System for detecting ramp lesion associated with anterior cruciate ligament injuries using deep learning technology and operating method thereof

Abstract

The present invention relates to a system for detecting a ramp lesion associated with anterior cruciate ligament injuries using deep learning technology for automatically detecting and diagnosing an inner ramp posterior rupture in an MRI image. The system includes a deep learning model based on a multi-layer neural network, and the system may preprocess image data through a preprocessing step for improving the resolution and contrast of an MRI image, and extract structural features of a ramp from the preprocessed image data through the deep learning model.

Description

A system for detecting meniscus lesions associated with anterior cruciate ligament injuries using deep learning technology and its operating method.

[0001] The present invention relates to a system that automatically detects and diagnoses a ramp lesion accompanying anterior cruciate ligament (ACL) injury in a patient using deep learning technology.

[0002] A ramp lesion is a longitudinal tear occurring at the meniscus and meniscocapsular junction, posterior to the medial meniscus. Ramped lesions are a relatively common injury, occurring in 9% to 40% of patients with anterior cruciate ligament (ACL) injury. When a ramp lesion is combined with an ACL injury, it is known to worsen anterior translation and rotational instability of the tibia. To address this issue, meniscus repair can be performed concurrently with ACL reconstruction.

[0003] However, arthroscopic suturing of meniscus lesions is technically challenging, increasing the surgical time required for ACL reconstruction. Furthermore, post-suturing rehabilitation, including joint angle restrictions and weight-bearing restrictions, must be performed more slowly than with ACL reconstruction. Therefore, preoperatively identifying whether an ACL injury and meniscus lesion coexist is crucial for both surgeons and patients.

[0004] The standard diagnostic method for diagnosing meniscus tears is magnetic resonance imaging (MRI). However, the accuracy of MRI is not high, with various studies reporting accuracy rates ranging from 53.9% to 84.6%. This leads to inconsistent diagnoses, and further improvement in accuracy is needed.

[0005] Orthopedic surgeons utilize imaging studies in the diagnosis of meniscus lesions, taking into account clinical risk factors. These include clinical indicators such as gender, age, mechanism of injury, and timing of injury. Radiographic indicators include osteophyte swelling of the proximal medial superior tibialis and the presence of an associated lateral meniscus tear. However, ACL injuries often present with other types of meniscus tears in addition to meniscus lesions, but for the same reasons, they are often not diagnosed preoperatively.

[0006] Advances in artificial intelligence technology are offering innovative opportunities in the medical field, and machine learning and deep learning technologies are already being successfully applied to the diagnosis and prediction of diseases. In particular, these technologies have demonstrated results comparable to those of experienced radiologists in the field of deep learning for image recognition. Against this backdrop, studies have also reported developing predictive models for risk factors associated with meniscus lesions using deep learning.

[0007] However, existing research using AI to diagnose meniscus lesions has several limitations. The typical research process begins with data measurement followed by data analysis. While previous studies have utilized AI for data analysis, the process of determining the presence of meniscus lesions in MRI, the process of measuring the data, relies on subjective human interpretation, making it difficult to guarantee consistent accuracy.

[0008] To address these issues, it is necessary to develop a system that can produce consistent results with high accuracy by applying deep learning technology not only in data analysis but also in the data measurement process.

[0009] [Prior Art Literature]

[0010] [Patent Document]

[0011] (Patent Document 1) Republic of Korea Patent Publication No. 10-2022-0167061 (published on December 20, 2022)

[0012] The purpose of the present invention is to provide a system for detecting meniscus lesions accompanying anterior cruciate ligament injury using deep learning technology, which can improve the accuracy of MRI diagnosis of meniscus lesions by utilizing deep learning technology and further increase the accuracy of diagnosis by fusing risk factors related to meniscus lesions with image data processed by deep learning.

[0013] A first aspect of the present invention for solving the above-described problem relates to a system for detecting meniscus lesions accompanying anterior cruciate ligament injury using deep learning technology. The system is a system for detecting meniscus lesions accompanying anterior cruciate ligament injury using deep learning technology for automatically detecting and diagnosing posterior rupture of the medial meniscus in MRI images, wherein the system includes a deep learning model based on a multilayer neural network, and the system preprocesses image data through a preprocessing step for improving the resolution and contrast of the MRI image, and extracts structural features of the meniscus from the preprocessed image data through the deep learning model.

[0014] According to an embodiment of the present invention, the system can improve the diagnostic accuracy of meniscus lesions by integrating risk factors including gender, age, injury mechanism, and injury time, and can improve the comprehensive accuracy of diagnosis by fusing the risk factors and image data through multivariate analysis.

[0015] According to an embodiment of the present invention, an area including a meniscus lesion is labeled in the MRI image, and the system increases the diversity and quantity of input data using data augmentation techniques including rotation, flipping, and scale adjustment, and verifies the quality of the labeled data so that it can be used for training the deep learning model.

[0016] According to an embodiment of the present invention, the system is integrated with a cloud-based platform to remotely analyze the MRI image in real time and provide a user interface that is easily usable by a doctor, and the user interface may include functions for interpretation of diagnostic results, customization, storage and sharing of diagnostic records.

[0017] A second aspect of the present invention relates to a method for operating a system for detecting meniscus lesions associated with anterior cruciate ligament injuries utilizing deep learning technology to automatically detect and diagnose posterior lesions of the medial meniscus in MRI images. The method may include: a) a preprocessing step for preprocessing image data through a preprocessing step for improving the resolution and contrast of the MRI image; and b) a feature extraction step for extracting structural features of the meniscus from the preprocessed image data.

[0018] According to an embodiment of the present invention, the step a) may include: a-1) a data collection step of collecting MRI images of patients with anterior cruciate ligament injuries who may have meniscus lesions; a-2) a data labeling step of analyzing the MRI images and labeling areas containing meniscus lesions; a-3) a data normalization step of consistently processing the MRI image data to optimize the learning performance of a deep learning model; and a-4) a data augmentation step of diversifying and increasing the amount of MRI image data to improve the learning performance of the deep learning model.

[0019] According to an embodiment of the present invention, the step a-4) increases the diversity and quantity of input data by using a data augmentation technique including rotation, flipping, and scale adjustment, and verifies the quality of the labeled data so that it can be used for training the deep learning model.

[0020] According to an embodiment of the present invention, the step b) may include: b-1) a preprocessed data preparation step of collecting MRI image data preprocessed in the step a) and converting it into a format suitable for deep learning model training; b-2) a region of interest setting step of setting a region in the MRI image that is likely to include a meniscus lesion as a region of interest; b-3) a model selection step of selecting an optimal deep learning model to accurately extract structural features of the meniscus lesion; and b-4) a feature extraction step of automatically extracting structural features of the meniscus lesion from the preprocessed MRI image data.

[0021] According to an embodiment of the present invention, step b) further includes step b-5) of integrating feature data to generate a single integrated feature vector by integrating feature vectors extracted from multiple sources, and step b-5) can increase the diagnostic accuracy of meniscus lesions by integrating risk factors including gender, age, injury mechanism, and injury time, and can improve the comprehensive accuracy of diagnosis by fusing the risk factors and image data through multivariate analysis.

[0022] According to an embodiment of the present invention, step b) further includes a final feature data utilization step b-6) for predicting the presence or absence of a meniscus lesion based on the integrated feature vector and deriving the result, and step b-6) can be integrated with a cloud-based platform to remotely analyze the MRI image in real time and provide a user interface that is easily usable by a doctor.

[0023] According to the present invention, deep learning technology can be utilized to dramatically improve the diagnostic accuracy and efficiency of meniscus lesions (ramp lesions) associated with anterior cruciate ligament (ACL) injuries. The deep learning diagnostic system according to the present invention integrates MRI images and clinical data to automatically detect and diagnose meniscus lesions, thereby providing the following various benefits.

[0024] First, the present invention can significantly improve diagnostic accuracy. By automatically detecting posterior medial meniscus tears in MRI images using a deep learning model, it can reduce variability resulting from subjective interpretation and increase consistency. Furthermore, by incorporating clinical risk factors such as gender, age, injury mechanism, and timing, it can enhance the overall diagnostic accuracy, enabling more precise diagnoses.

[0025] Second, it can contribute to shortening diagnosis times and increasing efficiency. Automated image analysis utilizing deep learning technology can save medical professionals time and effort and shorten diagnosis times. Specifically, it can enhance efficiency in the medical field by remotely analyzing MRI images in real time and providing diagnostic results through a cloud-based platform.

[0026] Third, it can improve patient management and treatment outcomes. Accurate diagnosis allows for a clear assessment of the patient's condition and the development of an optimal treatment plan. This contributes to the optimization of the rehabilitation process and the effective management of rehabilitation plans, including weight-bearing restrictions. Ultimately, this can improve patient outcomes and shorten recovery times.

[0027] Fourth, the present invention can also contribute to reducing medical costs. By reducing misdiagnosis through a highly accurate and automated diagnostic system, unnecessary additional testing and treatment costs can be reduced. Furthermore, through efficient use of medical resources, overall medical costs can be reduced.

[0028] Finally, the present invention offers significant economic and industrial benefits. Data collected through deep learning models can be utilized for future research and analysis, contributing to the development of disease prediction and diagnosis models based on medical data. Furthermore, the developed technology can be applied to the diagnosis of other types of meniscus tears and musculoskeletal disorders, and can be expanded into various medical fields. This will accelerate the development and distribution of AI-based medical solutions, create new markets, and significantly contribute to the advancement of the medical technology industry.

[0029] FIG. 1 is a diagram illustrating each step of a method for operating a system for detecting meniscus lesions accompanying anterior cruciate ligament injury using deep learning technology according to the present invention.

[0030] Figure 2 is a diagram illustrating a process of integrating MRI images and risk factor elements into a deep learning model in the present invention.

[0031] Hereinafter, specific details for implementing the present invention will be described with reference to the attached drawings. In describing the present invention, detailed descriptions of related known functions that are obvious to those skilled in the art and that may unnecessarily obscure the gist of the present invention will be omitted.

[0032] The present invention utilizes deep learning technology to detect meniscus lesions associated with anterior cruciate ligament injuries, and is a system for automatically detecting and diagnosing posterior medial meniscus tears in MRI images. The system comprises a deep learning model based on a multilayer neural network.

[0033] Figure 1 is a diagram illustrating each step of the method for operating a system for detecting meniscus lesions associated with anterior cruciate ligament injuries using deep learning technology according to the present invention. Figure 2 is a diagram illustrating the process of integrating MRI images and risk factors into a deep learning model according to the present invention.

[0034] Referring to FIGS. 1 and 2, the system performs a preprocessing step (S100) of preprocessing image data through a preprocessing step for improving the resolution and contrast of MRI images, and a feature extraction step (S200) of extracting structural features of the meniscus from the preprocessed image data through the deep learning model. Each step may be performed by a control unit such as a computer or processor.

[0035] The preprocessing step (S100) includes a data collection step (S110), a data labeling step (S120), a data normalization step (S130), and a data augmentation step (S140).

[0036] The data collection step (S110) is a step for collecting MRI images of patients with anterior cruciate ligament injury who have a possibility of having a ramp lesion.

[0037] First, we selected the target patients. Patients who visited our hospital and underwent ACL reconstruction for anterior cruciate ligament (ACL) injuries or revision reconstruction for re-injuries were included. MRI images obtained preoperatively were used if they had adequate resolution and quality for assessing meniscus lesions.

[0038] MRI images obtained preoperatively were excluded if they were inadequate for assessing meniscus lesions (e.g., low-resolution images, external film scan images) or if the medial meniscus status was not assessed at the time of surgery. The collected MRI images included T1-weighted images, T2-weighted images, proton density images, and sagittal views. The collected images were matched with the surgeon's surgical findings to enhance accuracy.

[0039] At this time, the patient's clinical risk factors, radiological risk factors, and MRI evaluation items were considered.

[0040] Clinical risk factors considered included gender, age, site of injury, time of injury, and degree of instability (anterior drawer grade, Lachman grade, pivot shift grade) identified in the preoperative physical examination.

[0041] Radiological risk factors considered included X-ray evaluation, degree of arthritis (Kellgrene-Lawrence grade), lower extremity alignment (hip-knee-ankle axis), and degree of anterior displacement identified on preoperative stress X-ray.

[0042] MRI evaluation items included the presence and degree of medial proximal tibial bone edema, the presence and degree of lateral compartment bone edema (pivot shift injury), medial and lateral tibial slope, and the presence of a longitudinal tear at the medial meniscus-synovial joint junction (ramp lesion).

[0043] The data labeling step (S120) is a step of analyzing the collected MRI images to indicate whether there is a meniscus lesion and labeling the corresponding area.

[0044] First, the area containing the meniscus lesion is labeled on the MRI image. The LabelMe program can be used for this. Orthopedic surgeons and radiologists who are not involved in the surgery can independently review the MRI images and participate in the labeling process. They can assess variables such as the presence of meniscus lesions, the presence and extent of bone edema in the lateral compartment, and the medial and lateral meniscus-tibial inclinations.

[0045] Next, the quality of the labeled MRI images is verified. This process verifies the accuracy and consistency of the labeling and, if necessary, performs image enhancements. Image enhancements include resolution enhancement and noise removal. These processes can improve the quality of the labeled data and enhance the model's learning performance.

[0046] Next, a range of interest (ROI) is established to most appropriately assess meniscus lesions. This helps the deep learning model accurately extract important features during training. The ROI is carefully determined, taking into account the location and size of the lesion, thereby maintaining data consistency.

[0047] The data normalization step (S130) consistently processes the collected MRI image data to optimize the learning performance of the deep learning model. This step applies various normalization techniques to ensure data quality and enable the deep learning model to learn more effectively. Conventional normalization methods can be used for the data normalization step.

[0048] The data augmentation step (S140) is a process for improving the learning performance of deep learning models by diversifying and increasing the volume of MRI image data. Deep learning models require a large amount of diverse data. In particular, medical image data is difficult and expensive to collect, making augmentation technology essential for transforming existing data to generate new data.

[0049] In the present invention, the diversity and quantity of data can be increased by using rotation, flipping, scaling, etc.

[0050] Rotation techniques rotate images at various angles, allowing deep learning models to learn from images from various viewpoints. For example, a new image can be created by rotating the original image at various angles, such as 10, 20, or 30 degrees.

[0051] Flipping techniques flip images horizontally or vertically, allowing deep learning models to learn symmetrical features. For example, a new image can be created by flipping an original image horizontally or vertically.

[0052] Scaling techniques allow deep learning models to learn from images at various locations by shifting the image left and right or up and down. For example, they generate images shifted 10 or 20 pixels from the original image in various directions.

[0053] The original MRI image can be randomly subjected to techniques such as rotation, flipping, and scaling. For each original image, multiple augmentation techniques are combined to create a new image. For example, ten augmented images can be generated for each original image, increasing the data volume tenfold. The generated augmented images are verified to ensure they retain the characteristics of the original image. Additional augmentation techniques can be applied or some images can be removed as needed.

[0054] This allows deep learning models to achieve higher accuracy on new data by learning from diverse augmented data. Furthermore, increasing data diversity prevents models from overfitting to specific data.

[0055] The feature extraction step (S200) extracts structural features of the meniscus from preprocessed MRI image data, enabling a deep learning model to effectively diagnose meniscus lesions. This step utilizes various techniques to accurately extract structural features of the meniscus, thereby improving the model's learning and prediction performance.

[0056] The feature extraction step (S200) includes a preprocessed data preparation step (S210), a region of interest (ROI) setting step (S220), a model selection step for feature extraction (S230), a feature extraction step (S240), a feature data integration step (S250), and a final feature data utilization step (S260).

[0057] The preprocessed data preparation step (S210) collects data organized through the preprocessing step (S100). The data consists of MRI sequences such as T1-weighted images, T2-weighted images, PD images, and sagittal images, as well as label information (presence or absence of meniscus lesions) for each image.

[0058] In the preprocessed data preparation step (S210), the collected data is converted into a format suitable for deep learning model training. First, the MRI image data is converted into a NumPy array or tensor format, preparing it for processing by the deep learning model. Next, unnecessary or erroneous data is removed, and the data is organized to include only valid data.

[0059] The cleaned dataset is divided into training, validation, and test sets. The training set is used for model training. The validation set is used to evaluate model performance and tune hyperparameters. The test set is used to evaluate the final model performance.

[0060] Next, data diversity is increased by including various forms of augmented data generated in the data augmentation step (S140). The augmented data includes various transformations of the original data using methods such as rotation, flipping, and scaling. This improves the model's generalization performance.

[0061] Verify the quality of the prepared data and perform additional cleaning if necessary. Review whether the data is suitable for model training. If no abnormalities are found, proceed to the next step: setting the Region of Interest (ROI) (S220).

[0062] While MRI images contain a wealth of information, the information needed to diagnose meniscus lesions is concentrated in specific areas. Therefore, the region of interest setting step (S220) sets a region of interest that includes only the critical areas for efficient and accurate analysis.

[0063] The initial region of interest (ROI) can be determined collaboratively by an orthopedic surgeon and a radiologist. Doctors identify areas with a high risk of meniscus lesions on the sagittal view of an MRI scan and designate these areas as ROIs. Based on their experience and knowledge, doctors can select specific areas, such as the posterior portion of the medial meniscus (meniscocapsular junction), that are likely to develop lesions.

[0064] At this time, labeling software such as LabelMe can be used to precisely define the region of interest. Areas likely to contain meniscus lesions can be easily marked and labeled on each MRI image. During this process, physicians can select the most appropriate region of interest based on factors such as the location, size, and shape of the lesion.

[0065] After this initial region of interest setting, the region of interest setting step (S220) uses a deep learning model to automatically set the region of interest. The deep learning model, trained based on the initial training data, can automatically set the region of interest in new MRI images. This enables consistent and rapid region of interest setting for large amounts of data.

[0066] Next, the established ROI is verified to accurately encompass the meniscus lesion. This verification is performed by comparing it with surgical findings to ensure high accuracy. If necessary, the ROI setting algorithm is modified and retrained to improve accuracy. The optimized ROI setting algorithm ensures consistent ROI setting across all MRI images. This algorithm ensures generalized performance across diverse patient data.

[0067] The established region of interest information is stored along with the data and used for subsequent feature extraction and model training. Datasets containing region of interest information play a crucial role in helping deep learning models learn to diagnose meniscus lesions.

[0068] The model selection step for feature extraction (S230) selects the optimal deep learning model to accurately extract the structural features of meniscus lesions. This step selects an initial set of model candidates.

[0069] Convolutional Neural Networks (CNNs) can be applied as deep learning models. CNNs are models that demonstrate excellent performance in image recognition and are suitable for extracting structural features of meniscus lesions. Various CNN-based architectures can be selected as initial model candidates.

[0070] For example, ResNet (Residual Network) is a model designed to learn without performance degradation even in deep network structures, effectively learning the complex patterns of meniscus lesions. U-Net is primarily used in medical image analysis and excels at extracting detailed features from high-resolution input images. EfficientNet optimizes model size and computational complexity while maintaining high performance, making it effective for extracting high-level features from MRI images.

[0071] Additionally, you can develop an initial model by selecting an architecture such as 3D Recurrent Residual U-Net or PSPNet, and conduct a pilot study to compare and learn various models.

[0072] In the model selection step (S230), the structure of each candidate model is analyzed and their performance is compared to select the most suitable model for feature extraction of meniscus lesions. Accuracy, sensitivity, and specificity are evaluated. Sensitivity and specificity are particularly important in medical image analysis.

[0073] Preprocessed MRI image data and labeled data are used for model training. Augmentation techniques such as rotation, flipping, and scaling are applied to ensure data diversity. This enables the model to recognize meniscus lesions in a variety of situations.

[0074] Next, the selected model is trained and its performance is evaluated using a validation dataset. Cross-validation can be performed to improve the model's generalization performance. This prevents overfitting and enhances model stability. Hyperparameters such as learning rate and batch size can be adjusted to optimize model performance. In particular, the AdamW optimizer can be used to prevent model overfitting and improve generalization performance. The AdamW optimizer is effective in reducing overfitting during the training process and maintaining stable model performance.

[0075] Based on the learning and validation results, a final model is selected. This model must be able to most effectively extract the structural features of meniscus lesions. The clinical validity of the selected model is verified through a final evaluation. The predictive performance of the deep learning model is verified by comparing it with actual surgical data.

[0076] Meanwhile, deep learning models can be improved as needed, and additional data can be collected for retraining. For example, an attention mechanism can be introduced to allow the deep learning model to focus more on important features. This can enable more accurate recognition of the complex morphology of meniscus lesions. Alternatively, ensemble learning can be used to further improve performance by combining multiple deep learning models. This allows for increased accuracy by combining the strengths of various deep learning models.

[0077] The feature extraction step (S240) automatically extracts structural features of meniscus lesions from preprocessed MRI image data. Various low-level and high-level features are extracted from MRI sequences to generate input data necessary for deep learning model training.

[0078] First, preprocessed MRI image data and data designated as regions of interest are input into the deep learning model. The input data includes MRI sequences such as T1-weighted images, T2-weighted images, PD images, and sagittal images. For each sequence, the model is provided with a 3D image, allowing for a variety of information to be utilized.

[0079] Next, a Convolutional Neural Network (CNN)-based model is used to extract initial features from the input MRI image. Convolutional layers can be used to extract low-level features (edges, corners, etc.) from the image. Furthermore, pooling layers can reduce the spatial dimensionality of the extracted features, retaining only important features and reducing model complexity.

[0080] At this time, the EfficientNetB0 architecture can be used to extract high-level abstract features. EfficientNetB0 optimizes the balance between model size and performance, effectively learning diverse patterns in MRI images. EfficientNetB0 efficiently extracts high-level abstract features to maximize model performance.

[0081] Next, structures like ResNet (Residual Block) can be used to increase the depth of deep learning models and learn more complex patterns. Furthermore, atrous convolutions can be used to extract features with wider receptive fields, enabling better recognition of the complex structures of meniscus lesions. Models like DeepLabv3+ can extract features at multiple scales using atrous convolutions and atrous spatial pyramid pooling (ASPP). ASPP extracts features from various image resolutions, enabling accurate recognition of even the complex morphology of meniscus lesions.

[0082] Additionally, by introducing an attention mechanism, the model can focus more on important areas, thereby extracting the features of meniscus lesions more accurately.

[0083] After high-level feature extraction, a feature map is generated to represent the location and shape of the meniscus lesion. A Feature Pyramid Network (FPN) can be used for this purpose. FPN generates feature maps of various resolutions, enabling the recognition of ramp lesions of various sizes and shapes.

[0084] Next, ROI pooling can be performed using the feature map extracted from the region of interest. This allows for more effective analysis of features within the region of interest. ROI pooling converts the feature map within the region of interest into a fixed size, providing consistent input for subsequent steps.

[0085] Next, the ROI-pooled data is input to a fully connected layer to generate the final feature vector. The fully connected layer converts this data into a high-dimensional feature vector, which can be used as input to predict the presence or absence of meniscus lesions. The Dropout technique can be used to prevent overfitting and improve the model's generalization performance.

[0086] Feature vectors extracted from various sequences and regions of interest can be integrated to generate a final vector representing the comprehensive characteristics of meniscus lesions. Multimodal fusion can be used to integrate features from T1-weighted images, T2-weighted images, PD images, and sagittal images, enabling more accurate predictions.

[0087] Additionally, multivariate task learning can enhance the model's versatility by simultaneously performing other related clinical tasks (e.g., tissue classification) in addition to the detection and segmentation of meniscus lesions. This allows the model to learn and perform multiple clinical tasks simultaneously, enabling more efficient and comprehensive diagnosis.

[0088] The feature data integration step (S250) integrates feature data extracted from various sources to build a comprehensive predictive model. The goal of this step is to improve the predictive performance of the deep learning model by combining MRI image data with clinical data.

[0089] MRI image data can include feature maps extracted from MRI sequences such as T1-weighted images, T2-weighted images, PD images, and sagittal images. Clinical data can include gender, age, injury mechanism, injury time, and radiological risk factors.

[0090] First, MRI image data and clinical data are each preprocessed and converted into a consistent format. MRI image data undergoes normalization and standardization. Clinical data undergoes normalization and one-hot encoding of categorical data to convert it into numerical data.

[0091] Next, features extracted from each data source are converted into vectors. For MRI image data, feature maps extracted through a CNN are vectorized through ROI pooling and a fully connected layer. For clinical data, normalized clinical data are prepared in vector form.

[0092] Next, feature vectors extracted from multiple sources are integrated to create a single, comprehensive feature vector. This can be achieved using concatenation, multimodal fusion, and / or an attention mechanism. Concatenation simply concatenates feature vectors from MRI image data and clinical data into a single vector. Multimodal fusion uses a multilayer perceptron (MLP) or a complex neural network structure to combine information from various data sources. Additionally, an attention mechanism can be applied to weight each feature during the data integration process to reflect its importance.

[0093] Next, the integrated feature vector is normalized to transform it into a form suitable for use as input for a deep learning model. This can be accomplished by normalization, which adjusts the values ​​of the integrated vector to a certain range, or standardization, which transforms the vector into a normal distribution using its mean and standard deviation.

[0094] Next, the integrated feature vector is analyzed to identify important features. Feature selection and dimensionality reduction can be used for this. Feature selection improves the learning efficiency of deep learning models by identifying important features. Dimensionality reduction uses dimensionality reduction techniques, such as principal component analysis (PCA), to reduce the dimensionality of feature vectors and improve computational efficiency.

[0095] Next, the integrated feature vector is used to train a prediction model for meniscus lesions. Using the integrated data, a deep learning model is trained to predict the presence of meniscus lesions. Furthermore, validation data can be used to evaluate the performance of the deep learning model and adjust hyperparameters to achieve optimal performance.

[0096] The final feature data utilization step (S260) predicts the presence or absence of a meniscus lesion based on the integrated feature vector and derives the result. Softmax or sigmoid activation and / or a classification layer may be used. Softmax or sigmoid activation converts the prediction result into a probability, indicating the likelihood of a meniscus lesion. The classification layer ultimately derives a classification result, classifying whether or not a meniscus lesion is present.

[0097] Prediction results must be interpreted and visualized to facilitate physician understanding. Results are presented as probability values, which can be used to indicate the presence or absence of a lesion along with a degree of confidence. Furthermore, visualization tools can be used to highlight the image data surrounding the predicted lesion.

[0098] Predictive results are provided to medical professionals to support clinical decision-making. These can be useful for preoperative planning, treatment strategy development, and rehabilitation planning. By integrating clinical risk factors (gender, age, injury mechanism, timing, etc.) with predictive results, comprehensive diagnostic information can be provided.

[0099] Additionally, this function can be supported by a remote diagnostic system that analyzes MRI images and diagnoses meniscus lesions on a cloud-based platform. The system integrates with the cloud-based platform to analyze MRI images in real time and provides a user interface for easy use by medical professionals. By providing remote diagnostic results in real time to medical staff and patients, clinical decision-making can be more effectively supported. The user interface can include functions for interpreting diagnostic results, customizing settings, and storing and sharing diagnostic records. This can significantly improve the efficiency and accuracy of the diagnostic process.

[0100] A detailed description of the ramp lesion and the clinical results of the ramp lesion detection system according to the present invention are as follows.

[0101] A ramp lesion is a longitudinal tear that occurs at the junction of the meniscus and the joint capsule. It is often associated with anterior cruciate ligament (ACL) injury, compromising the biomechanical stability of the knee.

[0102] Meniscus lesions increase anterior displacement and rotational instability of the tibia, significantly affecting the knee joint. If left untreated, joint function can worsen. If meniscus lesions are not treated during anterior cruciate ligament reconstruction, they can negatively impact future recovery.

[0103] MRI is the preferred tool for diagnosing meniscus lesions, but its sensitivity ranges widely, from 53.9% to 84.6%, and results can vary significantly depending on the interpreter. This is why clinicians consider clinical risk factors when interpreting MRI.

[0104] The diagnostic accuracy of meniscus lesions relies heavily on the experience of the MRI reader (physician). This dependence leads to inconsistent diagnostic results. The present invention improves diagnostic accuracy by considering both clinical risk factors (gender, age, injury mechanism, etc.) and radiological indicators.

[0105] With the recent introduction of artificial intelligence (AI) technology into the medical field, AI is demonstrating performance comparable to that of medical experts, particularly in image recognition-related diagnostics. AI can provide faster and more consistent performance than existing methods in disease diagnosis, classification, and prediction. Deep learning models, in particular, can train on large datasets and analyze subtle details that may be missed by human readers, resulting in high accuracy. These AI characteristics are invaluable in MRI image interpretation.

[0106] To address the issue of varying MRI results depending on the reader, AI-based diagnostic methods can improve accuracy. AI enables consistent diagnoses through deep learning based on MRI images. In this case, the present invention further enhances diagnostic accuracy by incorporating clinical risk factors into the system, providing additional information. This plays a crucial role in comprehensively evaluating diagnostic results.

[0107] In order to determine the accuracy of the meniscus lesion detection system according to the present invention, clinical tests were conducted as follows.

[0108] The objectives of this clinical trial were, first, whether MRI image processing using deep learning technology improves the accuracy of diagnosing ramp lesions, and second, whether combining clinical risk factors with images using deep learning technology can further improve diagnostic accuracy.

[0109] In other words, this clinical trial aims to verify whether AI technology can provide consistency in image analysis in the diagnosis of ramp lesions and how much more accurate the diagnosis can be when clinical risk factors are combined with a deep learning model.

[0110] The overall process of this clinical trial involves training MRI images using deep learning, then incorporating clinical risk factors to build a final model. This process compares the performance of an AI model trained solely on MRI images with a model that incorporates clinical risk factors, and evaluates how each factor contributes to diagnostic accuracy.

[0111] This clinical trial was conducted based on a list of patients who underwent arthroscopic surgery between 2005 and 2023. Patients with anterior cruciate ligament (ACL) injury and a posterior medial meniscus tear in their surgical records were selected. Patients with poor-quality MRIs or no MRIs, which were deemed unsuitable for the study, were excluded. Furthermore, patients with a normal posterior medial meniscus or other types of tears were excluded, resulting in a total of 222 cases.

[0112] They were divided into matched and non-matched groups based on preoperative MRI and surgical findings. Deep learning was performed on the matched group, and transfer learning was applied to the non-matched group to create the final model.

[0113] After selecting images from MRI images that could be used to evaluate MC junctions, a region of interest (ROI) was set and deep learning training was performed. The results of the trained model were compared with the interpretations of a radiologist at our hospital.

[0114] The AI ​​model analyzed MRI images using a conventional neural network, and its performance was evaluated by comparing the learning results with clinical judgment. Setting the ROI is a critical step in accurately learning the region containing the lesion.

[0115] Clinical risk factors included age, sex, type of surgery, date of injury, and time elapsed between surgeries. Posteromedial tibial osteoedema (BME) and lateral meniscus tear (LM tear) were identified as significant risk factors closely associated with ramp lesions.

[0116] The final model was trained using an Efficient-net-based CNN, and its performance was evaluated across 222 patients, divided into training (8), validation (1), and test (1) sets. The CNN-based deep learning model analyzes image data patch by patch and combines it with clinical risk factors to produce the final output. The data set was split 8:1:1 to ensure the model's generalization performance.

[0117] Clinical judgment accuracy, % Deep learning learning accuracy, % Accuracy (Accuracy) 77.787.7 Sensitivity (Sensitivity) 71.787.9 Specificity (Specificity) 81.494.5

[0118] In Table 1 above, accuracy refers to the proportion of correctly diagnosed cases among all tests. Sensitivity refers to the proportion of people with the disease who are correctly diagnosed. Specificity refers to the proportion of people without the disease who are correctly diagnosed as not having the disease.

[0119] A model utilizing deep learning technology outperformed conventional readers in detecting ramp lesions. While clinicians' accuracy was 77.7% based solely on images, applying deep learning improved the accuracy to 87.7%.

[0120] To analyze risk factors, the inventors divided all patients into patients who underwent primary ACL reconstruction (primary ACLR) and revision ACLR. They then divided the patients into groups with and without ramp lesions and conducted an analysis. The results showed that the group with ramp lesions had a higher incidence of bone marrow edema on the posteromedial tibia (BME) and a higher incidence of lateral meniscus tears (LM tears). This demonstrates a significant correlation between ramp lesions and posteromedial tibial bone edema and lateral meniscus tears, suggesting that these risk factors may play an important role in the diagnosis of ramp lesions.

[0121] The variables used in this study are age, sex, lateralization, chronicity, and type of surgery. Age represents the patient's age, and sex is categorized as male or female. Laterality refers to whether the injury was on the right or left side, and the duration from injury to surgery is categorized as less than or more than 3 months. Surgery type is categorized as primary ACLR, which refers to a first-time ACL reconstruction, and revision ACLR, which refers to a re-operation after a re-rupture. Finally, LM tear indicates the presence of a lateral meniscus tear. These variables are important factors in assessing risk factors associated with ramp lesions and can contribute to the development of a more accurate diagnostic model.

[0122] The inventor analyzed risk factors using three regression analysis methods and determined that logistic regression analysis was the most appropriate method in terms of area under the curve (AUC) and accuracy. This means that among various analysis techniques, logistic regression analysis provides the most reliable results for assessing risk factors and diagnosing ramp lesions. AUC is an indicator that evaluates the predictive performance of a model; a higher value indicates better model performance, while accuracy indicates how accurately the model predicted. Therefore, the inventor determined that logistic regression analysis performed the best in these evaluation indicators.

[0123] The inventors performed univariate and multivariate analyses of each variable using the previously selected logistic regression analysis. First, univariate analyses were performed individually for all variables, and only those with significant correlations were selected for multivariate analysis. Through this process, only bone marrow edema on the posteromedial tibia (BME) and lateral meniscus tear (LM tear) were evaluated as clinically significant risk factors among various variables. In other words, these two variables play a significant role in the diagnosis of ramp lesions, and compared to other variables, posteromedial tibial bone edema and lateral meniscus tear were confirmed to be the most reliable risk factors.

[0124] Accuracy in clinical judgment, % Accuracy in deep learning learning, % Model according to the present invention (image + risk factors) Accuracy (Accuracy) 77.787.793.0 Sensitivity (Sensitivit) y 71.787.991.4 Specificity (Specificity) 81.494.596.3

[0125] As shown in Table 2 above, when clinical risk factors were additionally combined with the image-based trained model, as in the model of the present invention, the diagnostic accuracy increased from 87.7% to 93%. This means that the combined model of the present invention showed the highest performance. In other words, the deep learning technology that combines clinical risk factors with image data significantly improved the diagnostic accuracy compared to simply analyzing the image. In conclusion, the deep learning technology of the present invention was able to increase the diagnostic accuracy by utilizing clinical risk factors in addition to image data. This suggests that deep learning technology can be a very useful tool for diagnosing ramp lesions, as it showed a higher accuracy than the clinician's direct judgment.

[0126] The scope of protection in this field is not limited to the description and expression of the embodiments explicitly described above. Furthermore, it should be noted that obvious modifications or substitutions within the technical field to which the present invention pertains may not limit the scope of protection of the present invention.

Claims

1. A system for detecting meniscus lesions associated with anterior cruciate ligament injury using deep learning technology to automatically detect and diagnose posterior rupture of the medial meniscus in MRI images. The system comprises a deep learning model based on a multilayer neural network, wherein the system preprocesses image data through a preprocessing step for improving the resolution and contrast of MRI images, and extracts structural features of a meniscus from the preprocessed image data through the deep learning model.

2. In paragraph 1, The system is characterized in that it improves the diagnostic accuracy of meniscus lesions by integrating risk factors including gender, age, injury mechanism, and injury timing, and improves the comprehensive accuracy of diagnosis by fusing the risk factors and image data through multivariate analysis.

3. In paragraph 1, A system characterized in that the system labels an area including a meniscus lesion in the MRI image, increases the diversity and quantity of input data using data augmentation techniques including rotation, flipping, and scaling, and verifies the quality of the labeled data to use it for training the deep learning model.

4. In paragraph 1, The system is characterized in that the system is integrated with a cloud-based platform to remotely analyze the MRI images in real time and provide a user interface that is easy for doctors to use, and the user interface includes functions for interpretation of diagnostic results, customization, and storage and sharing of diagnostic records.

5. A method for operating a system for detecting meniscus lesions associated with anterior cruciate ligament injury using deep learning technology to automatically detect and diagnose posterior rupture of the medial meniscus in MRI images. a) a preprocessing step of preprocessing image data through a preprocessing step to improve the resolution and contrast of the MRI image; and b) A method characterized by including a feature extraction step for extracting structural features of the meniscus from the preprocessed image data.

6. In paragraph 5, Step a) above, a-1) Data collection step for collecting MRI images of patients with possible meniscus lesions among patients with anterior cruciate ligament injury; a-2) A data labeling step for analyzing the MRI image and labeling an area containing a meniscus lesion; a-3) A data normalization step for consistently processing the MRI image data to optimize the learning performance of the deep learning model; and a-4) A method characterized by including a data augmentation step for improving the learning performance of a deep learning model by diversifying and increasing the amount of MRI image data.

7. In paragraph 6, The above step a-4) is characterized in that the method increases the diversity and quantity of input data using a data augmentation technique including rotation, flipping, and scaling, and verifies the quality of the labeled data to use it for training the deep learning model.

8. In paragraph 5, Step b) above, b-1) A preprocessed data preparation step of collecting preprocessed MRI image data from step a) above and converting it into a format suitable for deep learning model training; b-2) A region of interest setting step for setting a region with a high probability of including a meniscus lesion in the MRI image as a region of interest; b-3) A model selection step for selecting the optimal deep learning model to accurately extract the structural features of meniscus lesions; and b-4) A method characterized by including a feature extraction step for automatically extracting structural features of a meniscus lesion from the preprocessed MRI image data.

9. In paragraph 8, Step b) above, b-5) further includes a feature data integration step that integrates feature vectors extracted from multiple sources to create a single integrated feature vector; The above step b-5) is a method characterized in that it improves the diagnostic accuracy of meniscus lesions by integrating risk factors including gender, age, injury mechanism, and injury time, and improves the comprehensive accuracy of diagnosis by fusing the risk factors and image data through multivariate analysis.

10. In paragraph 9, Step b) above, b-6) It further includes a final feature data utilization step for predicting the presence or absence of a meniscus lesion based on the above integrated feature vector and deriving the result. The above step b-6) is a method characterized in that it is integrated with a cloud-based platform to remotely analyze the MRI image in real time and provide a user interface that is easily usable by a doctor.

Images