Chest X-ray image processing method for pneumoconiosis based on AI small shadow detection rate

Abstract

The invention relates to the technical field of pneumoconiosis image detection, and discloses a pneumoconiosis-oriented chest X-ray image processing method based on an AI small shadow detection rate. According to the method, a basic mask is generated through an initial segmentation network, and then an accurate lung contour is obtained through correction of a hierarchical optimization unit. The multi-scale feature extraction module analyzes gray features in the contour and constructs an initial small shadow probability distribution map. A morphological structure analyzer identifies discrete shadow regions and marks suspicious shadow clusters, and a dynamic confidence feedback mechanism evaluates its credibility to generate an optimized feature map. A shadow cluster topological incidence matrix is established by the spatial relation modeling network and matched with a pneumoconiosis pathological feature library to screen a target shadow area. The three-dimensional reconstruction engine generates a small shadow volume density thermodynamic diagram, the hierarchical fusion module integrates the thermodynamic diagram and original image space coordinates, an enhanced distribution map is output, and finally structured diagnosis report data is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of pneumoconiosis image detection, and in particular to a chest X-ray image processing method for improving the AI ​​small shadow detection rate of pneumoconiosis. Background Art

[0002] Chest X-rays are widely used in the diagnosis of pneumoconiosis, with the detection of small shadows being a crucial factor in determining the condition. Traditionally, the detection of small shadows relies primarily on manual interpretation, which can be influenced by subjective factors such as physician experience and fatigue. Different physicians may interpret the same image differently, resulting in low diagnostic consistency.

[0003] With the development of computer technology, several automatic detection methods based on image processing have emerged. However, existing methods often face the problem of inaccurate lung segmentation when processing chest X-ray images. The boundary between the lungs and surrounding tissue is blurred, resulting in inaccurate regions for subsequent small shadow analysis. In the feature extraction process, most methods use a single-scale analysis, which makes it difficult to fully capture the grayscale distribution characteristics of small shadows. This results in a high level of noise in the initial probability distribution map, which easily obscures the true shadow.

[0004] At the same time, existing technologies lack in-depth morphological analysis of small shadows, making it difficult to effectively identify discrete shadow regions and suspicious shadow clusters. Furthermore, they lack a dynamic assessment mechanism for the credibility of shadow clusters, making it easy to misidentify normal tissue as shadows or miss true lesions. Furthermore, there are deficiencies in modeling the spatial relationships between shadow regions, making it impossible to accurately establish topological associations between shadow clusters. This results in a low degree of match with the pathological feature library, affecting the accuracy of screening for target shadow regions. Finally, the generated diagnostic reports are often unstructured data, which is not conducive to subsequent statistics and analysis. Summary of the Invention

[0005] The object of the present invention is to provide a chest X-ray image processing method with an AI small shadow detection rate for pneumoconiosis, so as to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides a chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis, the method comprising: Acquire a chest X-ray original image to be processed, perform initial segmentation network processing on the original image, and generate a basic mask containing a lung area; A hierarchical optimization unit is used to refine and correct the boundaries of the basic mask to output an accurate lung contour; Analyzing the grayscale distribution characteristics within the precise lung contour through a multi-scale feature extraction module to construct an initial small shadow probability distribution map; Processing the initial small shadow probability distribution map based on a morphological structure analyzer, identifying discrete shadow areas and marking suspicious shadow clusters; Using a dynamic confidence feedback mechanism to evaluate the credibility of the morphological features of the suspicious shadow cluster, and generate an optimized small shadow feature map; Inputting the optimized small shadow feature map into a spatial relationship modeling network to establish a topological association matrix between shadow clusters; Matching the topological association matrix with a preset pneumoconiosis pathological feature library to screen target shadow areas that meet clinical pathological characteristics; Using a three-dimensional reconstruction engine to process the target shadow area to generate a small shadow volume density heat map; The small shadow volume density heat map and the spatial coordinate information of the original image are integrated through a hierarchical fusion module to output an enhanced small shadow distribution map; Structured diagnostic report data is generated based on the enhanced small shadow distribution map.

[0007] Preferably, the initial segmentation network processing specifically includes: constructing a dual-channel convolutional neural network as a segmentation backbone network, the segmentation backbone network includes a spatial path and a context path working in parallel; extracting high-resolution edge features of the original image through the spatial path to generate an edge feature tensor; using the context path to capture the global semantic features of the original image to generate a semantic feature tensor; using an adaptive feature fusion device to merge the edge feature tensor and the semantic feature tensor to generate a fused feature map; performing multi-level supervised training on the fused feature map, and updating the weight parameters of the segmentation backbone network through a back-propagation algorithm; and using the binary image output by the segmentation backbone network after training convergence as the basic mask.

[0008] Preferably, the operation of the hierarchical optimization unit includes: receiving the basic mask and executing the contour tracing algorithm to extract the lung boundary pixel chain; using the curvature analysis model to calculate the local curvature anomaly points of the lung boundary pixel chain; constructing a boundary correction vector field based on the local curvature anomaly points; optimizing the boundary correction vector field through an iterative energy minimization algorithm to drive the lung boundary pixel chain to move toward the anatomical standard position; and outputting accurate lung contour data containing smooth and continuous contours.

[0009] Preferably, the operating steps of the multi-scale feature extraction module are: establishing a multi-resolution pyramid structure within the precise lung contour; using a Gaussian difference filter group to process the lung area of ​​the original image to generate a multi-scale gradient amplitude map; extracting local texture features of the multi-scale gradient amplitude map to construct a rotation-invariant texture descriptor; calculating the similarity matrix between the rotation-invariant texture descriptor and a standard pneumoconiosis texture template; fusing the similarity matrix with the spatial weight of the multi-scale gradient amplitude map to generate initial small shadow probability distribution data.

[0010] Preferably, the morphological structure analyzer performs the following operations: performs adaptive threshold segmentation on the initial small shadow probability distribution map to obtain a binary candidate area; uses a connected domain labeling algorithm to process the binary candidate area to generate a discrete shadow area identifier; calculates the geometric morphological feature vector of each discrete shadow area, the geometric morphological feature vector including area irregularity and boundary fractal dimension; analyzes the spatial aggregation characteristics of the geometric morphological feature vector based on a density clustering algorithm, and marks suspicious shadow clusters that meet the aggregation constraints.

[0011] Preferably, the dynamic confidence feedback mechanism includes: constructing a shadow feature evaluation network, which includes a feature encoder and a confidence predictor; inputting the original image block of the suspicious shadow cluster into the feature encoder to generate a deep feature representation; calculating the morphological confidence score of the deep feature representation through the confidence predictor; using a non-maximum suppression algorithm to screen suspicious shadow clusters whose morphological confidence score is higher than a critical threshold; aligning the screening results with the original image coordinates to generate an optimized small shadow feature matrix.

[0012] Preferably, the operation of the spatial relationship modeling network includes: receiving the optimized small shadow feature matrix and extracting spatial position coordinates; Construct a graph neural network model with each shadow cluster as a node and the Euclidean distance between nodes as the edge weight; learn the spatial dependency between nodes through the graph convolution layer and update the node feature representation; calculate the topological correlation strength between the updated node feature representation and adjacent nodes; output the topological correlation matrix data containing the connection strength between nodes.

[0013] Preferably, the pathological feature library matching process includes: pre-constructing a standard pneumoconiosis shadow feature library, which stores the shadow distribution patterns of typical pneumoconiosis cases; extracting the key path feature vector in the topological association matrix; calculating the cosine similarity between the key path feature vector and the standard pneumoconiosis shadow feature library; screening the topological subgraph whose cosine similarity is greater than the matching threshold; mapping the topological subgraph back to the original image space, and outputting the target shadow area coordinate set.

[0014] Preferably, the operating steps of the three-dimensional reconstruction engine are: obtaining a sequence of X-ray images corresponding to the target shadow area in continuous layers; processing the X-ray image sequence using a voxel reconstruction algorithm to generate three-dimensional volume data; calculating the grayscale gradient characteristics of each voxel in the three-dimensional volume data; optimizing the signal-to-noise ratio of the three-dimensional volume data based on an anisotropic diffusion filter; and generating three-dimensional thermal map data that annotates the spatial distribution of small shadows.

[0015] Preferably, the hierarchical fusion module performs: receiving the spatial coordinate transformation parameters of the three-dimensional thermal map data and the original image; establishing a conversion function from small shadow density to color mapping; projecting the three-dimensional thermal map data to a two-dimensional diagnostic plane through an image registration algorithm; fusing the projection data with the anatomical structure features of the original image to generate an enhanced small shadow distribution map; and generating a structured report containing the number and distribution positions of shadows based on the density zoning statistics of the enhanced small shadow distribution map.

[0016] Compared with the prior art, the present invention has the following beneficial effects: Through the collaborative operation of multiple modules, precise detection and analysis of small pneumoconiosis shadows are achieved. After the initial segmentation network generates a basic mask, the hierarchical optimization unit refines and corrects its boundaries, resulting in a more accurate lung contour and reducing interference from surrounding tissue on subsequent analysis. The multi-scale feature extraction module extracts grayscale distribution features from the precise lung contour. The constructed initial small shadow probability distribution map can more comprehensively reflect possible shadow information and avoid feature omissions caused by single-scale analysis.

[0017] The morphological structure analyzer processes the initial probability distribution map, effectively identifying discrete shadow regions and marking suspicious shadow clusters, making the distribution of small shadows more clearly visible. A dynamic confidence feedback mechanism assesses the credibility of the morphological features of suspicious shadow clusters. The optimized small shadow feature map reduces the likelihood of misjudgments and missed detections, making the shadow information more realistic. A spatial relationship modeling network establishes a topological correlation matrix between shadow clusters and matches it with a pneumoconiosis pathology feature library. This allows the selection of target shadow regions that meet clinical pathology characteristics, making the selected regions more clinically valuable.

[0018] The 3D reconstruction engine processes the target shadow area to generate a heat map of the small shadow volume density, which can intuitively display the density distribution of the shadow. The hierarchical fusion module integrates the spatial coordinate information of the heat map with the original image. The output of the enhanced small shadow distribution map combines the density and spatial location information of the shadow, making the distribution of small shadows more intuitive and comprehensive. The structured presentation of diagnostic report data allows physicians to quickly obtain key information and facilitates comparison between different cases.

[0019] The entire process starts with precise segmentation of the lung area, then goes on to feature extraction, shadow recognition, credibility assessment, spatial relationship modeling, matching with the pathology library, three-dimensional reconstruction and atlas fusion, and finally generates a structured report, forming a complete processing chain. Each link cooperates with each other to improve the overall effect of small shadow detection and provide a more reliable image basis for the diagnosis of pneumoconiosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1This is a working principle diagram of the chest X-ray image processing method for AI small shadow detection rate for pneumoconiosis according to the present invention; Figure 2 Flowcharts for unit operations optimized for the hierarchy; Figure 3 Flowchart of the operation of the multi-scale feature extraction module; Figure 4 Flowchart of the dynamic confidence feedback mechanism; Figure 5 Flowchart of network operations for modeling spatial relationships; Figure 6 This is a flow chart for the detection and processing of small shadows in pneumoconiosis. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] See also Figure 1 The present invention provides a chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis, the method comprising: The system obtains an original chest X-ray image to be processed and performs an initial segmentation network on it, generating a base mask containing the lung region. A hierarchical optimization unit refines and modifies the boundaries of this base mask, outputting a precise lung contour. A multi-scale feature extraction module analyzes the grayscale distribution characteristics within this precise lung contour to construct an initial small shadow probability distribution map. This initial small shadow probability distribution map is processed using a morphological structure analyzer to identify discrete shadow regions and label suspicious shadow clusters. A dynamic confidence feedback mechanism assesses the credibility of the morphological features of these suspicious shadow clusters to generate an optimized small shadow feature map. This optimized small shadow feature map is input into a spatial relationship modeling network to establish a topological correlation matrix between shadow clusters. This topological correlation matrix is ​​then matched against a pre-defined pneumoconiosis pathology feature library to identify target shadow regions that meet clinical pathological characteristics. A 3D reconstruction engine is then used to process these target shadow regions, generating a small shadow volume density heat map. A hierarchical fusion module integrates this small shadow volume density heat map with the spatial coordinate information of the original image to output an enhanced small shadow distribution map. Structured diagnostic report data is generated based on this enhanced small shadow distribution map.

[0023] Raw chest X-ray images are acquired by medical imaging equipment and stored in a data processing system. Initial segmentation network processing involves a trained neural network model, and the hierarchical optimization unit uses a boundary correction vector field for fine-tuning. The multi-scale feature extraction module incorporates multi-resolution structures when processing grayscale distribution features. The morphological structure analyzer uses adaptive threshold segmentation and a connected domain algorithm. The dynamic confidence feedback mechanism uses a shadow feature evaluation network to calculate morphological confidence scores. The spatial relationship modeling network constructs a graph neural network model. Pathology feature library matching is based on cosine similarity calculations. The 3D reconstruction engine applies a voxel reconstruction algorithm when processing X-ray image sequences. The hierarchical fusion module performs an image registration algorithm.

[0024] See Figure 2 ,This embodiment involves the operation of the initial segmentation network processing and hierarchical optimization unit, which is applied to the processing flow of the original chest X-ray image.

[0025] The implementation process begins with the input of a chest X-ray image to be processed. This image is captured by medical imaging equipment and stored in a digital format, such as a DICOM file format, with a pixel resolution of 1024×1024 pixels and a grayscale value range of 0-255. The initial segmentation network constructs a two-channel convolutional neural network as the segmentation backbone. This segmentation backbone uses a pretrained architecture. Specifically, the spatial path consists of three convolutional layers: the first convolutional layer processes the input image using 32 3×3 convolutional kernels to extract high-resolution edge information; the second convolutional layer further refines the edge features using 64 3×3 convolutional kernels; and the third convolutional layer outputs an edge feature tensor using 128 3×3 convolutional kernels, which preserves the spatial details of the original image. The context path comprises a downsampling module: the initial layer uses 64 7×7 convolutional kernels to capture global context; subsequent layers combine max pooling operations to reduce the resolution and expand the output semantic feature tensor to 256 channels. This path integrates a global average pooling layer to aggregate image-level semantics. These two paths process the input image in parallel, with the spatial path focusing on local edges and the contextual path on overall structure. An adaptive feature fusion unit then merges the edge feature tensor generated by the spatial path with the semantic feature tensor generated by the contextual path. The fusion process incorporates a channel-wise attention mechanism: The average pooled value of the feature map is calculated as a weight, assigning higher weights to salient regions within the edge features. The weighted tensor is then combined using element-wise addition to generate a fused feature map, which is a three-dimensional array whose size is equal to the input image dimensions multiplied by the number of feature channels. Multi-level supervised training is performed on the fused feature map using a dataset of images labeled with lung regions. Training is repeated for 100 epochs. A backpropagation algorithm updates the weight parameters of the segmentation backbone network. The loss function used is binary cross-entropy, which measures the error between the predicted mask and the true label. The Adam optimizer is used with an initial learning rate of 0.001, which is reduced using an exponential decay strategy later in training. Multi-level supervision involves the addition of auxiliary losses at intermediate layers to ensure stable convergence. After training, the binary image output by the segmentation backbone network is used as the basic mask, which identifies the lung area in a black and white binary format, with the foreground pixel value being 1 and the background being 0.

[0026] The hierarchical optimization unit receives the base mask as input and executes a contour tracing algorithm to extract lung boundary pixel chains. Based on the boundary tracing method, the contour tracing algorithm scans the binary image of the base mask and identifies the outer boundary points of all connected regions. Freeman chain codes are used to encode the location information of boundary pixels, generating an ordered list of pixels, the lung boundary pixel chain, represented as a sequence of coordinate points. Subsequently, a curvature analysis model is used to calculate local curvature anomalies within the lung boundary pixel chain. This local curvature calculation involves moving a sliding window of 9 pixels across the pixel chain, assessing the curvature of the curve at each point. By comparing the directional changes in vectors of adjacent points, locations where the curvature exceeds a preset threshold of 0.1 are identified as outliers. These outliers correspond to discontinuities or jagged regions on the boundary. A boundary correction vector field is constructed based on the local curvature anomalies. Each outlier is assigned a correction vector whose direction points to a reference line at an anatomically standardized location. The reference line is predefined based on the lung anatomy, and the vector magnitude is calculated by distance difference. The boundary correction vector field is initialized as a field map covering the entire image area. This boundary correction vector field is optimized using an iterative energy minimization algorithm. The energy minimization algorithm defines an energy function consisting of two terms: a boundary smoothing term that measures the curvature consistency of boundary points, and a boundary constraint term that ensures that the boundary does not deviate too far from its original position. The optimization process uses gradient descent iterations, with the number of iterations set to 10. At each iteration, the gradient of the energy function is calculated, adjusting the direction and magnitude of the correction vector. Ultimately, the optimization process drives the chain of lung boundary pixels to smoothly move toward the anatomically standard position, eliminating abnormal protrusions or depressions. Upon completion of the optimization, the precise lung contour data is output, stored as a list of coordinate points, each containing x and y coordinate values, ensuring a continuous, smooth contour without breaks or overfitting.

[0027] In the overall implementation, the hardware platform uses a graphics processing unit (GPU) to accelerate computation. The segmentation backbone network is implemented in the PyTorch framework. The input image size is normalized to 512×512 pixels to speed processing. The channel attention mechanism of the adaptive feature fusion is implemented as an SE module. The contour tracing algorithm uses Python library functions to implement chain code operations. The threshold of the curvature analysis model is based on experimentally verified standard values. The reference lines of the boundary correction vector field are taken from a lung template database. The optimization iteration step size is set to 0.01 to ensure convergence. Data structures such as feature tensors are stored as NumPy arrays and coordinate point lists are stored in CSV format. The entire processing flow is executed sequentially, and intermediate results are stored in a memory cache to avoid repeated calculations and improve efficiency. After the precise lung contour is output, the system stores this data as a file or transmits it to subsequent modules.

[0028] The above implementation details the complete sequence of operations from the initial segmentation network processing and the hierarchical optimization unit, ensuring accurate lung segmentation in clinical applications. The initial segmentation network processing takes approximately 500 milliseconds to process image data, and the hierarchical optimization unit takes approximately 200 milliseconds. The overall system runs smoothly on standard hardware. Parameter adjustment of the segmentation backbone network is based on cross-validation, and the dataset is divided into training and validation sets. The contour tracing algorithm processes approximately 1,000 boundary points per scan. The curvature analysis model uses a sliding window with a one-pixel shift interval, analyzing each point individually. During the boundary correction vector field optimization process, gradient calculations use numerical methods to approximate derivatives. The energy minimization algorithm converges when the energy change is less than 0.001 or the maximum number of iterations is reached. The coordinate points of the output lung contour are evenly spaced, with spacing controlled to within 1 pixel. After integration, the system functions as a standalone processing unit, with input and output interfaces compatible with standard medical image formats. Training data augmentation techniques, including random rotation and scaling of the original image, enhance model robustness. A batch normalization layer is introduced during the feature fusion stage to stabilize the training process. The coordinate point list is serialized and stored as a binary file to save space. The execution engine uses multi-threaded processing to parallelize optimization calculations. During contour drawing, B-spline interpolation functions are used to smooth the output path. Input and output verification of the hierarchical optimization unit is performed through unit testing to ensure correct alignment. Finally, the processing results are integrated into the system log, recording the execution status of key steps.

[0029] See Figure 3 ,This embodiment involves the operation process of the multi-scale feature extraction module and the morphological structure analyzer, which is applied to the analysis and processing of the precise lung contour area.

[0030] The precise lung contour data output by Example 1 is obtained. The data exists in the form of a sequence of coordinate points in the format of an N×2 array, where N represents the number of boundary points. According to the contour coordinates, the lung area pixel blocks are extracted from the original chest X-ray image and cropped into a rectangular area image, and the size is automatically adapted to the maximum circumscribed rectangle. The multi-scale feature extraction module establishes a multi-resolution pyramid structure within the rectangular area. The pyramid structure is implemented as a three-level resolution hierarchy: the first layer maintains the original resolution, such as 1024×1024 pixels; the second layer generates a 512×512 pixel image by a downsampling factor of 2, using a bilinear interpolation algorithm; the third layer is downsampled again to a resolution of 256×256 pixels. Each layer of the image is processed independently.

[0031] A Difference of Gaussian filter bank is used to process the lung region of the original image, generating multi-scale gradient magnitude maps. The Difference of Gaussian filter bank consists of three independent filters: the first filter convolves the original layer image with a Gaussian kernel of σ = 1.0, outputting a smoothed image G1; the second filter processes the same layer image with σ = 1.6, producing G2; and the third filter, with σ = 2.0, produces G3. The differences between adjacent Gaussian smoothed images are calculated: D1 = |G1 - G2| and D2 = |G2 - G3|, forming two differential results. The Sobel gradient operator is applied to D1 and D2, extracting the X- and Y-direction derivatives, respectively, to synthesize the gradient magnitude maps M1 and M2. M1 and M2 are normalized to the range 0-255 to serve as scale-dependent gradient feature maps.

[0032] Local texture features are extracted from the gradient magnitude map. A sliding window method is used to scan the gradient map. The window size is set to 16×16 pixels, and the movement step is 8 pixels. At each window position, gray-level co-occurrence matrix properties, including contrast, correlation, and energy, are calculated. Texture patterns are calculated based on a rotation-invariant variant of the local binary pattern (LBP). The binary pattern is calculated for the eight-neighborhood region of the window's center pixel and circularly shifted to obtain the minimum decimal value as the invariant code. The gray-level co-occurrence matrix properties and the LBP code values ​​are combined to form a 128-dimensional rotation-invariant texture descriptor vector.

[0033] Calculate the similarity matrix between the descriptor vector and a standard pneumoconiosis texture template. A pre-established database of standard pneumoconiosis texture templates contains the mean texture feature vectors of 1,000 pneumoconiosis cases. Similarity is calculated using the cosine similarity metric: perform a dot product between the current descriptor vector V and the database template vector T, then divide the result by the modulus product. The result is stored in a similarity score matrix, where rows correspond to pixel window positions and columns correspond to different template indices. For each position, the highest similarity value is taken as the representative score S(x, y).

[0034] The spatial weights of the similarity matrix and the multi-scale gradient magnitude map are combined. The spatial weights are calculated based on the significance of the gradient magnitude map: a weighted sum of M1 and M2 is performed, with weights set to 0.6 and 0.4; the sum is normalized to produce the spatial weight map W(x,y). At the same spatial location, the texture similarity score S(x,y) is multiplied by the spatial weight W(x,y), and the product is mapped to a probability value P(x,y) using a sigmoid function. Bicubic interpolation is performed on all window positions to generate an initial small shadow probability distribution map at the same resolution as the original image, with probability values ​​ranging from 0 to 1.

[0035] The morphological structure analyzer operates on this probability distribution map. Adaptive threshold segmentation uses a local window algorithm: the image is divided into 32×32 pixel sub-blocks, and a threshold is calculated independently for each sub-block. The threshold calculation formula is based on the mean and standard deviation of the probability values ​​within the block, with the threshold set to the mean plus 0.5 times the standard deviation. The binarization operation sets pixels above the threshold to 1, and pixels otherwise to 0, generating a binary candidate region map.

[0036] The connected component labeling algorithm is used to process the binary candidate regions. The algorithm performs a four-neighborhood scan, traversing the binary image from left to right and top to bottom. When a foreground pixel (value = 1) is found, a seed-filling algorithm is initiated to mark the connected region. Starting from that pixel, the algorithm recursively checks the upper, lower, left, and right neighboring pixels. All connected foreground pixels are assigned the same label value. Label values ​​are assigned sequentially starting from 1, ultimately outputting a label matrix containing multiple discrete shadow regions.

[0037] Calculate the geometric morphological feature vector for each labeled region. The region area A is calculated as the total number of pixels within the region; the region boundary perimeter P is precisely calculated using a boundary tracing algorithm. The area irregularity index R = (P²) / (4πA). Calculate the boundary fractal dimension by using the box covering method at different scales s, counting the number of boxes N(s) required to cover the boundary. The slope of the fitted line in the log-log coordinate system is used as the dimension D. Output for each region is a feature vector containing the coordinate centroid, area, perimeter, irregularity, and fractal dimension.

[0038] The spatial distribution of eigenvectors was analyzed using a density clustering algorithm. The DBSCAN algorithm was implemented with the following input parameters: neighborhood radius ε = 5 pixels, minimum number of points minPts = 10. The algorithm uses the centroid coordinates of each region as data points; calculates the Euclidean distance between points; defines core points as points within a neighborhood of ε that has at least minPts points. Clusters are formed by expanding connected regions from the core points. Points that meet the clustering criteria are labeled as suspicious shadow clusters. Each cluster is assigned a unique identifier, ID, which records the original region label, centroid coordinates, feature mean, and spatial extent. The final output is a list of suspicious shadow clusters in the form of a structured array containing the ID, number of members, geometric statistic eigenvalues, and bounding box coordinates.

[0039] During implementation, all calculations were performed using single-precision floating-point arithmetic. The original image was loaded in 16-bit grayscale format. Pyramid generation utilized GPU parallel downsampling. Gaussian convolution employed separable convolution for performance optimization. Texture descriptor computation was accelerated using SIMD instructions. Similarity matrices were stored in a sparse format to conserve memory. A multi-layer pyramid fusion strategy was employed for probability distribution maps to enhance detail. The connected component labeling algorithm was optimized to a two-pass scanning method. Geometric feature calculations incorporated the addition of convex hull detection to improve accuracy. A density-clustered spatial index constructed a kd-tree to enhance search efficiency. The implementation system implemented an exception handling mechanism: when a region's area exceeds a threshold, the region segmentation strategy is automatically triggered to recalculate features. Variables such as ε and minPts are user-adjustable via a configurable interface. Data output is encapsulated in JSON format for compatibility with subsequent processing modules. Processing results are cached in an in-memory database, supporting breakpoint-resume processing. The morphological structure analyzer completes typical image analysis within 1.5 seconds.

[0040] Example 3: See Figure 4 , this embodiment relates to the operation process of the dynamic confidence feedback mechanism, which is a key component in the image processing process and is responsible for evaluating the credibility of the morphological features of suspicious shadow clusters. The implementation process starts with receiving the list of suspicious shadow clusters output by Example 2, which contains the ID identifier, the number of member regions, geometric feature statistics, and spatial bounding box coordinate information of each cluster. The construction of the shadow feature evaluation network adopts a modular design and includes two main components: a feature encoder and a confidence predictor. The feature encoder architecture is based on a convolutional neural network. The input layer receives 64×64 pixel image blocks, which are cropped from the original chest X-ray image. The cropping center position is determined by the centroid coordinates of the suspicious shadow cluster. When the shadow cluster spans multiple regions, the cropping area automatically adapts to the minimum circumscribed square.

[0041] The feature encoder's network architecture consists of five convolutional layers, each configured as follows: the first convolutional layer uses 32 5×5 convolution kernels with a stride of 2 and a ReLU activation function; the second convolutional layer uses 64 3×3 convolution kernels with a stride of 2 and a ReLU activation function; the third through fifth convolutional layers each use 128 3×3 convolution kernels with a stride of 1 and a ReLU activation function. Each convolutional layer is followed by a batch normalization layer to stabilize training. Max pooling is applied to the third and fifth convolutional layers, with a 2×2 pooling window and a stride of 2. A global average pooling layer is added at the end of the network to compress the feature maps into a fixed-length vector representation. When processing an input image patch, the feature encoder progressively extracts and abstracts visual features through hierarchical convolution operations, ultimately outputting a deep feature representation as a 512-dimensional vector that encapsulates the morphological characteristics of the image patch.

[0042] The confidence predictor design consists of two fully connected layers: the first fully connected layer has 256 neurons and uses ReLU activation; the second fully connected layer outputs a single neuron, and the output value is compressed to the (0, 1) interval through the sigmoid function to represent the morphological confidence score. The scoring calculation process can be expressed as:

[0043] in, Represents the input 512-dimensional deep feature vector, and Represent the weight matrix and bias term of the first fully connected layer respectively, represents the ReLU activation function, and are the parameters of the second fully connected layer, Represents the sigmoid function, which ultimately outputs the confidence score This score quantifies the confidence that the current image block contains a small pneumoconiosis shadow. A higher score indicates a higher degree of match between the morphological features and the typical pneumoconiosis shadow.

[0044] This implementation of the non-maximum suppression algorithm uses a spatial overlap criterion to process multiple suspicious shadow clusters. The algorithm first sorts all shadow clusters in descending order of their confidence scores. When processing the queue, the current highest-scoring cluster is selected as the benchmark, and the intersection over union (IoU) between its bounding box and the bounding boxes of the remaining clusters is calculated. When the IoU value exceeds a preset threshold of 0.3, the corresponding low-scoring cluster is marked as suppressed. This process is iterated until all clusters have been processed. The remaining shadow clusters constitute the final screening result set, which are sufficiently discriminative in their spatial distribution while maintaining high morphological confidence.

[0045] The coordinate alignment operation maps the filtered shadow clusters back to the original image space. This process is based on the calculation of a spatial transformation matrix, which records the conversion relationship from the coordinates of the cropped image block to the coordinates of the original image. For each retained shadow cluster, its centroid coordinates are relocated to the precise location in the original image through an inverse transformation. At the same time, all relevant geometric feature parameters (such as area and boundary shape) are adjusted accordingly based on the coordinate transformation ratio to ensure the consistency of feature values. The resulting optimized small shadow feature matrix uses a structured storage format and contains fields such as cluster ID, original image coordinates (x, y), confidence score, area adjustment value, and boundary complexity index.

[0046] At the implementation level, hardware acceleration utilizes the CUDA parallel computing architecture. The convolution operations of the feature encoder are optimized using the cuDNN library, with a batch size of 32. The confidence predictor's parameter initialization uses the Xavier method to avoid the vanishing gradient problem. The IoU calculation of the non-maximum suppression algorithm is accelerated through vectorized operations. Bilinear interpolation is used to maintain geometric accuracy during coordinate alignment. Regarding memory management, deep feature representations use float16 half-precision storage to save space. TensorRT optimization is enabled during the network inference phase to reduce computational latency. Input image patch preprocessing includes grayscale normalization and local contrast enhancement. An exception handling mechanism monitors the distribution of confidence scores, triggering recalculation when extreme values ​​occur. The parameter configuration interface allows adjustment of the confidence threshold and the overlap threshold for non-maximum suppression. The data output format is DICOM-compliant and includes original coordinate reference information. The processing log records the detailed computation path for each shadow cluster, enabling result traceability. Performance optimization measures include asynchronous data transmission and pipeline parallel processing to improve overall throughput. The quality control module regularly verifies coordinate alignment accuracy to ensure spatial positioning accuracy. System resource monitoring tracks GPU memory usage in real time to prevent overflow errors. During implementation, the weight parameters of the feature encoder are initialized through pre-training, and transfer learning is completed on the pneumoconiosis image dataset. The training of the confidence predictor adopts a balanced sampling strategy to ensure an appropriate ratio of positive and negative samples. The overall processing time of the dynamic confidence feedback mechanism is controlled to be completed within 800 milliseconds per image, meeting clinical real-time requirements. The result visualization component superimposes the optimized small shadow features on the original image, using color gradients to indicate the confidence level to assist physicians in review. System integration testing verifies the image compatibility of X-ray equipment from different manufacturers to ensure wide applicability. The maintenance interface provides online update function for the feature encoder and supports iterative optimization of the model. The security mechanism implements data encryption transmission and access control to protect patient privacy information.

[0047] See Figure 5 This embodiment involves the operational process of matching a spatial relationship modeling network with a pathology feature library, and its implementation is illustrated through a specific example. Taking a clinical chest X-ray image as an example, this image has a size of 2048×2048 pixels and a bit depth of 16 bits. The image has been processed through the aforementioned embodiment to obtain an optimized small shadow feature matrix containing feature data for six suspicious shadow clusters. The specific contents of the shadow feature matrix are shown in the following table:

[0048] After receiving this feature matrix, the spatial relationship modeling network first extracts the spatial coordinates of all shadow clusters. Taking cluster C1 (456,789) as an example, the system establishes a graph structure data model, treating each shadow cluster as a graph node. Node attributes include coordinate position, confidence score, and morphological characteristics. Edge connections between nodes are established based on the principle of spatial proximity. The Euclidean distance between all node pairs is calculated, and a connecting edge is established when the distance is less than 100 pixels. The edge weight value is set to the inverse of the distance, so that the closer the distance, the higher the connection strength between nodes. In this example, the distance between clusters C1 and C2 is 58 pixels, and the distance between clusters C1 and C3 is 34 pixels, both meeting the connection conditions. Therefore, two edges are established with weights of 0.017 and 0.029, respectively.

[0049] The graph neural network model uses a two-layer graph convolutional architecture to process the graph data. The first convolutional layer inputs node features with a dimension of 5 (including coordinates, confidence, area, boundary complexity, and grayscale value), and the output dimension is expanded to 32. The feature update process for each node aggregates the weighted features of its neighboring nodes, with the weights derived from the edge connection strength. The second convolutional layer compresses the feature dimension to 16, highlighting key spatial relationship features. After graph convolution processing, the feature vector of node C1 is updated to a 16-dimensional representation that contains spatial context information, where the values ​​of each dimension reflect the relative position and feature association of the node in the overall spatial distribution.

[0050] The topological association matrix is ​​generated based on the processed node features. The cosine similarity of each pair of node features is calculated, with a similarity threshold of 0.7. Node pairs above this threshold are considered to have a significant topological association. In this example, clusters C1 and C2 have a similarity of 0.83, and clusters C3 and C4 have a similarity of 0.79, both exceeding the threshold. Therefore, clusters C1 and C4 have a strong association in the matrix. However, the similarity with cluster C4 is only 0.32, indicating no association. The resulting topological association matrix is ​​a 6×6 symmetric matrix with binary values ​​(1 for association and 0 for no association).

[0051] During the pathological feature library matching phase, the system accesses a pre-built standard pneumoconiosis shadow feature library. This library contains three typical pneumoconiosis shadow distribution patterns: diffuse, nodular, and mixed. Each pattern stores a standard topological graph structure and feature vector. For example, the feature vector of the nodular pattern emphasizes the circularity index and spatial clustering density of small shadows. Key path features are extracted from the current topological association matrix. The specific operations are as follows: identifying the largest connected subgraph, which in this example is a triangular connection structure consisting of {C1, C2, C3}; and calculating the characteristic parameters of this subgraph, including an average node degree of 2.0, a clustering coefficient of 0.87, and an average path length of 1.0.

[0052] The extracted sub-image features are compared with the standard library for similarity. This similarity is calculated using a feature vector-by-feature vector comparison method. The match between the current example sub-image and the nodular pattern in the standard library reaches 82%, exceeding the preset 75% match threshold. Therefore, this group of shadow clusters is determined to be consistent with pneumoconiosis pathology. The system outputs the target shadow region coordinates, including the precise location information for clusters C1, C2, and C3. The matching pattern type is also recorded as nodular, which serves as a reference for subsequent diagnostic analysis.

[0053] In terms of implementation details, spatial relationship modeling is implemented using the PyTorch Geometric graph neural network framework. The pathology feature library is stored in a graph database, supporting fast topological queries. Node features are normalized to ensure consistent numerical ranges across all dimensions. The distance threshold for edge connectivity conditions is dynamically adjusted based on image resolution. Mean pooling is used as the aggregation function in the graph convolutional layer. The topological correlation matrix is ​​stored in a sparse format to optimize spatial efficiency. Parallel computing is implemented to accelerate feature comparison during the matching process. An exception handling mechanism monitors the numerical validity of node features to prevent overflow errors. The result visualization component colors related shadow clusters, visually displaying spatial distribution patterns. The quality control module regularly verifies the connectivity of the graph structure to ensure the accuracy of topological analysis. The system parameter configuration interface allows adjustment of distance and similarity thresholds. The data processing pipeline is automated, requiring no manual intervention from feature matrix input to target region output. Node feature updates are incrementally processed to reduce memory consumption. Edge weight calculations are cached to avoid redundant computations. The pathology feature library supports dynamic expansion, allowing the addition of new standard patterns at any time. Security mechanisms implement access control to protect sensitive patient data. Performance monitoring records the time taken for each step for system optimization analysis.

[0054] This embodiment relates to the operation process of the 3D reconstruction engine and the hierarchical fusion module.

[0055] A continuous slice X-ray image sequence is acquired, containing multi-angle projection images of the patient during inspiration and expiration. The system loads the spatial coordinate information of the target shadow area and parses the device scanning parameters through DICOM metadata. The 3D reconstruction engine uses a voxel reconstruction algorithm to process the image sequence. The algorithm first generates an empty voxel space with a resolution of 512×512×256 voxel cells, each initialized to zero. All input images are traversed, and the ray path is calculated based on the projection geometry parameters. Distance-weighted interpolation is applied as the ray traverses the voxel space, and the attenuation coefficient of the penetrating ray is accumulated for each voxel value. The reconstruction process implements a multi-resolution strategy: a low-resolution is used to quickly construct the global structure in the initial stage, and after two iterations, it switches to high-resolution refinement. The axial sampling spacing automatically adapts to the device slice thickness parameter, with a typical value of 0.5 mm.

[0056] Post-processing of 3D volume data involves grayscale gradient calculation and anisotropic diffusion optimization. Grayscale gradient features are extracted using a 3D Sobel operator, with the operator kernel configured with 3×3×3 convolution weights in the X, Y, and Z directions. The magnitude and direction of the gradient vector are calculated for each voxel, and the result is stored as a 4D array (X, Y, Z, Mag). An anisotropic diffusion filter is implemented based on this gradient data, with the diffusion coefficient function dynamically adjusted based on the gradient magnitude. The filter is iterated over five rounds: the first round uses a large diffusion scale to smooth uniform regions; subsequent iterations gradually narrow the diffusion range; and the final round smoothes only regions with weak gradients. The diffusion equation is solved using implicit discretization, with Neumann zero-flux conditions as the boundary conditions. The processed 3D data exhibits an optimized signal-to-noise ratio, and the boundary continuity of small shadow regions is enhanced.

[0057] A grayscale-to-density conversion model is established to generate density heatmaps of small shadow volumes. The model defines a segmented mapping relationship between gradient amplitude and density value: amplitudes of 0-50 correspond to densities of 0-0.3; amplitudes of 50-150 are mapped to densities of 0.3-0.7; and values ​​above 150 are mapped to densities of 0.7-1.0. This mapping is performed for each voxel, generating density distribution volume data. Heatmap color encoding uses HSV color space conversion: density values ​​are used as hue channel parameters, and a blue-green-yellow-red gradient represents low to high density; saturation is fixed at 90%; and the lightness channel is adjusted based on the original grayscale value. The resulting three-dimensional heatmap dataset contains a three-channel array of voxel coordinates, density values, and RGB color values.

[0058] The hierarchical fusion module performs spatial coordinate transformation operations. This module loads the original baseline X-ray image and parses its DICOM coordinate system parameters. It calculates the projection transformation matrix from the 3D thermal map space to the 2D image plane. The transformation integrates the device geometric parameters: source-image distance, source-object distance, and detector inclination. Using a perspective projection model, the 3D voxel coordinates are converted to 2D homogeneous coordinates, which are then aligned to the original image grid via an affine transformation. The projection process implements a ray tracing algorithm: virtual rays are emitted from the X-ray source point through the 3D voxel space, accumulating the density and color values ​​passed through to generate an intermediate projection image.

[0059] The fusion process utilizes multi-channel synthesis. The original image serves as the background layer, preserving the grayscale anatomical structure; the projected heat map serves as the foreground layer, adding pseudo-color information. Alpha blending is used for the fusion strategy, with transparency set to 70% for low-density areas and 30% for high-density areas. Color space conversion ensures that the fused image retains the grayscale contrast of the original image. The enhanced small shadow distribution map is generated, including a dual-view display mode: the primary view displays the fused image, while the secondary view provides 3D rotation of the heat map.

[0060] The generation of structured diagnostic reports is based on density zoning statistics. A grid density counting system is established in the projection coordinate system: the lung field is divided into three zones: upper, middle and lower, and each zone is further divided into three zones: inner, middle and outer. The system automatically identifies the distribution of small shadows in each zone: continuous areas with density values ​​greater than 0.3 are considered valid shadows. Statistical indicators include the total number of shadow pixels in each zone, the maximum continuous shadow size, the average density value and the coordinates of the distribution center of gravity. The report data is structured in XML format and contains patient information, equipment parameters, statistical data and image reference links. The report generation module also creates a thumbnail of the distribution map in JPG format and embeds it in the XML document for easy viewing.

[0061] The system is implemented using a layered architecture. The 3D reconstruction engine uses CUDA kernel functions to accelerate voxel calculations. Heat map generation deploys OpenGL shaders for real-time rendering. The coordinate transformation module integrates the ITK library to handle spatial registration. The exception handling mechanism monitors ray casting anomalies in voxel reconstruction. The quality control subsystem verifies the 3D reconstruction accuracy through phantom testing. The processing parameter configuration interface provides manual adjustment functions for diffusion intensity and color mapping curves. The performance monitoring log records the processing time of each image, with an average of 3.5 seconds for the reconstruction phase and 1.2 seconds for the fusion phase. The result data storage adopts a layered compression strategy, and the original projection data is archived independently. The security mechanism implements patient information desensitization processing to comply with medical data protection regulations. The output interface is compatible with the PACS system transmission protocol and supports automatic uploading of diagnostic reports and distribution maps. The system maintenance module provides offline calibration of the reconstruction engine to ensure long-term accuracy and stability.

[0062] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0063] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A chest X-ray image processing method for AI small shadow detection rate of pneumoconiosis, characterized by: The following steps are involved: Acquire a chest X-ray original image to be processed, perform initial segmentation network processing on the original image, and generate a basic mask containing a lung area; A hierarchical optimization unit is used to refine and correct the boundaries of the basic mask to output an accurate lung contour; Analyzing the grayscale distribution characteristics within the precise lung contour through a multi-scale feature extraction module to construct an initial small shadow probability distribution map; Processing the initial small shadow probability distribution map based on a morphological structure analyzer, identifying discrete shadow areas and marking suspicious shadow clusters; Using a dynamic confidence feedback mechanism to evaluate the credibility of the morphological features of the suspicious shadow cluster, and generate an optimized small shadow feature map; Inputting the optimized small shadow feature map into a spatial relationship modeling network to establish a topological association matrix between shadow clusters; Matching the topological association matrix with a preset pneumoconiosis pathological feature library to screen target shadow areas that meet clinical pathological characteristics; Using a three-dimensional reconstruction engine to process the target shadow area to generate a small shadow volume density heat map; The small shadow volume density heat map and the spatial coordinate information of the original image are integrated through a hierarchical fusion module to output an enhanced small shadow distribution map; Structured diagnostic report data is generated based on the enhanced small shadow distribution map.

2. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The initial segmentation network processing specifically includes: Constructing a dual-channel convolutional neural network as a segmentation backbone network, wherein the segmentation backbone network includes a spatial path and a context path operating in parallel; Extracting high-resolution edge features of the original image through the spatial path to generate an edge feature tensor; Capturing the global semantic features of the original image using the context path to generate a semantic feature tensor; Adopting an adaptive feature fuser to merge the edge feature tensor and the semantic feature tensor to generate a fused feature map; Performing multi-level supervised training on the fused feature map, and updating the weight parameters of the segmentation backbone network through a back-propagation algorithm; The binary image output by the segmentation backbone network after training convergence is used as the basic mask.

3. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 2, characterized in that: The operations of the hierarchical optimization unit include: receiving the basic mask and executing a contour tracing algorithm to extract a lung boundary pixel chain; Using a curvature analysis model to calculate the local curvature abnormal points of the lung boundary pixel chain; constructing a boundary correction vector field based on the local curvature outlier points; Optimizing the boundary correction vector field through an iterative energy minimization algorithm to drive the lung boundary pixel chain to move toward an anatomical standard position; The output is accurate lung contour data containing smooth continuous contours.

4. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The operation steps of the multi-scale feature extraction module are: establishing a multi-resolution pyramid structure within the precise lung contour; Processing the lung region of the original image using a Gaussian difference filter bank to generate a multi-scale gradient magnitude map; Extracting local texture features of the multi-scale gradient magnitude map and constructing a rotation-invariant texture descriptor; Calculating a similarity matrix between the rotationally invariant texture descriptor and a standard pneumoconiosis texture template; The similarity matrix and the spatial weight of the multi-scale gradient magnitude map are fused to generate initial small shadow probability distribution data.

5. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The Morphological Structure Analyzer performs the following operations: Performing adaptive threshold segmentation on the initial small shadow probability distribution map to obtain a binary candidate region; Using a connected component labeling algorithm to process the binary candidate region to generate a discrete shadow region identifier; Calculating a geometric feature vector of each discrete shadow area, wherein the geometric feature vector includes area irregularity and boundary fractal dimension; The spatial clustering characteristics of the geometric morphological feature vectors are analyzed based on a density clustering algorithm, and suspicious shadow clusters that meet the clustering constraints are marked.

6. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The dynamic confidence feedback mechanism includes: Constructing a shadow feature evaluation network, wherein the shadow feature evaluation network includes a feature encoder and a confidence predictor; Inputting the original image block of the suspicious shadow cluster into a feature encoder to generate a deep feature representation; Calculating a morphological confidence score of the deep feature representation by a confidence predictor; A non-maximum suppression algorithm is used to screen suspicious shadow clusters whose morphological confidence scores are higher than a critical threshold; Align the screening results with the original image coordinates to generate the optimized small shadow feature matrix.

7. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The operations of the spatial relationship modeling network include: receiving the optimized small shadow feature matrix and extracting spatial position coordinates; Construct a graph neural network model, with each shadow cluster as a node and the Euclidean distance between nodes as the edge weight; The spatial dependency between nodes is learned through the graph convolution layer, and the node feature representation is updated; Calculate the topological correlation strength between the updated node feature representation and the adjacent nodes; Outputs topological correlation matrix data containing the connection strengths between nodes.

8. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The pathological feature library matching process includes: Pre-building a standard pneumoconiosis shadow feature library, wherein the standard pneumoconiosis shadow feature library stores shadow distribution patterns of typical pneumoconiosis cases; Extracting a key path eigenvector from the topological association matrix; Calculating the cosine similarity between the key path feature vector and a standard pneumoconiosis shadow feature library; Filtering the topological subgraphs whose cosine similarity is greater than a matching threshold; The topological subgraph is mapped back to the original image space, and a target shadow region coordinate set is output.

9. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 1, characterized in that: The operation steps of the 3D reconstruction engine are as follows: Acquire a sequence of X-ray images of consecutive slices corresponding to the target shadow area; Processing the X-ray image sequence using a voxel reconstruction algorithm to generate three-dimensional volume data; Calculating the grayscale gradient characteristics of each voxel in the three-dimensional volume data; Optimizing the signal-to-noise ratio of the three-dimensional volume data based on an anisotropic diffusion filter; Generate 3D heat map data that annotates the spatial distribution of small shadows.

10. The chest X-ray image processing method for improving the AI ​​small shadow detection rate for pneumoconiosis according to claim 9, characterized in that: The hierarchical fusion module performs: Receiving spatial coordinate transformation parameters of the three-dimensional thermal map data and the original image; Establish a conversion function from small shadow density to color mapping; The three-dimensional thermal map data is projected onto the two-dimensional diagnostic plane through image registration algorithm; The projection data is integrated with the anatomical structure features of the original image to generate an enhanced small shadow distribution map; A structured report including the number and distribution locations of shadows is generated based on the density partition statistics of the enhanced small shadow distribution map.

Images