A smart lesion identification and tracking system for magnetic resonance imaging

Abstract

This invention relates to the field of medical image processing technology and discloses an intelligent lesion identification and tracking system for magnetic resonance imaging. The system includes: constructing a three-dimensional voting matrix to record the cumulative votes of voxels; generating proposals for suspected lesion regions and accumulating votes through a lightweight identification algorithm; performing image sequence registration based on skeletal reference to ensure spatial alignment; and finally determining high-confidence lesions based on a vote threshold. This invention uses a spatiotemporal voting mechanism to enable real lesions to form stable vote peaks over multiple scans, while noise cannot accumulate effectively due to random distribution, thus achieving robust identification under low signal-to-noise ratio conditions. Simultaneously, it combines morphological consistency verification to effectively distinguish structural artifacts, improving the accuracy and reliability of lesion identification.

Description

Technical Field

[0001] This invention relates to an intelligent lesion identification and tracking system for magnetic resonance imaging, belonging to the field of medical image processing technology. Background Technology

[0002] In the field of medical image processing, intelligent lesion recognition and tracking systems for magnetic resonance imaging are a key technology for improving diagnostic efficiency. Current mainstream methods rely on deep learning models to detect lesions in a single high-quality scan image. Such technologies extract complex spatial features through training on massive amounts of data to achieve accurate recognition. However, in clinical practice, scenarios where patients frequently undergo follow-up examinations (such as monitoring the efficacy of radiotherapy) often require the use of rapid scanning protocols to shorten examination time. The resulting images often suffer from significant decreases in signal-to-noise ratio, motion artifact interference, and positioning differences.

[0003] When existing technologies are applied to such low-quality images, the fine texture features that the model relies on are overwhelmed by noise, resulting in a precipitous drop in recognition performance. The core contradiction is that existing systems deeply bind diagnostic reliability to the quality of a single image and cannot extract effective information from multiple low-quality scans.

[0004] Specifically, existing technologies suffer from three fundamental limitations: 1. The design paradigm based on a single perfect acquisition makes it difficult to distinguish between real lesions and structural artifacts under low signal-to-noise ratio conditions; 2. The lack of a cross-time series collaborative analysis mechanism makes it impossible to utilize the spatial constancy of lesions to combat random noise; 3. The systematic impact of physiological state fluctuations on image quality is ignored, and a correlation model between data reliability and physiological stability has not been established. Although some studies have attempted to improve this by increasing network complexity or introducing specific artifact suppression algorithms, these efforts further increase the computational load and fail to solve the problem of multi-source interference coupling. Therefore, how to construct a system that does not rely on the quality of a single image and can achieve robust lesion identification and tracking from multiple low signal-to-noise ratio scans, and effectively distinguish the fundamental difference between the spatial consistency of anatomical structures and the randomness of artifacts, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides an intelligent lesion identification and tracking system for magnetic resonance imaging. Its main purpose is to solve the problems that existing technologies cannot achieve robust lesion identification and tracking in low signal-to-noise ratio multiple scan scenarios, and it is difficult to distinguish between the spatial constancy of real lesions and the randomness of artifacts.

[0006] To achieve the above objectives, the present invention provides an intelligent lesion identification and tracking system for magnetic resonance imaging, comprising:

[0007] A three-dimensional voting matrix is ​​configured to record the cumulative votes for each voxel in a three-dimensional space corresponding to the patient's anatomical structure.

[0008] A suspected target proposal module is configured to: receive the current cumulative vote count of a three-dimensional voting matrix, and run a lightweight recognition algorithm for each MRI image obtained from multiple scans to generate a suspected lesion region proposal; add votes to the corresponding positions in the three-dimensional voting matrix according to the voxel positions covered by the suspected lesion region proposal; the recognition threshold of the lightweight recognition algorithm is set to ensure that the recall rate of the suspected lesion region proposal is not less than 95%.

[0009] An image sequence registration module is configured to receive the currently scanned MRI image and, before the suspected lesion region proposal module generates a proposal, perform rigid body or affine transformation registration with the patient's skeletal structure as a stable reference to spatially align the coordinate system of the current scanned image with that of the first scanned image, thereby ensuring that the votes for all suspected lesion region proposals are accumulated in a unified anatomical coordinate system.

[0010] A physiological stability adaptive module is configured to: synchronously acquire electrical noise signals on the gradient coil power supply bus of the MRI equipment during each scan; determine the patient's physiological stability state corresponding to the scan based on the spectral entropy or energy variance of the electrical noise signals, including comparing the spectral entropy or energy variance with a predetermined physiological stability baseline range; and dynamically adjust the vote weight of the suspected lesion region proposal generated by the scan in the three-dimensional voting matrix according to the determined physiological stability state, with the vote weight set to 20% to 100%.

[0011] A final lesion determination module is configured to: after the votes from multiple scans have been accumulated, receive the final accumulated votes from the three-dimensional voting matrix, and identify connected voxel regions whose accumulated votes are higher than 75% of the total number of votes as high-confidence lesions.

[0012] Preferably, the system further includes a morphological consistency verification module, configured to: before the final lesion determination module determines the high-confidence lesion region as the final lesion, trace and obtain the morphological feature vectors of the suspected proposal image blocks from all previous scans that voted for the high-confidence lesion region; calculate a morphological consistency metric based on the set of morphological feature vectors, the morphological consistency metric being the variance of the Euclidean distance of the set of morphological feature vectors; and the final lesion determination module only determines the high-confidence lesion region as the final lesion when the cumulative number of votes for the high-confidence lesion region is higher than 75% of the total number of votes and the morphological consistency metric is lower than a predetermined morphological consistency threshold.

[0013] Preferably, the system further includes a systematic noise adaptive suppression module, configured to: analyze the spatial distribution of discarded votes with votes below a predetermined low threshold in the three-dimensional voting matrix to identify the existence of a topological pattern of systematic noise, the topological pattern including the periodic frequency peaks presented by the principal axis direction of the point cloud distribution or the projection of the point cloud onto a specific anatomical axis; and, if a topological pattern is identified, in the subsequent suspected target proposal stage of the scanned image, the suspected target proposal module adjusts the proposal logic according to the characteristics of the topological pattern to reduce the proposal weight of regions that conform to the topological pattern by at least 5% and no more than 20%.

[0014] Preferably, the system further includes a kinematic analysis and weighting module, configured to: obtain spatial transformation parameters calculated for aligning each scan image from the image sequence registration module, the spatial transformation parameters including three-dimensional translation parameters and three-dimensional rotation parameters; determine a motion stability index for each scan based on the time series of the spatial transformation parameters, the motion stability index reflecting the amplitude or smoothness of the patient's movement during the current scan; and, when accumulating votes, the suspected target proposal module dynamically weights the votes contributed by the suspected lesion area proposals generated by the current scan according to the motion stability index.

[0015] Preferably, the lightweight identification algorithm used by the suspect target proposal module is an image processing model based on convolutional neural networks or morphological filters. This model is optimized to quickly identify patches that differ from surrounding tissues in low signal-to-noise ratio images, and its computational complexity is limited to a level that can be executed in real time by a low-power central processing unit.

[0016] Preferably, the physiological stability adaptive module determines the patient's physiological stability state, including at least one of cardiac arrhythmia, respiratory rhythm abnormality, or hemodynamic pattern abnormality, and calculates the spectral entropy or energy variance of the electrical noise signal to quantify the degree of disorder or the severity of fluctuation of the signal.

[0017] Preferably, the cumulative vote threshold of 75 percent of the total votes used by the final lesion determination module is designed to maximize the accuracy and confidence of the final lesion identification, and this threshold can be calibrated according to different diseases or clinical needs.

[0018] Preferred motion stability index Through the first The second scan and the first The norm of the spatial transformation parameter difference between scans is calculated, and its calculation formula is as follows: ,in, Indicates the first Spatial transformation parameter vector for each scan Indicates the first Spatial transformation parameter vector for each scan represents the Euclidean norm of the vector; the vote weights are set as a negative correlation function of the motion stability index, where the negative correlation function ensures that the smaller the value of the motion stability index, the larger the corresponding vote weight.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. The system constructs a unified three-dimensional voting matrix and uses image sequence registration to ensure that multiple scan proposals are spatially aligned in a stable anatomical coordinate system. This allows the lightweight recognition algorithm to propose low-precision, high-recall suspected regions in a single scan. As the number of scans accumulates, the real lesion region will continuously receive votes due to its constant position, forming a significant peak. On the other hand, noise artifacts, due to their spatiotemporal randomness, have scattered votes that cannot be effectively accumulated. This mechanism naturally utilizes the statistical characteristics of noise, transforming the low-quality multiple scan data, which is traditionally regarded as interference, into an effective source of information for identifying lesions through statistical convergence. This improves the robustness of lesion recognition in low signal-to-noise ratio images in portable devices or rapid scanning scenarios.

[0021] 2. The system not only sets statistical thresholds at the level of vote accumulation, but also traces back all historical scan proposal image blocks that voted for candidate regions with high vote counts, and calculates the dispersion of their morphological feature vectors. True lesions should show inherent morphological consistency in multiple scans, while structural artifacts may fluctuate in morphology due to subtle changes even if their position is fixed. Therefore, only when a candidate region simultaneously meets the verification conditions of high vote accumulation and low morphological dispersion in two orthogonal dimensions is it finally confirmed as a lesion. This dual verification mechanism effectively distinguishes between true lesions with positional stability and structural artifacts, greatly reducing the risk of false positives and enhancing the reliability and specificity of diagnostic results.

[0022] 3. The system actively analyzes the spatial distribution pattern of votes discarded due to low vote counts in the 3D voting matrix. By identifying whether these discarded votes exhibit a specific topological structure, such as distribution along a specific direction or periodic characteristics, the system can infer potential systematic noise sources, such as motion artifacts in a specific direction or periodic interference from equipment. Once such a pattern is identified, the system will dynamically adjust the proposal logic in the subsequent suspected target proposal stage, and appropriately suppress proposals in areas that conform to the noise topological pattern. This self-learning mechanism based on historical discarded votes enables the system to perceive and adapt to systematic interference in the environment, continuously optimize proposal quality, and improve the adaptability and recognition accuracy of long-term operation.

[0023] 4. During each scan, the system synchronously acquires the electrical noise signal of the gradient coil power supply bus. By analyzing its disorder level (such as spectral entropy or energy variance), the system determines the patient's physiological stability (such as abnormal heart or respiratory rhythm). At the same time, the system also extracts transformation parameters reflecting the patient's motion amplitude and stability from the image registration process. Based on these two types of real-time perceived state information, the system dynamically adjusts the contribution weight of the suspected proposals generated in this scan in the voting matrix. When the physiological state is unstable or the motion amplitude is too large, the voting weight of the corresponding proposals will be significantly reduced or even ignored. This dynamic weighting mechanism ensures that the final voting result relies more on high-quality scan data obtained under good physiological and motion conditions, thereby improving the fairness and reliability of the overall decision-making.

[0024] 5. The core computational operations of the entire system—lightweight suspected region proposal, fast image registration based on stable references, integer voting accumulation, morphological feature extraction and consistency calculation, noise topology analysis, state awareness and weight adjustment—are all implemented using mature and computationally efficient algorithms. The resolution of the 3D voting matrix is ​​matched with the original image or obtained through interpolation and is persistently stored. This low-complexity design enables the system to run efficiently on general-purpose computing hardware, making it particularly suitable for deployment in primary healthcare institutions or long-cycle treatment scenarios requiring frequent follow-ups. By continuously updating the voting matrix and tracing historical morphological features, the system provides doctors with visualized information on lesion location, cumulative confidence, and morphological evolution, supporting long-term, dynamic, and low-cost monitoring and tracing of lesions. Attached Figure Description

[0025] Figure 1 This is a block diagram of the intelligent lesion identification and tracking system for magnetic resonance imaging of the present invention;

[0026] Figure 2 This is a graph showing how the vote weight of the system of the present invention changes with the number of scans;

[0027] Figure 3 This is a sequence diagram of the information flow and processing of the system of the present invention.

[0028] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail.

[0030] This application provides an intelligent lesion identification and tracking system for magnetic resonance imaging. Its overall architecture is built on a unified three-dimensional spatial representation and operates collaboratively through image sequence registration, weighted voting accumulation, and multi-dimensional verification. The entire workflow is designed as a closed-loop, adaptive data processing and decision-making system.

[0031] The core of this system lies in a three-dimensional voting matrix that spatially corresponds precisely to the patient's anatomical structure. The voxel resolution of this matrix is ​​configured to match the original acquisition resolution of the MRI images or obtained through linear interpolation. Serving as the central hub for information fusion, it is periodically stored and updated in non-volatile memory to continuously receive and accumulate spatial confidence information generated by the suspect proposal module. Before information input, the image sequence registration module first ensures that all MRI images from all sources are aligned to a unified anatomical coordinate system, laying the spatial foundation for effective vote accumulation. Simultaneously, the physiological stability adaptive module and the kinematic analysis and weighting module evaluate the data quality of each scan in parallel and dynamically adjust the contribution weight of the votes generated in that scan. Finally, after the votes from multiple scans have accumulated, the final lesion determination module, combined with morphological consistency verification, makes a final determination on the high-vote regions in the three-dimensional voting matrix, outputting a high-confidence lesion identification. To ensure no potential lesions are overlooked, in clinical scenarios where the signal-to-noise ratio of MRI images is significantly reduced due to the use of fast scanning protocols, the suspected target proposal module in this system employs a high-recall detection strategy. Its built-in lightweight recognition algorithm, such as an image processing model based on convolutional neural networks or morphological filters, with computational complexity limited to real-time execution by a low-power CPU, has its recognition threshold calibrated as follows: The system runs on an offline standard database containing various known lesion and artifact samples. The algorithm's internal decision threshold is scanned from high to low, and its recall rate in the database is calculated in real time. When the recall rate first reaches and stabilizes above 95%, this internal decision threshold is locked as the system's operating parameter. Thus, when processing each new MRI image frame, this module can generate suspected lesion region proposals with extremely high sensitivity, thereby maximizing the probability of capturing real lesion signals in the initial stage.

[0032] The core parameters upon which this system depends for operation are all determined through an offline calibration and online optimization procedure before deployment. First, by performing continuous scanning on a standard homogeneous phantom, the electrical noise of the gradient coil is collected and the statistical distribution of its energy variance is calculated. The sum of its mean and twice the standard deviation is taken as the upper limit of the hardware physiological stability baseline. Secondly, noise data from a group of healthy subjects at rest were collected, and the sum of the mean energy variance and twice the standard deviation was taken as the upper limit of the clinical physiological stability baseline. The smaller of the two values ​​is set as the final physiological stability threshold, while the weighting function... It is then concretized as a piecewise function, when the energy variance The value is 1.0 when it does not exceed the threshold, and the value is 1.0 when it exceeds the threshold. ,in The reference energy variance mean was obtained by collecting data from subjects under known strong perturbation physiological conditions. Meanwhile, the determination of the morphological consistency threshold was based on a validation database containing labeled real lesions and structural artifacts. The system calculated the Euclidean distance variance of the 27-dimensional morphological feature vector set of each sample in the database for each scan of the proposed image patch, thereby obtaining two variance value distributions corresponding to real lesions and structural artifacts, respectively. The threshold was finally set as the variance value corresponding to the maximum value of the Youden statistic of these two distributions, namely (sensitivity + specificity - 1), an indicator that measures the performance of the classification model.

[0033] The weight adjustment procedure of the systematic noise adaptive suppression module also follows deterministic computational logic. After identifying the potential principal axis directions of the discarded voting point cloud through principal component analysis, the system calculates the directional salience index. This metric is defined as the largest eigenvalue in principal component analysis of point clouds. Sum of all eigenvalues The ratio of the two values ​​is used to determine the proposal suppression weight. It is calculated as a function related to this indicator. ,in, This is a directional decision threshold set to 0.75. This function binds the suppression weights to the quantization intensity of the noise mode, and its output value range is strictly limited to between 5% and 20%. Finally, when accumulating votes on the three-dimensional voting matrix, the final vote weight is the contribution of the suspected lesion region proposal generated by each scan. It is determined by physiological stability weights Motion stability weights and systematic noise suppression weights The product determined by both factors is calculated using the following formula: This integrates quality assessments from different dimensions into a unified quantitative factor that affects the final vote.

[0034] To ensure a defined operational procedure for the suspected target proposal module and its subsequent morphological consistency verification, the lightweight recognition algorithm embedded in this module can be implemented through a combination of morphological operators. The specific process is as follows: First, a 3x3 median filter is applied to each aligned image to suppress discrete noise; second, a morphological closing-opening operation with a radius of 10 voxels is used to estimate and subtract the image background, thereby highlighting high-signal regions in the residual image; subsequently, all connected components in the residual image with an area between 20 and 500 voxels are identified as suspected proposals. The upper and lower limits of this area range are determined by the source... The distribution range, determined by statistical analysis of a standard reference database containing various known lesions, covers 99% of the actual lesion volume. In parallel, a morphological feature vector is extracted for each suspected proposal image patch. This vector is a 27-dimensional fixed-length vector, consisting of: seven Hu invariant moments characterizing region contour invariance, a 16-level normalized grayscale histogram describing pixel intensity distribution, and four Haralick texture features calculated from the grayscale co-occurrence matrix, quantifying image energy, contrast, correlation, and homogeneity, respectively. Each dimension is Z-score normalized using the mean and standard deviation calculated from all real lesion samples in the aforementioned standard reference database to form a dimensionless, standardized morphological descriptor that can be directly used for subsequent Euclidean distance variance calculation. To ensure the effective operation of the voting mechanism based on high-recall proposals, the spatial inconsistencies caused by differences in patient positioning or minor movements between multiple scans must be addressed; otherwise, the votes for real lesions will fail to form an effective peak due to spatial misalignment. Therefore, the system is equipped with an image sequence registration module. Before the suspected target proposal module intervenes, this module performs rigid body or affine transformation registration using the patient's skeletal structure as a stable reference. Specifically, the system uses the main skeletal contours in the first scan image as the reference three-dimensional anatomical coordinate system. For each subsequent scan, the system employs a fast registration algorithm to calculate the spatial transformation parameters required to align the skeletal structure in the current scan image with the reference coordinate system and applies this transformation to the entire image. Through this mandatory spatial alignment step, all subsequent suspected lesion region proposals and their voting behaviors can be accurately accumulated under a unified anatomical coordinate system.

[0035] Furthermore, considering the inherent fluctuations in data quality from a single scan, particularly the instability of the patient's physiological state (e.g., abnormal heart rhythm, respiratory rhythm, or hemodynamic patterns), which can directly lead to unpredictable artifacts in the images, the system integrates a physiological stability adaptive module. This module synchronously acquires the electrical noise signal on the gradient coil power supply bus of the MRI equipment during each scan and determines the patient's physiological stability based on the quantification of the signal's disorder or volatility—that is, the calculation of spectral entropy or energy variance. The system establishes a predetermined physiological stability baseline range—the spectral entropy or energy variance—representing physiological stability during the baseline calibration phase. In clinical applications, this... If the calculated value corresponding to a certain scan deviates from the baseline range, the system will dynamically reduce the vote weight contributed by the suspected lesion area proposal generated by that scan according to the degree of deviation. The weight is set in the range of 20% to 100%, thus giving higher decision weight to high-quality data obtained under good physiological conditions. Similar to physiological conditions, the patient's macroscopic movement during the scanning process is also a key factor affecting image quality and registration accuracy. To this end, the system has added a kinematic analysis and weighting module. This module directly uses the spatial transformation parameters calculated by the image sequence registration module when aligning each scan image, that is, the spatial transformation parameter vector including three-dimensional translation parameters and three-dimensional rotation parameters. A motion stability index is obtained by calculating the difference in the transformation parameter vector between two consecutive scans. ,in Indicates the first The spatial transformation parameter vector of the next scan, and The Euclidean norm of a vector is used by the system to quantify the amplitude or stability of patient movement during a given scan; a smaller... The value characterizes a relatively stable static state; correspondingly, when accumulating votes, the suspect proposal module is configured to apply a value related to this motion stability index. A negatively correlated function is used to dynamically set the vote weights, ensuring that proposals generated during strenuous activity have a significantly weakened influence on the final decision. After data accumulation and weighted voting across multiple scans, the final lesion identification module intervenes. Its primary task is to apply a high-confidence statistical threshold for preliminary screening, set at 75% of the total votes. This threshold aims to identify connected voxel regions with strong spatiotemporal stability that consistently receive votes in the vast majority of scans. This threshold can be calibrated according to different diseases or clinical needs. However, simply meeting the high vote count does not completely rule out the possibility of disease-related or special lesions. To address the structural artifacts generated by fixed anatomical structures and exhibiting positional stability, the system activates a morphological consistency verification module. For each high-confidence lesion region that passes the vote threshold screening, this module traces and acquires all suspected proposal image blocks from previous scans that voted for that region, extracts morphological feature vectors for each image block, and then calculates an Euclidean distance variance as a morphological consistency metric based on this set of morphological feature vectors. Finally, a high-confidence lesion region is only definitively identified as a lesion when its cumulative vote count is higher than the threshold and its morphological consistency metric is lower than a predetermined morphological consistency threshold.

[0036] To ensure that the dynamic weighting of physiological and motor states follows a closed mathematical procedure, the final vote weights for each scan are determined. Weights based on physiological stability Weights related to motion stability Multiplying them together yields the result, i.e. Among them, motion stability weights Based on a motion stability index The piecewise function is determined when: Less than or equal to the stable motion threshold determined during the hardware baseline calibration phase hour, The value is 1.0, when Greater than But less than the maximum tolerable motion threshold hour, ,when Greater than or equal to hour, The value is set to 0.2; similarly, the physiological stability weight... Also based on the current electrical noise energy variance With baseline stability range The piecewise function is determined when When within this range, The value is set to 1.0. When it exceeds this range, its weight value will be linearly reduced according to the relative magnitude of the deviation, but the minimum will not be lower than 0.2. In addition, the clinical calibration procedure for the cumulative vote threshold (e.g., 75%) used by the final lesion determination module is as follows: For a specific disease, firstly, a test sequence set containing at least 50 confirmed positive cases is constructed. Then, the threshold is scanned in the interval [50%, 95%] with a step size of 1%, and the true positive rate and false positive rate on the test set are calculated for each threshold point to plot the receiver operating characteristic curve (ROC curve). Finally, the curve is used to determine the ROC curve. Statistic, which is defined as The percentage of votes corresponding to the point that reaches the maximum value is determined as the optimal working threshold for that specific disease. Furthermore, to address certain persistent systemic interferences, the system also includes a systemic noise adaptive suppression module. This module actively identifies the topological patterns of systemic noise by analyzing the spatial distribution of votes discarded in the three-dimensional voting matrix due to vote counts falling below a predetermined low threshold. Its analysis process aims to reveal the intrinsic structure of systemic artifacts, including principal component analysis of the discarded vote spatial point cloud to identify its principal axis direction, and Fourier transform of the point cloud projection onto a specific anatomical axis to identify periodic frequency peaks. If such a topological pattern is identified, the system adjusts the proposal logic during the subsequent suspected target proposal stage, reducing the proposal weight of regions conforming to the topological pattern by at least 5% and no more than 20%. This self-learning and suppression mechanism enables the system to adapt to specific hardware and clinical environments, continuously optimizing its recognition accuracy.

[0037] Example 1: This example demonstrates the operation of the general technical solution described in a specific clinical monitoring scenario. In a long-term follow-up application for the adjuvant radiotherapy effect after surgery in a brain tumor patient, the patient needs to undergo up to fifteen rapid MRI scans over several months to monitor potential micro-recurrences. This scenario presents two challenges: First, to shorten the time per scan, the rapid scanning protocol inevitably leads to a decrease in the image signal-to-noise ratio, resulting in a decline in the performance of traditional recognition algorithms that rely on fine texture features; Second, a metal implant in the patient's mouth will create a highly positionally stable structural feature at a specific scan level. Artifacts, whose morphology resembles early, small lesions, pose a significant challenge to automated identification due to false positives. At the start of this monitoring cycle, the system uses the patient's first high-quality scan as a baseline, locking the skull as a stable anatomical coordinate system through an image sequence registration module. In each subsequent rapid scan, the system first aligns spatially with the baseline coordinate system based on the skeletal structure. Then, a suspect target proposal module employing a high recall strategy generates proposals for suspected lesion regions on each low signal-to-noise ratio image, and the corresponding votes are counted in a unified three-dimensional voting matrix. This mechanism resolves the traditional contradiction between detection sensitivity and specificity, allowing for single-scan detection... The identification process is low-specificity. By shifting the burden of identification from the spatial domain judgment of a single high-quality image to the spatiotemporal statistics of multiple low-quality images, a foundation is laid for achieving high-specificity identification through statistical convergence without sacrificing initial detection sensitivity. Midway through the monitoring cycle, if a patient's respiratory rhythm becomes abnormal during a scan due to anxiety, the physiological stability adaptive module analyzes the spectral entropy of the electrical noise signal on the gradient coil power supply bus to determine that the scan data is significantly affected by physiological noise. Therefore, based on a preset weighting function, the weight of all proposal votes generated in that scan is reduced to 30%. Then, in another scan… When the patient makes a slight head movement, the motion stability index value calculated by the kinematic analysis and weighting module also increases significantly. The system then reduces the vote weight for that scan accordingly. Here, the two independent verification dimensions of physiological state perception and kinematic analysis show a synergistic effect. They act as two parallel quality control checkpoints, and their output dynamic weight values ​​directly modulate the contribution of vote accumulation. Together, they ensure that the votes injected into the three-dimensional voting matrix mainly come from high-quality and high-stability scan data. This greatly purifies the statistical sample, allowing the cumulative peak of the real lesion signal to stand out more quickly and clearly from the statistical background of random noise.

[0038] After all fifteen scans were completed, the final lesion identification module analyzed the three-dimensional voting matrix and found that the cumulative votes for two brain regions both exceeded the 75% threshold of the total votes. One region was a real, small recurrent lesion, and the other was a stable artifact caused by metal implants. At this point, the system automatically triggered the morphological consistency verification module. This module traced and acquired all historical suspected proposal image patches that voted for these two high-vote regions and calculated the Euclidean distance variance of their respective morphological feature vector sets. The results showed that the real lesion region, due to the constancy of its inherent biological morphology, had its fifteen morphological feature vectors highly clustered in the feature space, exhibiting extremely low variance values. In contrast, although the spatial position of the metal artifact region was stable, its morphology underwent subtle changes due to slight differences in the patient's head tilt during each scan, resulting in significant dispersion and high variance values ​​in its morphological feature vectors. Ultimately, only the real lesion region... Simultaneously meeting the dual criteria of high vote count and low morphological variance, it was identified by the system as a high-confidence lesion. The artifact region was excluded because its morphological consistency metric failed to fall below the predetermined morphological consistency threshold, thus avoiding a potential clinical misjudgment. The overall architecture of the system did not attempt to combat noise by increasing the complexity of the recognition algorithm for a single image. Instead, it redefined the problem itself, changing the challenge from how to find an uncertain spatial pattern in a single noisy image to how to verify a signal that is constant in both time and space and in morphology in a series of image sequences with matching errors and quality fluctuations. It does not rely on the perfection of any single acquisition, but rather transforms the uncertainty itself in multiple acquisitions into a source of certainty for the final decision through mechanisms such as statistical voting, dynamic weighting, and multi-dimensional verification. This reflects a design principle of combating multi-source interference through information fusion and consistency testing.

[0039] Example 2: This example aims to verify, through a controlled experiment, the performance gain of the present invention's technical solution compared to traditional single-image recognition methods when faced with signal-to-noise ratio (SNR) degradation and motion artifact interference. To achieve this, the experiment was conducted on a semi-virtualized test platform combining real clinical data and simulated artifacts. The baseline data originated from a set of high-resolution, high SNR brain MRI images of healthy volunteers, verified by repeated scans to be free of abnormalities. Three spherical simulated lesions, with diameters between 3 and 5 mm and signal intensity characteristics consistent with early gliomas, were digitally implanted at three different anatomical locations in the three-dimensional image data, thus constructing a gold standard dataset with accurate ground reality. Based on this, different levels of Ricean noise were applied to this gold standard dataset to simulate SNR degradation, and random, minute affine transformations were applied to each frame of the sequence to simulate involuntary patient movements, generating a test sequence containing twenty frames. Frames 7 and 15 were additionally superimposed with localized high-intensity artifacts to simulate transient interference caused by physiological instability. The number of frames in this test sequence... The fundamental technical consideration in setting this up lies in achieving an optimal balance between the sufficiency of the statistically convergent sample and the computational efficiency of the experiment. The decision rule is that the total number of frames should be significantly greater than the preset number of frames with strong interference to ensure that the effective data is statistically dominant. Given that this experiment presets 2 frames with strong interference and several frames with moderate motion interference, The settings provide a reasonable sample basis for verifying the robustness of this scheme.

[0040] The experiment compared the performance of the intelligent lesion identification and tracking system for MRI of this invention with a representative conventional lesion identification method based on deep learning. System A used a three-dimensional convolutional neural network pre-trained on a high-quality image dataset to independently identify lesions in each frame of the test sequence and performed logical union processing on the identification results of all frames. System B deployed all modules of this invention and applied them to the entire test sequence. The core observation indicators of the experiment were the detection sensitivity of the two systems for the three implanted lesions and the number of false positive lesions under different interference levels. During the experiment, it was observed that as the image signal-to-noise ratio decreased and motion artifacts were introduced, the performance of System A showed a sharp degradation, with a significant decrease in its detection sensitivity and a substantial increase in the number of false positive lesions. In contrast, the performance of System B showed high stability, with its sensitivity maintained at a high level and the number of false positives effectively suppressed. Table 1 shows the key performance data obtained under two typical interference conditions.

[0041] Table 1: Performance comparison of the two systems under different interference conditions.

[0042]

[0043] The analysis of the data in Table 1 clarifies the intrinsic mechanism of System B. Under conditions of low signal-to-noise ratio and motion artifacts, the average vote weight of System B decreases from 92.5% to 78.3%. This weight change quantitatively reflects the role of the kinematic analysis and weighting module, which accurately identifies the inter-frame inconsistencies introduced by affine transformation and generally reduces the vote contribution of frames with motion interference. For the strongly interfering frame #7 with superimposed local high-intensity artifacts, System B assigns a fixed vote weight of 25.0%. This value is the direct output of the physiological stability adaptive module after quantifying and evaluating the abnormal electrical noise signal of the frame. This minimizes the impact of the strongly interfering frame on the final decision. It is this synergistic effect of accumulating real signals through spatiotemporal voting and suppressing noise and artifact interference through a dynamic weighting mechanism that enables System B to robustly extract high-confidence lesion information from a series of low-quality images.

[0044] Example 3: This example combines Figures 1 to 3 This paper describes the implementation of an intelligent lesion recognition and tracking system for magnetic resonance imaging, such as... Figure 1As shown, the system acquires multi-sequence image data from the first to the Nth MRI scan. After data input, the image sequence registration module processes the images. This module uses the skeletal structure as a reference and performs rigid body / affine transformation registration to achieve alignment of the unified anatomical coordinate system, ensuring that all scanned images are spatially aligned. The registered image data is then transmitted to the suspected target proposal module, which runs a lightweight recognition algorithm and sets a high recall strategy (≥95%) to generate suspected lesion regions. Simultaneously, during each scan, the physiological stability adaptive module collects gradient coil electrical noise and calculates spectral entropy / energy variance to determine the physiological stability state, thereby dynamically adjusting the vote weights (20%-100%). In parallel, the kinematic analysis and weighting module extracts spatial transformation parameters and calculates motion stability indices, thereby dynamically weighting the vote contribution. The suspected lesion region proposals generated by the above modules and the weighted votes are accumulated in a three-dimensional voting matrix. This matrix records the accumulated votes for each voxel, the unified anatomical coordinate system, and is continuously updated. The system stores and accumulates votes in a spatiotemporal voting mechanism. The weight adjustment results of the physiological stability adaptive module and the kinematic analysis and weighting module are fed back to the weight adjustment stage, affecting the contribution of the proposal of the suspected lesion region in the three-dimensional voting matrix. After the voting is completed, the final lesion determination module intervenes, screens and identifies connected voxel regions or performs analysis based on the accumulated vote threshold (≥75%), and performs double verification standard, finally outputting high-confidence lesions. Before this, the morphological consistency verification module traces historical proposal image blocks, extracts morphological feature vectors, calculates Euclidean distance variance, and verifies morphological consistency. Its results are part of the double verification performed by the final lesion determination module. In addition, the systematic noise adaptive suppression module analyzes the spatial distribution of discarded votes, identifies topological patterns (including principal component analysis / Fourier transform), and dynamically adjusts the proposal weights. The feedback optimization of this module also affects the proposal process of the suspected target. Finally, the system outputs high-confidence lesion identification results, including lesion location information, accumulated confidence, morphological evolution process, and long-term dynamic monitoring information.

[0045] like Figure 2 As shown, Figure 2The graph depicts the changes in system voting weight (vote weight (%)) with the number of scans under different interference conditions. The horizontal axis represents the number of scans, and the vertical axis represents the vote weight (%). The graph shows points with normal scans (white circles), points with strong physiological interference (black circles), and points with motion interference (gray circles). At the 7th scan, strong interference (25%) occurred, causing the vote weight to drop significantly to 25%. At the 11th scan, motion interference (30%) occurred, and the vote weight dropped to 30%. The right side of the graph also indicates an average of 78.3%, meaning that the average vote weight for all scans is 78.3%. This graph intuitively shows how the physiological stability adaptive module and the kinematic analysis and weighting module dynamically adjust the vote weight based on perceived interference (such as physiological instability or motion), thereby reducing the impact of low-quality data on the final lesion identification results.

[0046] like Figure 3 As shown, the process begins with the MRI device issuing a start scan command and activating the gradient field. During the scan, the gradient coils process and transmit electrical noise signals in parallel to the physiological stability module. Simultaneously, the MRI device transmits the original image to the image registration module and the physiological stability module. After receiving the electrical noise signal, the physiological stability module calculates the spectral entropy / energy variance and determines the physiological stability state. Then, it transmits the vote weights (20%-100%) to the suspect proposal module. After receiving the original image, the image registration module identifies the skeletal structure, calculates the spatial transformation parameters, and performs image alignment. It then transmits the aligned image to the suspect proposal module. After receiving the aligned image and the vote weights, the suspect proposal module generates the suspected lesion region, applies weight adjustments, and weighted cumulative voting into the three-dimensional voting matrix.

[0047] Example 4: This example elaborates on the internal operating procedures and calibration logic of key parameters of the systematic noise adaptive suppression module. In an application scenario of a portable MRI device deployed in an intensive care unit environment, due to the continuous operation of other life support systems nearby, a weak periodic stripe artifact originating from electromagnetic interference and distributed along a specific spatial direction is present in the image sequences acquired by the device. This artifact has extremely low signal strength in a single frame image, but because its position and direction are highly repeatable in multiple scans, it poses a potential challenge to the spatiotemporal voting accumulation-based identification mechanism in this invention, that is, this systematic artifact may be incorrectly accumulated. For regions with high vote counts, to address this challenge, the systematic noise adaptive suppression module is configured to periodically initiate its self-learning and suppression program after the tenth scan, following a specific number of initial scans. This program begins by defining discarded votes. In this embodiment, the definition procedure is as follows: all voxel locations in the 3D voting matrix with a cumulative vote count of 1 are identified as discarded vote points. This rule is based on the fact that regions receiving only a single vote after ten scan opportunities lack spatiotemporal constancy and should be classified as random noise. Furthermore, the system extracts the 3D coordinates of all these discarded vote points to form an input point cloud for analyzing potential topological patterns. .

[0048] The module's analysis process includes two parallel computational paths: one is directional pattern analysis, where the system analyzes the point cloud. Perform principal component analysis to calculate its three eigenvalues. and the corresponding eigenvectors, where the largest eigenvalue is... The corresponding feature vector is the potential principal axis direction of the point cloud distribution. To determine whether this directionality is significant, the system calculates the directional significance index, which is defined as follows: The system only identifies a significant directional noise pattern when the index exceeds a directional judgment threshold set at 0.75; the second is periodic pattern analysis, where the system analyzes the point cloud... Projected to , , Three standard anatomical axes generate three one-dimensional density distribution curves. A Fast Fourier Transform (FFT) is performed on each curve to identify any significant periodic frequency peaks. The significance of a frequency peak is determined by its peak-to-noise ratio (PSNR), meaning the peak amplitude must be higher than five standard deviations above the average level of the spectral background noise to be considered a valid periodic pattern. If any of the above analysis paths identifies a significant topological pattern, the system quantifies the pattern's strength and determines a dynamic suppression weight accordingly. If a directional pattern is confirmed as significant, its suppression weight... The calculation logic is set as a linear function positively correlated with the directional significance index. This function ensures that the suppression weight ranges between 5% and 20%. In all subsequent scans, after generating a suspected lesion region proposal, the suspected target proposal module first calculates the principal direction of the proposal region. If the angle between this direction and the principal axis direction of the identified systemic noise is less than a matching tolerance set to 10 degrees, the votes contributed by this proposal will be multiplied by [the specified value]. By attenuating the artifacts, this module can autonomously learn the spatial topological features of stripe artifacts caused by electromagnetic interference in the aforementioned intensive care unit application scenario. In the subsequent recognition process, it can suppress the regions that match the artifact features, thereby avoiding interference from systemic environmental noise without affecting the ability to detect real lesions.

[0049] Example 5: After integrating the system software into a new MRI device, before its clinical use, a hardware baseline calibration procedure is first performed. This procedure is completed by performing multiple rapid scans on a standardized, homogeneous phantom containing no target objects. During this process, the physiological stability adaptive module continuously acquires the electrical noise signal on the gradient coil power supply bus of the device under stable load and calculates the statistical distribution of its spectral entropy or energy variance. Based on the mean and standard deviation of this distribution, a predetermined physiological stability baseline range specific to this device is established. At the same time, the systematic noise adaptive suppression module analyzes the spatial distribution of discarded votes generated in this series of scans, originating from the device's own dark current and environmental interference, to identify and record any weak systematic artifact topology patterns inherent to the device, providing an initial background reference for subsequent adaptive suppression.

[0050] After completing the hardware baseline calibration, the system enters an initial online parameter optimization phase. This phase covers the initial fifty clinical patient scan sequences performed by the device. During this period, the morphological consistency verification module continuously collects all real lesion areas manually confirmed by radiologists and their corresponding suspected proposal image blocks in each scan. Based on this continuously expanding, highly confident positive sample database, statistical methods are used to iteratively optimize and eventually converge to a predetermined morphological consistency threshold that can best distinguish between real lesions and structural artifacts. The systematic noise adaptive suppression module also utilizes a large amount of discarded voting data generated in these fifty sequences to perform its first complete topological pattern analysis. Based on the analysis results, dynamic suppression weights and their triggering conditions are set. These are all extended implementation methods known to those skilled in the art.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent lesion identification and tracking system for magnetic resonance imaging, characterized in that, include: A three-dimensional voting matrix is ​​configured to record the cumulative votes for each voxel in a three-dimensional space corresponding to the patient's anatomical structure. A suspected target proposal module is configured to: receive the current cumulative vote count of a three-dimensional voting matrix, and run a lightweight recognition algorithm for each MRI image obtained from multiple scans to generate a suspected lesion region proposal; add votes to the corresponding positions in the three-dimensional voting matrix according to the voxel positions covered by the suspected lesion region proposal; the recognition threshold of the lightweight recognition algorithm is set to ensure that the recall rate of the suspected lesion region proposal is not less than 95%. An image sequence registration module is configured to: receive the currently scanned MRI image and, before the suspected lesion region proposal module generates a proposal, perform rigid body or affine transformation registration using the patient's skeletal structure as a stable reference to spatially align the coordinate system of the currently scanned image with that of the first scanned image. A physiological stability adaptive module is configured to: synchronously acquire electrical noise signals on the gradient coil power supply bus of the MRI equipment during each scan; determine the patient's physiological stability state corresponding to the scan based on the spectral entropy or energy variance of the electrical noise signals, including comparing the spectral entropy or energy variance with a predetermined physiological stability baseline range; and dynamically adjust the vote weight of the suspected lesion region proposal generated by the scan in the three-dimensional voting matrix according to the determined physiological stability state, with the vote weight set to 20% to 100%. A final lesion determination module is configured to: after the votes from multiple scans have been accumulated, receive the final accumulated votes from the three-dimensional voting matrix, and identify connected voxel regions whose accumulated votes are higher than 75% of the total number of votes as high-confidence lesions.

2. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 1, characterized in that, The system also includes a morphological consistency verification module, configured to: before the final lesion determination module identifies a high-confidence lesion region as the final lesion, trace and obtain the morphological feature vectors of the suspected proposal image blocks from all previous scans that voted for the high-confidence lesion region; calculate a morphological consistency metric based on the set of morphological feature vectors, where the morphological consistency metric is the variance of the Euclidean distance of the set of morphological feature vectors; and the final lesion determination module only identifies the high-confidence lesion region as the final lesion when the cumulative number of votes for the high-confidence lesion region is higher than 75% of the total number of votes and the morphological consistency metric is lower than a predetermined morphological consistency threshold.

3. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 1, characterized in that, The system also includes a systematic noise adaptive suppression module, configured to: analyze the spatial distribution of discarded votes with a number of votes below a predetermined low threshold in the three-dimensional voting matrix to identify the existence of topological patterns of systematic noise, including periodic frequency peaks presented by the principal axis direction of the point cloud distribution or the projection of the point cloud onto a specific anatomical axis. Furthermore, if a topological pattern is identified, in the subsequent suspected target proposal stage of the scanned image, the suspected target proposal module adjusts the proposal logic according to the characteristics of the topological pattern, so as to reduce the proposal weight of the region that conforms to the topological pattern by at least 5% and no more than 20%.

4. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 1, characterized in that, The system also includes a kinematic analysis and weighting module, configured to obtain spatial transformation parameters calculated for aligning each scanned image from the image sequence registration module. The spatial transformation parameters include three-dimensional translation parameters and three-dimensional rotation parameters. Based on the time series of spatial transformation parameters, a motion stability index is determined for each scan. The motion stability index reflects the amplitude or stability of the patient's movement during the scan. Furthermore, when accumulating votes, the suspected target proposal module dynamically weights the votes contributed by the suspected lesion area proposals generated by the scan based on the motion stability index.

5. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 1, characterized in that, The lightweight identification algorithm used in the suspect target proposal module is an image processing model based on convolutional neural networks or morphological filters. This model is optimized to quickly identify patches that differ from surrounding tissues in low signal-to-noise ratio images, and its computational complexity is limited to a level that can be executed in real time by a low-power central processing unit.

6. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 1, characterized in that, The physiological stability adaptive module determines the patient's physiological stability state, including at least one of cardiac arrhythmia, respiratory rhythm abnormality, or hemodynamic pattern abnormality. The calculation of the spectral entropy or energy variance of the electrical noise signal is used to quantify the degree of disorder or the severity of fluctuation of the signal.

7. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 1, characterized in that, The cumulative vote threshold of 75 percent of the total votes used by the final lesion determination module is designed to maximize the accuracy and confidence of the final lesion identification, and this threshold can be calibrated according to different diseases or clinical needs.

8. The intelligent lesion identification and tracking system for magnetic resonance imaging according to claim 4, characterized in that, Motion stability index Through the first The second scan and the first The norm of the spatial transformation parameter difference between scans is calculated, and its calculation formula is as follows: ,in, Indicates the first Spatial transformation parameter vector for each scan Indicates the first Spatial transformation parameter vector for each scan represents the Euclidean norm of the vector; the vote weights are set as a negative correlation function of the motion stability index, where the negative correlation function ensures that the smaller the value of the motion stability index, the larger the corresponding vote weight.

Images