Deep learning based rib fracture ct image intelligent detection system

Abstract

The application discloses a rib fracture CT image intelligent detection system based on deep learning, which comprises a rib CT image acquisition module, a tissue adaptive progressive enhancement module, a density threshold dynamic adjustment module, a CT perception loss constraint module, a rib CT image detection model design module and a rib fracture CT image detection module. The application belongs to the field of image processing and specifically refers to a rib fracture CT image intelligent detection system based on deep learning. The tissue adaptive progressive enhancement mechanism is designed to gradually strengthen the fracture line features and avoid overexposure caused by single strong enhancement. The density threshold dynamic adjustment mechanism is designed to further widen the density gap between subtle fractures and bone marrow. The bone tissue feature enhancement activation function is designed to avoid insufficient response to rib CT features. The skewness feature balanced loss design is based on to solve the skewness distribution of rib fracture data. And then, the rib fracture CT image detection effect is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of image processing, in particular to a rib fracture CT image intelligent detection system based on deep learning. BACKGROUND

[0002] The rib fracture CT image intelligent detection system is based on computer vision and deep learning technology, automatically analyzes the gray scale features, anatomical structure and density distribution of rib CT images, and realizes automatic identification, positioning and grading analysis of rib fractures (including subtle fractures). However, the general rib fracture CT image intelligent detection system has the problems that the gray scale difference between the rib and the soft tissue is not distinguished, the subtle fracture and the bone marrow are easily confused, and the detection result reliability is poor; the general rib fracture CT image intelligent detection system has the problems that focusing on the local part loses the integrity of the rib anatomical structure, the feature distribution of the rib fracture CT image is unbalanced under the skew data, and the detection effect is poor. SUMMARY

[0003] In view of the above problems, in order to overcome the defects of the prior art, the rib fracture CT image intelligent detection system based on deep learning is provided, which is designed to solve the problems that the general rib fracture CT image intelligent detection system has the problems that the gray scale difference between the rib and the soft tissue is not distinguished, the subtle fracture and the bone marrow are easily confused, and the detection result reliability is poor. The organization adaptive progressive enhancement mechanism is designed, the differential enhancement logic is designed, the multiple progressive enhancement is carried out, the fracture line features are gradually strengthened, and the overexposure or noise amplification caused by single strong enhancement is avoided; the density threshold dynamic adjustment mechanism is designed, the density gap between the subtle fracture and the bone marrow is further widened; and the final rib fracture CT image detection reliability is improved. In view of the problems that the general rib fracture CT image intelligent detection system has the problems that focusing on the local part loses the integrity of the rib anatomical structure, the feature distribution of the rib fracture CT image is unbalanced under the skew data, and the detection effect is poor. The bone tissue feature enhancement activation function is designed to accurately strengthen the bone fracture features and suppress invalid interference, and the response of the rib CT features is avoided. Based on the skew feature balance loss design, the skew distribution of the rib fracture data is solved through the feature anchoring, targeted negative sample and dynamic parameter triple mechanism, the bone fracture feature recognition degree is improved, the missed detection rate is reduced, and the pseudo-feature interference caused by the bone texture noise is suppressed. And the final rib fracture CT image detection effect is improved.

[0004] The technical scheme adopted by the application is as follows: the rib fracture CT image intelligent detection system based on deep learning provided by the application comprises a rib CT image acquisition module, an organization adaptive progressive enhancement module, a density threshold dynamic adjustment module, a CT perception loss constraint module, a rib CT image detection model design module and a rib fracture CT image detection module.

[0005] The rib CT image acquisition module acquires historical rib fracture CT images, and constructs a preliminary rib CT image set;

[0006] The tissue adaptive progressive enhancement module takes the mean gray value of the normal rib region as the reference, and performs progressive enhancement through a basic enhancement function and a pixel-level enhancement matrix;

[0007] The density threshold dynamic adjustment module optimizes the enhanced rib CT image through a dynamic weight enhancement function;

[0008] The CT-aware loss constraint module constructs a weighted total loss function of bone structure continuity, bone density uniformity, and tissue boundary sharpness, and combines tissue adaptive progressive enhancement and density threshold dynamic adjustment to realize preliminary rib CT image set enhancement and obtain a final rib CT image set;

[0009] The rib CT image detection model design module designs a skewness feature balance loss and a pseudo-contour frame optimization mechanism based on the final rib CT image set, and establishes a rib CT image detection model;

[0010] The rib fracture CT image detection module detects the real-time acquired rib CT image.

[0011] Further, the rib CT image acquisition module acquires historical rib fracture CT images; and performs preprocessing and labeling; to obtain a preliminary rib CT image set.

[0012] Further, the tissue adaptive progressive enhancement module specifically includes:

[0013] A tissue reference forward reference point is established, and the mean gray value of all normal rib regions in the preliminary rib CT image set is calculated as the reference forward reference point of the enhancement intensity;

[0014] A basic enhancement function is constructed, and the initial gray calibration is performed on each image in the preliminary rib CT image set.

[0015] Progressive enhancement is introduced, and a pixel-level enhancement matrix with the same dimension as the CT image is introduced to perform step-by-step progressive enhancement on the preliminary rib CT image set.

[0016] Further, the density threshold dynamic adjustment module specifically includes:

[0017] Customized density threshold extraction; for each image in the enhanced preliminary rib CT image set, the pre-trained ResNet-18 network is used to extract the customized rib-cancellous bone area threshold of the image;

[0018] Local density mean value calculation, each enhanced image is traversed through a 3×3 sliding window, and the local density mean value in each window is calculated;

[0019] The dynamic weight enhancement function design dynamically adjusts the enhancement weight according to the relationship between the local density mean value and the customized threshold value, and finally optimizes the image set.

[0020] Further, the CT-aware loss constraint module specifically includes:

[0021] Bone structure continuity loss design;

[0022] Bone density uniformity loss design;

[0023] Tissue boundary sharpness loss design;

[0024] Get the total loss function;

[0025] Perform rib CT image enhancement to obtain the final rib CT image set.

[0026] Further, the rib CT image detection model design module specifically includes:

[0027] Bone tissue feature enhanced backbone network construction; based on ResNet as the basic architecture, a bone tissue feature enhanced backbone network is constructed, and a bone tissue feature enhanced activation function is designed;

[0028] Design a bone tissue feature enhanced activation function: including a dynamic enhancement coefficient, a soft threshold filter and a scale-aware factor; calculate the enhanced residual block feature;

[0029] Skewness feature balance loss design; construct a feature balance mechanism with double storage feature pool + dynamic parameter adjustment, calculate the class-level feature center as the forward reference point; targeted negative sample screening; and dynamic parameter adaptation: adjust the coefficient inversely with the class sample proportion;

[0030] Construct a critical value screening pseudo-contour frame optimization mechanism, based on the initial pseudo-contour frame output by the unit, set a double-dimensional screening standard, and finally construct a bone fracture area back-propagation loss;

[0031] Collaborative training optimization; establish a basic classification loss; finally get the total loss function.

[0032] Further, the rib fracture CT image detection module is to collect rib fracture CT images in real time, after preprocessing and rib CT image enhancement, input into the rib CT image detection model, and realize rib fracture CT image detection based on the model output result.

[0033] The above scheme has the following beneficial effects:

[0034] (1) In view of the problem that the general rib fracture CT image intelligent detection system cannot distinguish the gray difference between the ribs and soft tissues, the subtle fractures and the cancellous bone are easily confused, and the reliability of the detection result is poor, the self-adaptive progressive enhancement mechanism is designed, the differential enhancement logic is designed, and the multi-round progressive enhancement is carried out, so that the fracture line features are gradually strengthened, and the overexposure or noise amplification caused by single strong enhancement is avoided; the density threshold dynamic adjustment mechanism is designed, and the density gap between the subtle fractures and the cancellous bone is further widened; and the reliability of the final rib fracture CT image detection is improved.

[0035] (2) In view of the problem that the general rib fracture CT image intelligent detection system loses the integrity of the rib anatomy structure by focusing on the local, the feature distribution of the rib fracture CT image is unbalanced under the skew data, and the detection effect is poor, the bone tissue feature enhancement activation function is designed, the bone fracture feature is accurately strengthened, the invalid interference is suppressed, and the response to the rib CT feature is avoided; based on the skew feature balance loss design, through the feature anchoring, the targeted negative sample and the dynamic parameter triple mechanism, the skew distribution of the rib fracture data is solved, the fracture feature recognition is improved, the missed detection rate is reduced, and the pseudo-feature interference caused by the bone texture noise is suppressed; and the detection effect of the final rib fracture CT image is improved. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The flowchart of the rib fracture CT image intelligent detection system based on deep learning provided by the present application is shown.

[0037] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0039] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore cannot be understood as a limitation on the present application.

[0040] Embodiment one, refer toFigure 1 The application provides a deep learning-based rib fracture CT image intelligent detection system, which comprises a rib CT image acquisition module, a tissue adaptive progressive enhancement module, a density threshold dynamic adjustment module, a CT perception loss constraint module, a rib CT image detection model design module and a rib fracture CT image detection module.

[0041] The rib CT image acquisition module acquires historical rib fracture CT images and constructs a preliminary rib CT image set; and sends data to the tissue adaptive progressive enhancement module.

[0042] The tissue adaptive progressive enhancement module takes the gray mean value of a normal rib region as a reference, performs progressive enhancement through a basic enhancement function and a pixel-level enhancement matrix, and sends data to the density threshold dynamic adjustment module.

[0043] The density threshold dynamic adjustment module optimizes the enhanced rib CT images through a dynamic weight enhancement function, and sends data to the CT perception loss constraint module.

[0044] The CT perception loss constraint module constructs a weighted total loss function of bone structure continuity, bone density uniformity and tissue boundary sharpness, realizes preliminary rib CT image set enhancement in combination with tissue adaptive progressive enhancement and density threshold dynamic adjustment, and obtains a final rib CT image set; and sends data to the rib CT image detection model design module.

[0045] The rib CT image detection model design module designs a skewness feature balance loss and a pseudo-contour frame optimization mechanism based on the final rib CT image set, establishes a rib CT image detection model, and sends data to the rib fracture CT image detection module.

[0046] The rib fracture CT image detection module detects real-time collected rib CT images.

[0047] Example two, refer to Figure 1 Based on the above-mentioned example, the rib CT image acquisition module acquires historical rib fracture CT images, performs continuous tomographic scanning on the chest rib region through a clinical standard CT scanning device, outputs a CT image sequence in DICOM format, the image resolution is not less than 512*512, and the layer thickness is controlled to be 1-2mm; and pre-processing and labeling are performed; the image pre-processing includes denoising (median filtering) and gray scale normalization (mapping the image gray value to the [0, 255] interval); the labeled class label includes normal ribs, slight fractures and severe fractures; and a preliminary rib CT image set is obtained.

[0048] Example three, refer to Figure 1, the embodiment is based on the above-mentioned embodiment, the problem of significant gray difference between ribs (high density) and soft tissue (low density) in the preliminary rib CT image set is solved by the adaptive progressive enhancement module, and the gray contrast between the fracture line and the surrounding tissue is gradually strengthened, while the overexposure of the bone edge and the covering of the fine fracture line by noise are avoided; specifically comprising:

[0049] Establishing a tissue reference positive reference point, the gray mean value of all normal rib regions in the preliminary rib CT image set is calculated as the reference positive reference point of the enhancement intensity, denoted as ;

[0050] Constructing a basic enhancement function, for each image in the preliminary rib CT image set, the initial gray calibration is performed to widen the gray gap between the ribs and the soft tissue, which is represented as: ; wherein, is the pixel coordinate; is the output result of the basic enhancement function; is the gray value; is the initial enhancement coefficient, taking the value of 0.1~0.3;

[0051] Progressive enhancement, gradually strengthening the features, introducing a pixel-level enhancement matrix with the same dimension as the CT image, step-by-step progressive enhancement is performed on the preliminary rib CT image set, gradually improving the saliency of the fracture line, which is represented as: ; wherein, and are the enhancement results of the tth progressive and (t-1)th progressive, respectively, and the initial progressive result is ; t is the number of progressions, with a maximum value of 3~5; is the pixel-level enhancement matrix, which is obtained by tissue density clustering analysis, and high enhancement weight is given to the rib region and low enhancement weight is given to the soft tissue region, realizing differential enhancement.

[0052] Embodiment four, refer to Figure 1 , the embodiment is based on the above-mentioned embodiment, the density threshold dynamic adjustment module is aimed at the problem that after the preliminary rib CT image set is subjected to the adaptive progressive enhancement, there may still be "fine fracture and cancellous bone density similar, easy to confuse" problem, the enhanced image set is further optimized, the gray feature saliency of fine fracture is targetedly improved, and the confusion between fine fracture and cancellous bone is avoided; the specific operation is:

[0053] Customized density threshold extraction; for each image in the enhanced preliminary rib CT image set, the rib-cancellous bone differentiation threshold S of the image is extracted by using the pre-trained ResNet-18 network, and the threshold is adapted to the overall density distribution of the image;

[0054] Local density mean calculation, traverse each enhanced image by 3x3 sliding window, calculate the local density mean v(x) in each window, reflect the regional tissue density level;

[0055] Dynamic weight enhancement function design, according to the relationship between local density mean and custom threshold, dynamically adjust the enhancement weight, optimize the final enhancement of image set, expressed as: ;

[0056] The progressive order of the preliminary rib CT image set: tissue adaptive progressive enhancement → density threshold dynamic adjustment → next progressive.

[0057] Example five, see Figure 1 , this example is based on the above example, CT perception loss constraint module is to the preliminary rib CT image set after tissue adaptive progressive enhancement and density threshold dynamic adjustment, construct bone structure-density uniformity-tissue boundary loss function, constrain the quality of enhanced image; Ensure that the enhanced image not only retains the continuity of the fracture line, the uniformity of the normal bone tissue density, but also strengthens the boundary of the fracture line, and does not appear semantic confusion between ribs and soft tissues; Specifically includes:

[0058] Bone structure continuity loss design, ensure that the fracture line is not interrupted, maintain the rib anatomy, expressed as: ; wherein, is the bone structure continuity loss value; M is the number of key anatomical points of the rib, obtained by image edge detection; is the neighborhood point set of the i-th anatomical point, the neighborhood size is 4~8; and are the gray values of the i-th anatomical point and the j-th anatomical point of the enhanced image respectively, j is the index of the neighborhood anatomical point; and are the gray values of the i-th anatomical point and the j-th anatomical point of the preliminary rib CT image respectively;

[0059] Bone density uniformity loss design, control the normal bone tissue gray value in the ideal range (avoid overexposure), expressed as: ; wherein, is the bone density uniformity loss value; P is the total number of pixels in the preliminary rib CT image marked as normal rib region; is the gray value of the i-th pixel in the normal rib region of the enhanced image; Q is the ideal gray value of normal bone tissue, based on the statistics of clinical CT images;

[0060] Tissue boundary definition loss design, strengthen the boundary of the fracture line and the surrounding tissue, expressed as: ; wherein, is the loss value of the organization boundary clarity; H and W are the image height and image width (pixel number) respectively; and are the horizontal direction gradient operator and the vertical direction gradient operator (using the Sobel operator) respectively; is the enhanced image gray value;

[0061] The total loss function is obtained , which is expressed as: ;

[0062] Rib CT image enhancement process:

[0063] Initialization of enhancement parameters: set the maximum number of iterations, the initial value of the initial enhancement coefficient, and obtain the initial value of the pixel-level enhancement matrix through tissue density clustering;

[0064] Enhanced image generation: enhance the preliminary rib CT image set with the initial parameters (one iteration of tissue adaptive progressive enhancement and dynamic adjustment of density threshold), to obtain the enhanced image;

[0065] Parameter optimization: calculate the total loss, and adjust the enhancement parameters through gradient descent method to minimize the total loss;

[0066] Progressive convergence: repeat the above process until the total loss converges (the fluctuation of the progressive loss value is less than 0.001); if the maximum number of iterations is reached, go to the initialization of the enhancement parameters to re-progressive; the converged enhancement parameters are the optimal enhancement parameters;

[0067] Final enhancement: perform a complete enhancement (t times) on the preliminary rib CT image set with the optimal enhancement parameters to obtain the final rib CT image set.

[0068] In view of the problem that the general rib fracture CT image intelligent detection system cannot distinguish the gray difference between ribs and soft tissues, and the subtle fracture and bone marrow are easily confused, and further lead to poor reliability of the detection result, the scheme designs a tissue adaptive progressive enhancement mechanism, designs a differentiated enhancement logic, and performs multi-round progressive enhancement, gradually strengthens the fracture line features, and avoids overexposure or noise amplification caused by single strong enhancement; a density threshold dynamic adjustment mechanism is designed to further narrow the density gap between subtle fractures and bone marrow; and further improve the reliability of the final rib fracture CT image detection.

[0069] Embodiment six, refer to Figure 1 , the rib CT image detection model design module is based on the final rib CT image set, uses ResNet50 as the backbone network, and designs a CT feature enhancement activation function to optimize the feature transmission efficiency of the residual block, so that the network can capture the subtle gray changes on the rib surface; specifically including:

[0070] Construction of a bone tissue feature-enhanced backbone network: Based on ResNet architecture, a bone tissue feature-enhanced backbone network is constructed. A bone tissue feature enhancement activation function is designed and replaced with the original activation function to strengthen effective feature responses, optimize the residual block structure for stable gradient propagation, and differentiate between shallow and deep layers. This allows the network to retain the overall rib morphology while focusing on fine-grained features of minor fractures. The residual block structure is optimized by performing batch normalization (BN) and activation functions sequentially after convolutional layers, followed by fusion with the shortcut branch to further stabilize gradients. The shallow layers retain multi-channel features to capture the overall rib morphology, while the deep layers focus on fine-grained features to locate minor fractures.

[0071] Design a bone tissue feature enhancement activation function: including dynamic enhancement coefficients, soft threshold filtering, and scale-aware factors, to accurately identify bone fracture features, suppress ineffective interference, and adapt to the division of labor between deep and shallow layers of the network, expressed as: ; ;in, It is a bone tissue feature enhancement activation function; yes Derivative; u is the input to the activation function, which is the original feature value of the final rib CT image output in the convolutional layer; dynamic enhancement coefficient. Scale perception factor ;c is the tissue type identifier corresponding to the feature channel (1=bone cortex / bone trabeculae, 2=soft tissue, 3=noise);s is the network level where the feature is located (1=shallow layer, 2=deep layer). It is the soft suppression coefficient, with a value ranging from 0.2 to 0.4; and These are the positive and negative feature thresholds, respectively, with values ​​ranging from 0.08 to 0.12 and -0.06 to -0.04; the enhanced residual block feature (optimized gradient propagation and feature fusion) is calculated as follows: ; ; ;in, and These are the intermediate feature maps after the first and second convolutional layers (followed by BN and activation), respectively; F is the final output feature map of the optimized residual block. , and It is the convolution kernel of the three convolutional layers in the residual block, with a size of 3×3; It is the input feature map of the residual block; It is a batch normalization operation;

[0072] Skew feature balance loss design; Construct a feature balance mechanism with double storage feature pool + dynamic parameter adjustment, including: Double storage feature pool construction: high confidence feature library stores high confidence rib feature of labeled / unlabeled image (high confidence is prediction probability not less than 0.7), category label library stores corresponding label (labeled data stores true value, unlabeled data stores network prediction value), calculate category level feature center as positive reference point; Feature pool is updated every 10 rounds; Targeted negative sample screening: select samples with confidence exceeding threshold (0.6~0.8) and non-fracture class from labeled data, select samples with fracture confidence not in the top 3 from unlabeled data, avoid bone texture noise interference; Dynamic parameter adaptation: adjust coefficient inversely with category sample proportion, negative sample weight adapts to batch size, stabilize loss calculation;

[0073] Category level positive reference point calculation (ensure positive reference point reliability) is expressed as: Skew feature balance loss (Optimize feature distribution) is expressed as: Dynamic adjustment coefficient is expressed as: Dynamic negative sample weight is expressed as: Wherein, is the category level positive reference point of the kth category; is the total number of samples of the kth category in the high confidence feature library; is the nth sample feature vector of the kth category in the high confidence feature library; is the feature set of the kth category in the high confidence feature library; B is the total number of samples in the current training batch; K is the total number of categories, taking 3; is the number of samples of the kth category in the current batch, and b is the sample index; is the feature vector of the bth sample in the current batch; is the dynamic adjustment coefficient; is the basic adjustment coefficient, taking value 0.05~0.3; is the smoothing term; is the negative sample feature vector of the kth category in the current batch; T is the total training iteration number; t is the previous training iteration number; is the total number of samples of the cth category; is the temperature adjustment coefficient, taking value 0.5~1.0; is the total number of negative samples of the kth category in the current batch; is the vector inner product;

[0074] The pseudo-bounding box optimization mechanism of critical value screening is constructed to guide the initial pseudo-bounding box output by the unit (the guiding unit is a pre-positioning unit in the rib CT image detection model responsible for generating the initial pseudo-bounding box, which is a core functional branch with preliminary identification capability of the fracture area), and a double-dimension screening standard is set: the pseudo-bounding box with foreground score higher than the critical value of 0.6-0.8 and the reverse deviation lower than the critical value of 3-10 is reserved, and the non-stable pseudo-boxes such as deviation boxes caused by interference signals are accurately screened out, so that the pseudo-bounding box participating in training has reliable positioning accuracy, and finally the fracture area reverse loss is constructed; the refined box calculation (ensuring full coverage of the fracture area) is integrated and represented as: ; fracture area reverse loss (optimize positioning accuracy) is represented as: ; wherein, is the final pseudo-bounding box of the i-th fracture area after critical value screening; is a double-dimension critical value screening function; is the initial pseudo-bounding box set of the i-th fracture area output by the guiding unit; is the number of effective foreground pseudo-bounding boxes reserved after double-dimension critical value screening; is a basic positioning loss function, and the mean absolute error is selected; is the i-th effective foreground pseudo-bounding box after screening, is a qualified box filtered by self-adaption; is a reverse target reference box, which is a true positioning reference box;

[0075] Cooperative training optimization; establish basic classification loss (compute category prediction error), represented as: ; total loss function is represented as: ; wherein, is the batch size of the labeled sample participating in the current iteration, and b is the current batch sample index; is a cross-entropy function; is a category prediction probability distribution; is the output feature of the feature extraction network for the b-th sample; is a true category label; is an indicator function, when t is greater than the threshold value, , otherwise , the threshold value is 20%-30% of the total training times;

[0076] Regarding model training: the final rib CT image set is divided into a training set, a validation set and a test set; the training set is used for model parameter iterative update, the validation set is used for monitoring indicators, adjusting hyperparameters and triggering early stopping during the training process, and the test set is used for final model performance evaluation; the Adam optimizer (initial learning rate 1e-4) is used to update all parameters based on the total loss back propagation; the learning rate is attenuated to 0.1 of the original value every 20000 steps; the core indicators (classification: severe fracture recall rate, three-class F1-score; positioning: fracture region IoU) are evaluated every 1000 steps on the validation set; if the severe fracture recall rate does not improve for 30 consecutive rounds and the total loss fluctuates by less than 5%, early stopping is triggered; finally, the test set is used to test the classification accuracy to evaluate the model performance.

[0077] By performing the above operations, for the general rib fracture CT image intelligent detection system, the problems of focusing on the local and losing the integrity of the rib anatomical structure, the imbalance of the feature distribution of the rib fracture CT image under the skewed data, and the poor detection effect are solved. The bone tissue feature enhancement activation function is designed to accurately strengthen the bone fracture feature and suppress invalid interference, avoiding insufficient response to the rib CT feature. Based on the skewed feature balance loss design, the skewed distribution of the rib fracture data is solved through the three mechanisms of feature anchoring, targeted negative samples and dynamic parameters, the bone fracture feature recognition is improved, the missed detection rate is reduced, and the pseudo-feature interference caused by bone texture noise is suppressed. The detection effect of the final rib fracture CT image is improved.

[0078] Embodiment seven, see Figure 1 This embodiment is based on the above-mentioned embodiments. The rib fracture CT image detection module is used to collect the rib fracture CT image in real time, and after preprocessing and rib CT image enhancement, the rib fracture CT image is input into the rib CT image detection model to realize rib fracture CT image detection based on the model output result. The rib CT image detection model extracts the deep and shallow features of the rib CT image, combines the classification and positioning branches optimized by cooperative training, and outputs the classification results of three types of rib states (normal, minor fracture, and severe fracture) and the accurate positioning frame of the corresponding fracture region.

[0079] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application.

[0080] The above describes the present application and its embodiments, which are not limited, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed, which shall belong to the protection scope of the present application.

Claims

1. A deep learning-based intelligent detection system for rib fracture CT images, characterized in that: The system includes a rib CT image acquisition module, a tissue adaptive progressive enhancement module, a density threshold dynamic adjustment module, a CT perception loss constraint module, a rib CT image detection model design module, and a rib fracture CT image detection module. The rib CT image acquisition module acquires historical rib fracture CT images and constructs a preliminary rib CT image set. The tissue adaptive progressive enhancement module uses the average gray level of the normal rib region as a benchmark and performs progressive enhancement through a basic enhancement function and a pixel-level enhancement matrix. The density threshold dynamic adjustment module optimizes and enhances rib CT images through a dynamic weight enhancement function; The CT perception loss constraint module constructs a weighted total loss function based on bone structure continuity, bone density uniformity, and tissue boundary clarity. Combined with tissue adaptive progressive enhancement and dynamic adjustment of density threshold, it achieves initial rib CT image set enhancement and obtains the final rib CT image set. The rib CT image detection model design module is based on the final rib CT image set, and designs a skewed feature balance loss and pseudo contour box optimization mechanism to establish a rib CT image detection model. The rib fracture CT image detection module detects rib CT images acquired in real time.

2. The intelligent rib fracture CT image detection system based on deep learning according to claim 1, characterized in that: The organization adaptive progressive enhancement module specifically includes: Establish a positive reference point for tissue baseline, and statistically analyze the gray-scale mean of all normal rib regions in the preliminary rib CT image set as a positive reference point for enhancement intensity; Construct a basic enhancement function to perform initial grayscale calibration on each image in the preliminary rib CT image set; Progressive enhancement.

3. The intelligent detection system for rib fracture CT images based on deep learning according to claim 2, characterized in that: The progressive enhancement involves introducing a pixel-level enhancement matrix of the same dimension as the CT images to progressively enhance the initial rib CT image set in stages.

4. The intelligent rib fracture CT image detection system based on deep learning according to claim 3, characterized in that: The density threshold dynamic adjustment module specifically includes: Customized density threshold extraction: For each image in the enhanced preliminary rib CT image set, a customized rib-cancellous bone distinction threshold is extracted using a pre-trained ResNet-18 network. Local density mean calculation: Each enhanced image is traversed through a 3×3 sliding window, and the local density mean within each window is calculated. Design of dynamic weight enhancement function.

5. The intelligent rib fracture CT image detection system based on deep learning according to claim 4, characterized in that: The dynamic weight enhancement function design dynamically adjusts the enhancement weights based on the relationship between the local density mean and a customized threshold, and performs final enhancement optimization on the image set.

6. The intelligent rib fracture CT image detection system based on deep learning according to claim 5, characterized in that: The CT sensing loss constraint module specifically includes: Design for loss of bone structure continuity; Design for loss of bone density uniformity; Design for organizational boundary clarity loss; The total loss function is obtained.

7. The intelligent rib fracture CT image detection system based on deep learning according to claim 6, characterized in that: The rib CT image detection model design module specifically includes: Construction of a bone tissue feature-enhanced backbone network: Based on ResNet architecture, a bone tissue feature-enhanced backbone network is constructed, and a bone tissue feature enhancement activation function is designed. Design an activation function to enhance bone tissue features, including dynamic enhancement coefficients, soft thresholding, and scale-aware factors; calculate enhanced residual block features; The design incorporates a skewed feature balance loss mechanism; it constructs a dual-feature pool and a dynamic parameter adjustment mechanism, calculates category-level feature centers as positive reference points, performs targeted negative sample screening, and performs dynamic parameter adaptation: the adjustment coefficients are adjusted inversely according to the proportion of category samples. Construct a pseudo-contour box optimization mechanism for critical value filtering; Collaborative training optimization; establishing a basic classification loss; finally obtaining the total loss function.

8. The intelligent rib fracture CT image detection system based on deep learning according to claim 7, characterized in that: The pseudo-contour box optimization mechanism for constructing critical value screening is based on the initial pseudo-contour box output by the guidance unit, sets a two-dimensional screening standard, and finally constructs the back-inference loss of the fracture area.

9. The intelligent rib fracture CT image detection system based on deep learning according to claim 8, characterized in that: The rib CT image acquisition module acquires historical rib fracture CT images; performs preprocessing and annotation; and obtains a preliminary rib CT image set.

10. The intelligent rib fracture CT image detection system based on deep learning according to claim 9, characterized in that: The rib fracture CT image detection module acquires rib fracture CT images in real time, preprocesses and enhances the rib CT images, and then inputs them into the rib CT image detection model. Based on the model's output results, rib fracture CT image detection is achieved.

Images