundary errors.

[0107] V. Active Learning Sample Selection: Active learning occurs in step 3, and its evaluation object must be explicitly defined as the predicted labels of the MRI output by 3D V-Net, not the pseudo-labels already generated in step 2. Specifically, after the current round of training, 3D V-Net infers on candidate MRIs to obtain MRI bone segmentation prediction labels; the system then calculates the uncertainty, error risk, and difference from pseudo-labels / labeled labels for these prediction labels, and selects cases or local ROIs for physician processing accordingly. The model calculates a unified comprehensive score for the predicted MRI bone segmentation mask. :

[0108] in: Uncertainty in MRI bone segmentation models can be obtained by predictive entropy, low maximum probability voxel ratio, or MCdropout / test-time enhanced consistency.

[0109] The difference between the 3D V-Net predicted labels and the pseudo-labels or confirmed labels from step 2.

[0110] False tags pose a low credibility risk; Sample diversity; , , , As weight.

[0111] Executable definitions for each item in the active learning scoring: Segmentation uncertainty:

[0112] The prediction-pseudo-label discrepancy is used to identify cases or local areas of interest (ROIs) where the current 3D V-Net prediction is inconsistent with the pseudo-labels or confirmed labels generated in step 2 on the MRI co-space. Significant discrepancies indicate a high risk of error on at least one side of the model prediction or the existing pseudo-label, requiring physician confirmation.

[0113]

[0114] Sample diversity is used to avoid selecting only similar difficult cases in each round.

[0115] The feature extraction function can take the global average pooling features from the bottleneck layer of the encoder in the MRI bone segmentation model, or the vector obtained by pooling the features from the penultimate layer of decoding. The feature dimension is 256. During calculation, the nearest neighbor distance between the candidate case or ROI features and the existing doctor-annotated sample set is used as the diversity score. The larger the distance, the more it can supplement the existing training set distribution.

[0116] Active learning scoring weights Desirable This prioritizes uncertainty and the difference between prediction and pseudo-labels, with low credibility risk and diversity of pseudo-labels serving as auxiliary constraints; the Dice gain can be adjusted in each round based on the validation set.

[0117]

[0118] The composite score is used after normalization. In each round, the case or local ROI with the highest score is selected to enter the doctor's treatment queue. Parameter description: or This indicates a case to be evaluated or a local three-dimensional region of interest (ROI). The scoring weights are divided into four categories; This is a 0-1 normalization based on the candidate pool statistics; This indicates an existing set of doctor-labeled samples; Represents image features or intermediate layer features of a model; This represents a set of bone categories.

[0119] During the active learning training phase, 3D V-Net prioritizes high-scoring samples for doctors to process. Doctors can discard data unsuitable for training or confirm or refine predicted labels. Confirmed / refined labels are added to the labeled pool and given higher weights in subsequent training. Discarded data is excluded from the current training batch and the weakly supervised pool and recorded as samples with abnormal quality or unsuitable applications. In the inference and deployment phase, the MRI bone labels output by the trained 3D V-Net are used directly.

[0120] VI. Main technical means adopted in this invention: By training an MRI-to-CT image translation network using paired CT-MR, unlabeled MRI images can generate virtual CT images that are in the same spatial location as the original images.

[0121] In step 2, MRI cospatial pseudo-labels are generated through two paths: for cases with paired CT, paired CT labels or CT bone segmentation results are used directly; for cases without paired CT, a virtual CT is generated first and then the CT bone segmentation model is called.

[0122] The pseudo-labels undergo bone structure quality control, probability confidence assessment, and uncertainty assessment. When the model supports the algorithm, MC dropout or model ensemble is used; otherwise, softmax entropy, low-confidence voxel ratio, and enhanced consistency during testing are employed. High-quality pseudo-labels are added to the pseudo-label pool, while samples with high uncertainty or high information content are added to the labeled pool after physician confirmation / refinement.

[0123] In step 3, 3D V-Net is trained using a pseudo-label pool and an labeled pool. Through active learning, cases / ROIs with high uncertainty, high error risk, or high clinical value predicted by 3D V-Net are selected for doctors to discard, confirm, or refine. The current validation goal is to achieve near-fully supervised training model performance with approximately 10% of the sample size reviewed by doctors. The specific savings need to be quantified by combining experiments with different labeling ratios and statistics on doctor labeling time.

[0124] The final trained model is an independent 3D V-Net MRI bone segmentation model. When deployed, it only requires MRI input and does not require real-time running of an MRI-to-CT image translation network or a CT bone segmentation model.

[0125] VII. Technical differences from existing methods: The technical differences between this invention and existing methods are shown in Table 1.

[0126] Table 1. Differences from existing methods

[0127] As shown in Table 1 above, compared with existing methods, the present invention has significant technical differences and advantages.

[0128] VIII. Verification Results Completed: This section summarizes the validation results that have been completed so far. The validation included virtual CT quality, pseudo-label quality, active learning benefits, the performance of the final MRI bone segmentation model, and the deployment pipeline. The validation results are shown in Table 2.

[0129] Validation metrics and closed-loop return quantification: Dice is used to evaluate the overlap of voxels in various bone structures.

[0130]

[0131] HD95 is used to measure boundary errors and reduce the impact of extreme outliers.

[0132]

[0133] The annotation saving rate can be defined as follows, which is used to support the technical effect of "achieving near-full supervision effect with low annotation cost".

[0134]

[0135] Benefits of active learning It can be written as follows, when Less than the threshold Stop adding new rounds of doctor review; The Dice value for the validation set can be preferably increased by 0.001 to 0.005, or set according to the physician's annotation budget.

[0136]

[0137] Parameter description: P represents the model's predicted label, and Y represents the reference label. P and Y represents the predicted / reference label boundary, and d(a,b) represents the distance between boundary points; This represents the minimum reward threshold at which active learning iterations should cease.

[0138] Table 2 Validation Results

[0139] Dice, HD95, bone structural continuity, consistency, and puncture path-related indices were statistically analyzed on an independent MRI test set. Based on the aggregated test results, the median Dice for each bone-related category, after averaging, was approximately 0.84, and the overall median Dice for all target structures was approximately 0.86.

[0140] IX. Closed-loop logic: Overall conclusion: The methodological theoretical chain is basically closed-loop. Its closed-loop logic is as follows: "Training the MRI-to-CT image translation network using paired CT-MRI -> For cases with paired CT, generating MRI cospatial pseudo-labels directly from paired CT labels / CT bone segmentation results; for cases without paired CT, generating MRI cospatial pseudo-labels by calling the CT bone segmentation model after generating a virtual CT from MRI -> Quality control, confidence level, and MC dropout information content evaluation -> Constructing a pseudo-label pool and a labeled pool after physician confirmation / refinement -> Training the 3D V-Net MRI bone segmentation model using the two data pools -> Assessing uncertainty and error risk for 3D V-Net predicted labels -> Physicians discard or refine high-uncertainty predictions -> Refined labels are returned to the labeled pool -> Iteratively updating 3D V-Net -> Only the 3D V-Net MRI bone segmentation model is retained during the deployment phase."

[0141] This approach explains why we don't directly perform simple CT-MRI label registration and transfer, and why deployment doesn't require real CT, real-time virtual CT, or CT bone segmentation models. The core theory is: during the training phase, the CT domain bone label capability is converted into MRI co-spatial pseudo-labels, and then through physician confirmation / refinement and active learning, the pseudo-label noise is gradually converted into high-confidence MRI supervision signals, ultimately resulting in an independent 3D V-Net MRI bone segmentation model.

[0142] Key links that have already been closed in the loop: Data loop: Paired CT-MRI is used to learn image translation; MRI with and without paired CT are used together to expand the pseudo-label pool; labels confirmed or refined by doctors are returned to the labeled pool; the pseudo-label pool and the labeled pool are used together to train the 3D V-Net.

[0143] Feasibility of acquiring paired CT-MRI data: Paired data can come from CT and MRI scans acquired by the same patient within the same treatment cycle for diagnosis, preoperative evaluation, or interventional planning, without requiring completely synchronous scanning. Training pairs are formed through rigid registration, deformable registration when necessary, resampling, and anatomical region trimming. For MRI cases without paired CTs, they can be used as unlabeled pools in the virtual CT generation, pseudo-label screening, and active learning processes.

[0144] Domain generalization processing explanation: When clinical MRI sequences differ from training sequences, the quality of virtual CT can be improved through multi-sequence training, intensity standardization, histogram matching, sequence type conditional encoding, or fine-tuning the translation network with a small amount of target sequence paired data; simultaneously utilizing... and Identify samples from outside the region and prioritize their review by doctors.

[0145] Spatial closed loop: Virtual CT is directly generated from MRI, so the CT bone segmentation output inherits the MRI coordinate system, avoiding label offset caused by insufficient precise registration of CT-MRI for each case.

[0146] Training loop: Step 2 generates a pseudo-label pool and a labeled pool, not the final model; the final 3D V-Net is trained in step 3. Doctor confirmation labels serve as strong supervision signals, while pseudo-labels serve as weak supervision signals weighted by quality scores, jointly driving the iterative updates of the 3D V-Net.

[0147] Active learning closed loop: The higher the uncertainty of the 3D V-Net predicted label, the greater the difference from the pseudo label, and the more representative the sample, the higher the priority for the doctor to process; the doctor can discard the data or refine the predicted label, and the refined result enters the labeled pool and retrains 3D V-Net.

[0148] Deployment closed loop: During the training phase, image translation networks, paired CT, CT bone segmentation tools, pseudo-label pools, and physician closed loops can be used. During the deployment phase, only the trained 3D V-Net is retained, MRI is input, and bone labels are output, resulting in a clear clinical link.

[0149] In summary, this invention provides an MRI bone segmentation method and system based on CT domain pseudo-label active learning. By using cross-modal pseudo-label generation technology, it effectively solves the problems of weak bone structure signals and high annotation costs in MRI images. Furthermore, it significantly improves model training efficiency and segmentation accuracy by utilizing closed-loop active learning. The final deployed model only needs to input MRI to output accurate bone segmentation results.

[0150] Additionally, refer to Figure 2 As shown, embodiments of the present invention also provide an electronic device that can execute the above-described method or run the above-described system. The electronic device may include a processor, a memory, a communication bus, and a communication interface, and may also include a computer program stored in the memory and executable on the processor.

[0151] In some embodiments, the processor may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units, microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control core of the electronic device, connecting various components of the device through various interfaces and lines. It executes programs or modules stored in memory and calls data stored in the memory to perform various functions and process data within the electronic device.

[0152] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0153] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0154] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0155] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An MRI bone segmentation method based on CT domain pseudo-label active learning, characterized in that, The method includes the following steps: S1. Obtain paired CT-MRI datasets and train an MRI-to-CT image translation network based on these datasets; S2. A dual-path bone pseudo-label generation mechanism is used to generate bone pseudo-labels in the same MRI space, wherein: For MRI samples with paired CT scans, MRI spatial bone pseudo-labels are generated directly using paired CT labels or CT bone segmentation results, after spatial registration, resampling, and quality control. For MRI samples without paired CT images, a virtual CT image in the same space is generated using the MRI-to-CT image translation network; and a pre-trained CT bone segmentation model is used to segment the virtual CT image to generate bone pseudo-labels in the same space as the MRI. A pseudo-label pool is constructed based on bone pseudo-labels generated by dual paths, and quality control and confidence assessment are performed to screen out high-confidence pseudo-labels. These high-confidence pseudo-labels are used as weakly supervised samples in the initial training and subsequent iterative training of the MRI bone segmentation model. Labels confirmed or corrected by doctors are added to the labeled pool and used as strongly supervised samples in training. S3. Based on the pseudo-label pool and the labeled pool, a closed-loop active learning strategy is used to iteratively train the MRI bone segmentation model. During training, the MRI bone segmentation model outputs predicted labels for candidate MRIs, calculates the inconsistency between the predicted labels and pseudo-labels / labeled labels, filters difficult samples and sends them to the doctor for discarding, confirmation or correction, and adds the confirmed or corrected labels to the labeled pool; iterative training continues until the preset convergence condition is met, and the final MRI bone segmentation model is obtained. S4. Use the final MRI bone segmentation model to segment the MRI image to be segmented into bone structures and output the segmentation results.

2. The method according to claim 1, characterized in that, In step S2, the quality control specifically includes at least one of the following: small connected domain removal, hole filling, left-right symmetry inspection, rib continuity inspection, vertebral body sequence continuity inspection, and anatomical adjacency relationship inspection.

3. The method according to claim 1, characterized in that, In step S2, the confidence assessment specifically includes: Extract the intermediate layer feature map or output probability map of the pre-trained CT bone segmentation model when generating the bone pseudo-label; Based on the intermediate layer feature map or output probability map, calculate the segmentation uncertainty measure at the voxel level; A confidence score is generated based on the uncertainty measure, and a confidence threshold is set. Pseudo-labels that are higher than the threshold are marked as high-confidence pseudo-labels.

4. The method according to claim 1, characterized in that, In step S3, the predicted labels are sorted according to the prediction uncertainty, the difference with pseudo-labels / labeled labels, the quality risk of pseudo-labels and sample diversity. Difficult samples are selected and pushed to the doctor for disposal, confirmation or correction. Confirmed / corrected labels are put into the labeled pool and participate in the next round of training with preset weights. Discarded samples are excluded from the current training queue and the reason is recorded.

5. The method according to claim 1, characterized in that, In step S1, the MRI-to-CT image translation network uses a 3D ResUNet-L network.

6. The method according to claim 1, characterized in that, In step S2, before generating a virtual CT image in the same space using the MRI-to-CT image translation network, the following steps are also included: spatial standardization and grayscale normalization preprocessing of the MRI samples to eliminate data distribution differences caused by acquisition parameters of different MRI devices.

7. An MRI bone segmentation system based on CT domain pseudo-label active learning, characterized in that, The system includes: an MRI-to-CT image translation module, a CT bone segmentation pseudo-label module, a pseudo-label quality control module, an MRI bone segmentation module, and an active learning interaction module; the system uses the method described in any one of claims 1 to 6 to obtain the final MRI bone segmentation model, performs bone structure segmentation on the MRI image to be segmented, and outputs the segmentation results.

8. An electronic device, characterized in that, The device includes a processor, a memory, and a display device, wherein the memory stores a program that, when executed by the processor, implements the method according to any one of claims 1 to 6.

Images