A lung nodule auxiliary detection method based on multi-channel nonlinear density adaptive mapping
By using a multi-channel nonlinear density adaptive mapping method, multi-channel inputs are generated for candidate ROI image blocks of lung nodules, which solves the problems of missed and false detection of low-contrast lung nodules and improves the accuracy and efficiency of lung nodule detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN AGRI UNIV
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-21
AI Technical Summary
In the early screening of lung cancer, low-dose spiral CT is prone to missing or misdetecting low-contrast, poorly defined ground-glass opacities. Furthermore, existing methods are computationally intensive or lack targeted enhancement, resulting in poor detection performance.
A multi-channel nonlinear density adaptive mapping method is adopted to generate a first reference channel and multiple enhancement mapping channels for candidate ROI image blocks of lung nodules. The density interval of lung nodules is separated from the background interference interval through a differential continuous mapping mechanism, and then input into a deep neural network for detection.
In the input characterization stage, background information is compressed, and the density range of lung nodules is enhanced in a targeted manner to improve the discrimination ability of the downstream detection network, reduce computational costs, and improve detection accuracy.
Smart Images
Figure CN122048932B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image data processing, and in particular to a lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping. Background Technology
[0002] In early lung cancer screening, low-dose spiral CT is the most commonly used imaging method. Lung nodules, especially ground-glass opacities, often exhibit low contrast, blurred boundaries, and partial overlap with normal alveolar tissue and small blood vessel density, easily leading to missed and false detections. Existing technologies mainly suffer from the following problems: One type of approach relies on deeper or more complex network structures, such as cascaded convolutional networks, attention modules, or multi-stage detection pipelines. While these can improve performance to some extent, they have high computational and deployment costs and do not address the low signal-to-noise ratio problem of low-contrast lung nodules at the input representation level. Another type of approach uses conventional window width and level adjustments, fixed HU range truncation, or multi-width linear normalization. A common characteristic of these methods is that they apply a uniform linear mapping slope to all density intervals, failing to provide targeted amplification for the density interval containing the lung nodule. There are also some general image enhancement methods, such as histogram equalization or ordinary segmented stretching. While these can improve overall contrast, they typically do not incorporate functional segment design based on the physical density distribution of the lung nodule, easily introducing enhancements unrelated to the detection target, or even amplifying non-target structures. Therefore, there is an urgent need for a preprocessing and input construction method that can combine the density distribution characteristics of lung nodules, especially low-contrast lung nodules, so that lung nodules can obtain a grayscale representation that is more suitable for detection before entering the downstream neural network. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a lung nodule auxiliary detection method based on multi-channel nonlinear density adaptive mapping, thus solving the deficiencies of the prior art.
[0004] The objective of this invention is achieved through the following technical solution: a lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping, the method comprising:
[0005] S1. Process the original chest CT data to generate candidate ROI image blocks for lung nodules;
[0006] S2. For each candidate ROI image block of a lung nodule, simultaneously generate a first reference channel and an enhancement mapping channel;
[0007] S3. For the enhanced mapping channel, the physical density space is divided into functional segments according to the physical density distribution of lung nodules, and a differentiated continuous mapping mechanism is adopted to make the target density interval where the lung nodules are located occupy a larger representation bandwidth.
[0008] S4. The first reference channel and the enhanced mapping channel are used as the joint input method to input into the deep neural network, and the output is one or more of the following: lung nodule candidate probability map, candidate score, and candidate region.
[0009] S1 specifically includes the following:
[0010] Read the original CT image data of the case to be processed, and convert the original pixel values of the original CT image data into HU values. HU stands for Huntsfield unit, which is used to characterize the degree of radiation attenuation of different tissues relative to water and air.
[0011] The center point and bounding box of the lung nodule can be provided by manual annotation, or the candidate coordinates can be output by an existing CAD system, or the doctor can obtain the candidate center point of the lung nodule by clicking on the specified area of the suspicious lung nodule on the workstation;
[0012] Based on the candidate center point of the lung nodule, a fixed-size candidate ROI image block of lung nodule is extracted from the original CT image data. If there are multiple candidate lung nodule regions in the same case, multiple candidate ROI image blocks of lung nodule are generated respectively.
[0013] The first reference channel is used to preserve the lung field contour, interlobar structure, vascular orientation, and anatomical background around the lung nodule, thereby providing spatial context for the deep neural network to determine whether the region truly corresponds to a lung nodule.
[0014] The functional segmentation of the physical density space includes dividing the physical density space into three continuous functional segments: a low-density suppression area, a target enhancement area, and a high-density suppression or saturation area. Through functional segmentation, the density intervals related to lung nodules are separated from the background and interference.
[0015] The low-density suppression zone is used for compressed air and low-density lung parenchyma background information;
[0016] The target enhancement region is used to cover the density distribution range of lung nodules, so that they can obtain a larger representation bandwidth in the output grayscale space;
[0017] The high-density inhibition or saturation zone is used to compress blood vessels, bronchial walls, soft tissues, and other high-density interference structures.
[0018] The differential continuous mapping mechanism includes:
[0019] Apply suppression to the low-density suppression region, the target enhancement region, and the high-density suppression or saturation region respectively. , , The mapping gain, which satisfies , and The requirement is to increase the mapping gain corresponding to the target enhancement region. Higher than the other two sections;
[0020] when At this time, in the low-density suppression region, the normalized output value after mapping... ,in, The input HU value is the pixel to be mapped. To truncate the lower bound globally, This marks the boundary between the low-density suppression region and the target enhancement region.
[0021] when <x≤ At this time, in the target enhancement region, the normalized output value after mapping... Where μ represents the statistical center of the density distribution of the target lung nodules, s represents the scale parameter of the target enhancement region, and σ() is the Sigmoid function. This is the boundary point between the target enhancement region and the high-density suppression or saturation region;
[0022] when At this time, in the high-density suppression or saturation region, the normalized output value after mapping... ,in, The upper bound is used for global truncation.
[0023] The enhancement mapping channels include wide-area enhancement mapping channels, basic enhancement mapping channels, and focused enhancement mapping channels; a multi-channel input tensor with C=4 channels is formed by using one first reference channel and three enhancement mapping channels and input into the deep neural network;
[0024] The wide-area enhanced mapping channel: through the... , This was achieved by extending the sample outwards to preserve more of the peripheral density changes associated with lung nodules. This marks the boundary between the low-density suppression region and the target enhancement region. This is the boundary point between the target enhancement region and the high-density suppression or saturation region;
[0025] The underlying enhancement mapping channel: adopts the standard , Mapping is performed to stabilize and highlight the main density regions of lung nodules;
[0026] The focus enhancement mapping channel: through the... , It is obtained by inward contraction, which is used to increase the difference between the core density region of the lung nodule and the adjacent background.
[0027] The output of the lung nodule candidate probability map includes:
[0028] After encoding and decoding the input tensor, the deep neural network obtains the probability value of each pixel belonging to the lung nodule region after the single channel is activated by Sigmoid, thus forming a lung nodule candidate probability map. This probability map reflects the network's confidence that different locations within the current ROI belong to lung nodules.
[0029] The output of the candidate region includes:
[0030] The candidate probability map of lung nodules is binarized according to a preset threshold to obtain the initial candidate mask of lung nodules;
[0031] Perform connected component analysis on the candidate mask to extract one or more connected candidate regions;
[0032] By combining rules such as area threshold, shape constraint, whether it is located inside the lung field, and its prior location relationship with known lung nodules, false candidate regions that do not conform to the morphological characteristics of lung nodules are filtered out.
[0033] The remaining connected regions were used as candidate regions for lung nodules.
[0034] The output of the candidate scores includes:
[0035] The average probability of all pixels within a candidate region is used as the candidate score for that region, or the maximum probability within a candidate region is used as the candidate score, to reflect the highest confidence level of the deep neural network in that candidate region.
[0036] If the deep neural network includes an auxiliary classification branch, the output of the classification branch is fused with the segmentation probability pooling result to obtain the candidate score.
[0037] The present invention has the following advantages: a lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping, which compresses background information that is weakly associated with the detection target during the input characterization stage, performs targeted enhancement on the density range where lung nodules are mainly distributed, suppresses high-density interference structures such as blood vessels and soft tissues, and improves the discrimination ability of the downstream detection network through the joint input of the first reference channel and one or more enhanced mapping channels. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application provided below with reference to the accompanying drawings is not intended to limit the scope of protection of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. The present invention will be further described below with reference to the accompanying drawings.
[0040] This invention specifically relates to a lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping, addressing the problems of existing technologies such as the inability of uniform linear windowing to effectively highlight lung nodule density ranges, the high computational cost of complex network schemes, and the lack of lung nodule specificity in general enhancement methods. The core of this invention does not lie in simply changing the network, but in altering the representation of lung nodules in the input grayscale space. For the same CT region of interest image block, a first reference channel and one or more enhancement mapping channels are simultaneously generated. These multiple generated channels are stacked along the channel dimension to form an input tensor with at least two channels, preferably multiple channels. This tensor is then input into a deep neural network to output a lung nodule candidate probability map, candidate score, candidate region, or a combination thereof.
[0041] like Figure 1 As shown, it specifically includes the following:
[0042] S1, Data Acquisition and Channel Generation;
[0043] This step is used to extract basic input information directly related to the auxiliary detection of lung nodules from the raw chest CT data. First, the raw CT image data of the case to be processed is read. This image data can be a DICOM sequence or other medical image data containing pixel physical density information. The raw pixel values are preferably converted to HU values before subsequent processing, where HU is an abbreviation for Hounsfield Unit, used to characterize the degree of radiation attenuation of different tissues relative to water and air.
[0044] In pulmonary nodule-assisted detection scenarios, the region to be analyzed can be determined in the following ways: manually annotating the center point and bounding box of the pulmonary nodule, or using a CAD system to output candidate coordinates, or having a doctor click to specify a suspected pulmonary nodule region on a workstation. Based on the candidate pulmonary nodule centers, a fixed-size region of interest (ROI) image block is extracted from the original chest CT data, preferably 96×96 pixels. If multiple candidate pulmonary nodule regions exist in the same case, multiple candidate ROI image blocks can be generated separately.
[0045] For each candidate ROI image patch for lung nodules, a first reference channel and one or more enhancement mapping channels are generated.
[0046] The primary function of the first reference channel is to preserve the lung field contour, interlobar structures, vascular orientation, and the anatomical background surrounding the lung nodules, thereby providing spatial context for subsequent network determination of whether the region truly corresponds to a lung nodule. In a preferred embodiment of lung CT, the first reference channel can use lung window parameters with a window width of 1500 HU and a window level of -600 HU, or equivalent linear normalization parameters, to generate a first reference channel matrix ranging from [0,1].
[0047] One or more enhanced mapping channels: Based on the physical density distribution of lung nodules, the density space is divided into functional segments, and differentiated continuous mapping is used to make the target density interval where the lung nodules are located occupy a larger effective grayscale bandwidth.
[0048] S2. Divide the physical density space into functional zones;
[0049] This invention does not consider all HU density values as equally important. Instead, based on the density distribution relationship between lung nodules and background and interfering tissues, the physical density space is divided into at least three continuous functional segments: a low-density inhibition zone, a target enhancement zone, and a high-density inhibition or saturation zone.
[0050] The low-density suppression area is used to compress air and background information of very low-density lung parenchyma; the target enhancement area is used to mainly cover the density distribution range of lung nodules (especially low-contrast lung nodules) so that they can obtain a larger representation bandwidth in the output grayscale space; the high-density suppression or saturation area is used to compress blood vessels, bronchial walls, soft tissues and other high-density interference structures.
[0051] In the preferred embodiment of low-contrast pulmonary nodules, the cutoff threshold meets the ranges suggested in Table 1 below:
[0052] Table 1. Threshold Range Table
[0053]
[0054] in, Its function is to distinguish pixels that mainly belong to air or low-density background from pixels that may belong to the target region of lung nodules; Its function is to distinguish between the density zones of lung nodules that need to be highlighted and the high-density interference zones such as blood vessels and soft tissues; and These are used to limit the impact of excessively low and high density values on the mapping results, respectively.
[0055] To avoid limiting the present invention to a single fixed empirical value, the present invention preferably employs a statistical adaptive method to determine the aforementioned threshold. Specifically, the HU values of the target lung nodule region can be collected from training samples, based on the lower quantile of the lung nodule HU distribution (recommended). ) and upper quantile (recommended) Automatic determination and In the currently formally validated main implementation, the automatically estimated typical value is approximately... , .
[0056] By dividing the functional areas as described above, this invention can separate the density regions related to lung nodules from the background and interference during the input preprocessing stage, providing a foundation for subsequent multi-channel mapping and deep network detection.
[0057] S3, Differentiated Continuous Mapping Mechanism;
[0058] After completing the global truncation, the present invention applies different mapping gains to the three functional segments mentioned above, so that the gain corresponding to the target enhancement area is higher than that of the two side segments, so as to allocate a larger representation bandwidth for the lung nodule target interval in the output gray space [0,1], thereby improving the separability of lung nodules (especially low-contrast lung nodules) from background, blood vessels and other interfering structures.
[0059] The gain coefficients corresponding to the low-density suppression region, the target enhancement region, and the high-density suppression or saturation region are denoted as follows: , , And preferably satisfying the following relationship:
[0060] ,
[0061] ,
[0062] ,
[0063] Typical value examples: In the main implementation method .
[0064] The purpose of this mapping formula is to redistribute the HU values in the original CT image to the normalized grayscale space of [0,1], so that the main density range of lung nodules is stretched to the medium-high grayscale range that the network is more sensitive to, while the air background and high-density interference are compressed to a narrower grayscale band, thereby improving the separability of the target.
[0065] Let the input HU value of the pixel to be mapped after global truncation be x. Then the output y is calculated according to the following piecewise formula (ensuring that it is within the boundary point). , (continuous, without jumps)
[0066] when Time (low-density suppression region):
[0067] ,
[0068] when <x≤ (Target Enhancement Zone):
[0069] ,
[0070] when Time (high-density suppression region):
[0071] .
[0072] Where μ represents the statistical center of the density distribution of the target lung nodule, s represents the scale parameter of the target enhancement region, and σ() is the Sigmoid function. This function performs continuous, smooth, non-linear stretching on the main density range of the lung nodule within the target enhancement region, resulting in density values near the statistical center of the lung nodule that have a more suitable grayscale representation for network recognition, while avoiding significant abrupt changes near segment boundaries. Compared to simple linear stretching, this function makes the grayscale distribution within the target enhancement region smoother, thereby improving the separability of lung nodules from the background and high-density interference structures.
[0073] Through this continuous segmented mapping, the target density range of lung nodules, especially low-contrast lung nodules, is stretched, while the air background and high-density vascular interference are compressed, thereby enabling lung nodules to achieve a clearer contrast representation in the input grayscale space. The focus of this design is not on a fixed curve itself, but on achieving a larger representation bandwidth for the density range corresponding to the target lung nodule in the output grayscale space through functional segmentation and differentiated gain configuration.
[0074] S4. Constructing input tensors using multiple channels;
[0075] The first reference channel provides the overall anatomical context, enabling the downstream neural network to recognize the spatial relationships between lung nodules and the lung parenchyma, blood vessels, and pleura. The enhancement mapping channel highlights the target density region where the lung nodule is located, making it easier for the network to distinguish real lung nodules from background or interfering structures.
[0076] Enhanced mapping channels include three types: wide-area enhanced mapping channels: through... , Outward expansion is used to preserve more peripheral density changes associated with lung nodules; Baseline enhancement mapping channel: using standard , Mapping is performed to stabilize and highlight the main density region of lung nodules; focused enhancement mapping channel: by... , It is obtained by inward contraction, which is used to further increase the difference between the core density area of the lung nodule and the adjacent background, and is particularly suitable for low-contrast, blurred-border ground-glass-related lung nodules.
[0077] Each channel originates from the same candidate ROI image patch for lung nodules, thus possessing a natural spatial alignment relationship. They can be directly stacked along the channel dimension without additional registration. This preserves the original lung anatomy background while enabling the network to simultaneously observe the same candidate lung nodule region at different enhancement scales, thereby improving its ability to identify lung nodules.
[0078] This invention covers both at least dual-channel and multi-channel implementations. The currently formally validated main implementation uses one first reference channel and three enhanced mapping channels (wide-domain, basic, and focused), ultimately forming a multi-channel input tensor with C=4 channels. This multi-channel input tensor can provide richer and more targeted input representations for subsequent lung nodule-assisted detection.
[0079] S5, Deep Neural Network Inference;
[0080] The constructed multi-channel input tensor is fed into a deep neural network for inference. The network can be a segmentation network, a detection network, a joint segmentation and classification network, or other convolutional neural networks, preferably a two-dimensional U-Net-like network. In one implementation, a channel attention module consisting of dilated convolutions and global pooling can also be combined to adjust the feature weights of different input channels, enabling the network to pay more attention to key channels and key regions related to lung nodule-assisted detection.
[0081] After encoding and decoding the input Region of Interest (ROI) tensor, the deep neural network outputs a single-channel lung nodule probability map corresponding to the input spatial size. This single-channel output, after sigmoid activation, yields the probability value of each pixel belonging to a lung nodule region, thus forming a lung nodule candidate probability map. This probability map reflects the network's confidence level in identifying different locations within the current ROI as lung nodules.
[0082] Based on the lung nodule candidate probability map, lung nodule candidate regions and lung nodule candidate scores can be further obtained. The specific process includes:
[0083] 1. Binarize the probability map according to a preset threshold to obtain the initial lung nodule candidate mask;
[0084] 2. Perform connected component analysis on the candidate mask to extract one or more connected candidate regions;
[0085] 3. By combining rules such as area threshold, shape constraint, whether it is located inside the lung field, and its prior positional relationship with known lung nodules, false candidate regions that obviously do not conform to the morphological characteristics of lung nodules are filtered out.
[0086] 4. The remaining connected regions are used as candidate regions for lung nodules.
[0087] For the obtained candidate lung nodule regions, candidate scores can be calculated by averaging, maximizing, or weighted averaging the pixel probabilities within each candidate region. For example, the average probability of all pixels within a candidate region can be used as the candidate score, or the highest probability within the candidate region can be used as the candidate score to reflect the network's highest confidence level for that lung nodule candidate region. If the network also includes an auxiliary classification branch, the output of the classification branch can be fused with the segmentation probability pooling result to obtain the final lung nodule candidate score.
[0088] When only one region with the highest score is retained within a Region of Interest (ROI) (i.e., the region with the highest score obtained after sorting these candidate regions by average probability, maximum probability, or fusion score), this region can be used as a candidate region for lung nodules within that ROI. When multiple connected candidate regions are retained within the same ROI after probability map binarization, connected component analysis, and rule-based filtering, the top k results can be output after sorting by candidate scores. Subsequently, the candidate regions for lung nodules can be remapped back to the original CT image coordinate system and displayed on the original image as contours, borders, or semi-transparent masks, thus forming the final auxiliary detection results for lung nodules, which can be used by doctors for auxiliary image interpretation and subsequent analysis.
[0089] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and improvements, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping, characterized in that: The method includes: S1. Process the original chest CT data to generate candidate ROI image blocks for lung nodules; S2. For each candidate ROI image block of a lung nodule, simultaneously generate a first reference channel and an enhancement mapping channel; S3. For the enhanced mapping channel, the physical density space is divided into functional segments according to the physical density distribution of lung nodules, and a differentiated continuous mapping mechanism is adopted to make the target density interval where the lung nodules are located occupy a larger representation bandwidth. S4. The first reference channel and the enhanced mapping channel are used as the joint input method to input into the deep neural network, and the output is one or more of the following: lung nodule candidate probability map, candidate score, and candidate region. The first reference channel uses lung window parameters with a window width of 1500 HU and a window level of -600 HU, or equivalent linear normalization parameters, to generate a first reference channel matrix in the range of [0,1]. The enhanced mapping channel divides the density space into functional segments based on the physical density distribution of lung nodules and uses differentiated continuous mapping to make the target density interval where the lung nodules are located occupy a larger effective grayscale bandwidth. The functional segmentation of the physical density space includes dividing the physical density space into three continuous functional segments: a low-density suppression area, a target enhancement area, and a high-density suppression or saturation area. Through functional segmentation, the density intervals related to lung nodules are separated from the background and interference. The low-density suppression zone is used for compressed air and low-density lung parenchyma background information; The target enhancement region is used to cover the density distribution range of lung nodules, so that they can obtain a larger representation bandwidth in the output grayscale space; The high-density inhibition or saturation zone is used to compress blood vessels, bronchial walls, soft tissues, and other high-density interference structures. The differential continuous mapping mechanism includes: Apply suppression to the low-density suppression region, the target enhancement region, and the high-density suppression or saturation region respectively. , , The mapping gain, which satisfies , and The requirement is to increase the mapping gain corresponding to the target enhancement region. Higher than the other two sections; when At this time, in the low-density suppression region, the normalized output value after mapping... ,in, The input HU value is the pixel to be mapped. To truncate the lower bound globally, This marks the boundary between the low-density suppression region and the target enhancement region. when <x≤ At this time, in the target enhancement region, the normalized output value after mapping... Where μ represents the statistical center of the density distribution of the target lung nodules, s represents the scale parameter of the target enhancement region, and σ() is the Sigmoid function. This is the boundary point between the target enhancement region and the high-density suppression or saturation region; when At this time, in the high-density suppression or saturation region, the normalized output value after mapping... ,in, The upper bound is used for global truncation.
2. The lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping according to claim 1, characterized in that: S1 specifically includes the following: Read the original CT image data of the case to be processed, and convert the original pixel values of the original CT image data into HU values. HU stands for Huntsfield unit, which is used to characterize the degree of radiation attenuation of different tissues relative to water and air. The center point and bounding box of the lung nodule can be provided by manual annotation, or the candidate coordinates can be output by an existing CAD system, or the doctor can obtain the candidate center point of the lung nodule by clicking on the specified area of the suspicious lung nodule on the workstation; Based on the candidate center point of the lung nodule, a fixed-size candidate ROI image block of lung nodule is extracted from the original CT image data. If there are multiple candidate lung nodule regions in the same case, multiple candidate ROI image blocks of lung nodule are generated respectively.
3. The lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping according to claim 1, characterized in that: The first reference channel is used to preserve the lung field contour, interlobar structure, vascular orientation, and anatomical background around the lung nodule, thereby providing spatial context for the deep neural network to determine whether the region truly corresponds to a lung nodule.
4. The lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping according to claim 1, characterized in that: The enhancement mapping channel includes a wide-area enhancement mapping channel, a basic enhancement mapping channel, and a focus enhancement mapping channel; A multi-channel input tensor with C=4 channels is formed by using one first reference channel and three enhancement mapping channels and input into the deep neural network. The wide-area enhanced mapping channel: through the... , This was achieved by extending the sample outwards to preserve more of the peripheral density changes associated with lung nodules. This marks the boundary between the low-density suppression region and the target enhancement region. This is the boundary point between the target enhancement region and the high-density suppression or saturation region; The underlying enhancement mapping channel adopts the standard. , Mapping is performed to stabilize and highlight the main density regions of lung nodules; The focus enhancement mapping channel: through the focus enhancement mapping channel: , It is obtained by inward contraction, which is used to increase the difference between the core density region of the lung nodule and the adjacent background.
5. The lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping according to claim 1, characterized in that: The output of the lung nodule candidate probability map includes: After encoding and decoding the input tensor, the deep neural network obtains the probability value of each pixel belonging to the lung nodule region after the single channel is activated by Sigmoid, thus forming a lung nodule candidate probability map. This probability map reflects the network's confidence that different locations within the current ROI belong to lung nodules.
6. The lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping according to claim 1, characterized in that: The output of the candidate region includes: The candidate probability map of lung nodules is binarized according to a preset threshold to obtain the initial candidate mask of lung nodules; Perform connected component analysis on the candidate mask to extract one or more connected candidate regions; By combining rules such as area threshold, shape constraint, whether it is located inside the lung field, and its prior location relationship with known lung nodules, false candidate regions that do not conform to the morphological characteristics of lung nodules are filtered out. The remaining connected regions were used as candidate regions for lung nodules.
7. The lung nodule-assisted detection method based on multi-channel nonlinear density adaptive mapping according to claim 1, characterized in that: The output of the candidate scores includes: The average probability of all pixels within a candidate region is used as the candidate score for that region, or the maximum probability within a candidate region is used as the candidate score, to reflect the highest confidence level of the deep neural network in that candidate region. If the deep neural network includes an auxiliary classification branch, the output of the classification branch is fused with the segmentation probability pooling result to obtain the candidate score.