Lung ct image nodule detection method based on selection state space and convolution model

Abstract

This invention provides a lung CT image nodule detection method based on a selection state space (SSM) and convolutional model. It deeply integrates the SSM with a 3D convolutional neural network (3DCNN), fully leveraging the powerful long-range spatial dependency modeling capabilities of the SSM to capture the global contextual relationship between lung nodules and surrounding tissues in CT sequences, providing crucial spatial prior knowledge. Simultaneously, its dynamic selection mechanism adaptively filters the most valuable features for nodule identification (such as minute edges and ground-glass opacities), dynamically providing crucial supplementary information. This dynamic selection mechanism highly simulates the behavior of doctors prioritizing key areas, making the model more closely resemble actual diagnostic scenarios. This invention significantly improves the sensitivity (reducing the false negative rate) and specificity (reducing the false positive rate) of lung nodule detection, enabling rapid and accurate assistance to doctors in locating and identifying lung nodules, providing efficient technical support for early lung cancer screening.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more particularly to a method for detecting lung CT nodules based on a selected state space and a convolution model. Background Technology

[0002] Lung cancer, as one of the deadliest malignant tumors worldwide, poses a significant threat to human health. Early detection and accurate diagnosis of pulmonary nodules are crucial for improving patient survival rates. Low-dose computed tomography (CT) has become the gold standard for lung cancer screening, generating high-resolution three-dimensional images of the lungs. However, radiologists face significant challenges in manually interpreting these images: the sheer volume of CT slides leads to a heavy workload and visual fatigue, and small, ground-glass opacities, or occult nodules are prone to missed diagnoses (low sensitivity) and misdiagnoses (high false positives).

[0003] To address these challenges, computer-aided diagnostic systems have emerged, aiming to assist doctors in improving diagnostic efficiency and accuracy. Utilizing deep learning, particularly convolutional neural network (CNN) technology, these systems have demonstrated significant potential in medical image analysis. In lung nodule detection, 3D CNNs have become the mainstream method for computer-aided diagnosis of lung nodules due to their ability to directly process CT volume data and effectively capture the three-dimensional spatial morphological features of nodules (such as size, shape, and texture). Their performance is generally superior to 2D CNN models that only process two-dimensional slices. For example, Dou et al. have improved detection accuracy by designing more complex 3D CNN structures (such as multi-level context encoding and attention mechanisms) and attempted to alleviate the problem of positive and negative sample imbalance. However, despite significant progress in 3D CNNs, their performance in complex real-world scenarios still faces bottlenecks. Insufficient modeling of long-range spatial dependencies is a major challenge. The kernel size of 3D CNNs limits their receptive field, making it difficult to efficiently model the global information of a single slice in CT images and the long-distance spatial contextual relationships spanning multiple slices (e.g., the relationship between nodules and distal blood vessels or the pleura). In scenarios with complex backgrounds and weak nodule features, models struggle to adaptively focus on key clues and are susceptible to irrelevant noise. This global information is crucial for distinguishing between benign and malignant nodules or ruling out false positives (such as vascular crossings).

[0004] In recent years, Selective State Space Models (SSMs) (such as Mamba) have made groundbreaking progress in the field of long sequence modeling. Their core advantage lies in their efficient long-range dependency modeling; SSMs can efficiently capture ultra-long-range dependencies in sequence data (such as CT slice sequences) with near-linear complexity. The core innovation of SSMs lies in their selection mechanism. This mechanism allows the model to dynamically decide which historical information should be retained, updated, or ignored based on the current input, thereby achieving adaptive focusing on key contextual information. This capability is crucial for identifying subtle features with diagnostic value.

[0005] Deeply integrating the efficient long-range modeling capabilities and dynamic selection characteristics of SSM with the powerful local spatial feature extraction capabilities of 3DCNN offers a promising new approach to addressing the aforementioned bottlenecks of 3DCNN. Currently, there is relatively little exploration in the field of this type of deeply integrated SSM and 3DCNN architecture for lung nodule detection tasks.

[0006] In summary, addressing the shortcomings of current 3DCNN-based lung nodule detection models in long-range spatial context modeling and dynamic key information filtering, this invention proposes an innovative approach that deeply fuses a Selective State Space Model (SSM) with a 3D Convolutional Neural Network (3DCNN). Through a carefully designed fusion mechanism, SSM provides 3DCNN with long-range spatial dependencies and dynamically filtered key contextual information (i.e., powerful prior and supplementary knowledge) that are difficult for 3DCNN to obtain. This synergistically enhances the model's ability to characterize complex lung nodules and its robustness in discrimination, ultimately achieving high-sensitivity, low-false-positive-rate lung nodule detection and contributing to accurate early screening of lung cancer. Summary of the Invention

[0007] This invention addresses the problem of low efficiency in lung nodule detection in real-world clinical scenarios. Specifically, existing deep learning-based lung nodule detection methods are insufficient in effectively modeling the global spatial dependencies of CT sequences and dynamically filtering key contextual information. This often results in limited ability of the model to identify complex nodules (such as micronodules, ground-glass opacities, or nodules in adjacent structures), exhibiting low sensitivity and a high false positive rate. To overcome the shortcomings of existing technologies, this invention provides a lung CT image nodule detection method based on a selected state space and a convolutional model.

[0008] This invention is achieved through the following technical solution: a method for detecting lung CT nodules based on a selected state space and a convolution model, specifically including the following steps:

[0009] S1: Dataset preprocessing and grouping; Integrate the 10 lung CT image data contained in the luna16 dataset, preprocess the CT image data, convert it into a machine-readable storage type, and save it separately; Group the preprocessed image data into training set, validation set and test set.

[0010] S2: Feature extraction from the training data; first, a feature encoder based on ResNet50 is used to extract features from the preprocessed training set to obtain feature maps of lung nodule images; and the output feature maps of each layer are then processed. Retain and pass on to the next stage;

[0011] S3: Design the state space selection module; the module consists of two parts: the selection state space block (SSM) and the frequency domain information enhancement block (FIE); it will process the feature map extracted by the encoder. First, select the state space block. Processing to obtain results Then via frequency domain information enhancement block After processing, the final output of the selected state space module is obtained as follows: ;

[0012] S4: Design a cross-scale semantic feature information fusion module; the module consists of two parts: cross-scale feature fusion (CSF) and multi-channel feature concatenation (MCF); the results extracted by the selected state space module are then processed. First, cross-scale feature fusion is performed. get Then through multi-channel feature cascade , obtain the output result ;

[0013] S5: The final output result obtained from the cross-scale semantic feature information fusion module After passing through two different branches The convolution is used to obtain classification and regression results; then patches are truncated based on the predicted nodule center position. Then, the fully connected layer extraction method is used to judge false positives and obtain the final prediction result.

[0014] As a preferred approach, in step S1, the preprocessed image data is divided into groups, with 70% of the data used as the training set, 10% as the validation set, and 20% as the test set.

[0015] As a preferred option, step S3 specifically includes the following steps:

[0016] S3-1: Design the state space selection module; the module consists of two parts: the state space selection block SSM and the frequency domain information enhancement block FIE;

[0017] S3-2: The selected state space block is composed of convolutional modules, employing a linear normalization layer + 2D selective scan block (SS2D) + Se block and dual residual connections. The normalization layer, the self-attention mechanism SS2D and SeBlock are sequentially input, and the target state space is selected by using self-attention and 2D selective scan blocks.

[0018] S3-3: The frequency domain information enhancement block also uses convolutional modules. A learnable wavelet decomposition layer simulated by convolutional blocks distinguishes between high-frequency (HF) and low-frequency (LF) information. This is then passed through a high-frequency attention enhancement module consisting of residual convolutional blocks, a 3D batch normalization layer, and an adaptive max-3D pooling layer, and a low-frequency attention enhancement module consisting of residual convolutional blocks, a 3D instance normalization layer, an adaptive average 3D pooling layer, and a pose normalization block. The enhanced high-frequency and low-frequency features are added together along the channel dimension using a cat expression, and then restored to the original channel size through several convolutional layers, resulting in the final output of the selected state space module. .

[0019] As a preferred option, step S4 specifically includes the following steps:

[0020] S4-1: Design a cross-scale semantic feature information fusion module; this module includes two parts: cross-scale feature fusion (CSF) and multi-channel feature concatenation (MCF);

[0021] S4-2: The cross-scale feature fusion part first selects the state space module. The above-ground processing is performed using bilinear interpolation to obtain the above-ground data. To make its size and Consistency; To Band matrix processing is performed to extract detailed information. ,right Diagonal matrix processing is performed to extract global information. Then and Perform dot product processing. and Perform a dot product, then use a learnable parameter. The processed output is then fused at the pi-level to obtain the result. The specific formula is expressed as follows:

[0022]

[0023] in These are learnable parameters. It is a dot product in dimensions. It involves multiplying element-wise; by integrating detailed features extracted from high-level semantics into low-level semantics, and simultaneously integrating global features extracted from low-level semantics into high-level semantics, the two are then fused more evenly using learnable parameters to obtain the output of the cross-scale feature fusion part. ;

[0024] S4-3: The multi-channel feature cascade MCF part establishes the correlation between feature information channels; firstly, the cross-scale feature fusion part obtains... The feature information is processed by using the chunks function to divide it into three equal parts, obtaining... And then The output of the convolution process enters the second branch. The results are added together, then processed by dilated convolution with a dilation rate of 2, and the output enters the third branch. Add them together, then... The output is obtained by superimposing along the channel dimension. Then through convolution and The results are added together and then convolved to obtain the final output. The specific formula is expressed as follows:

[0025]

[0026] in, It's a regular convolution. It is a superposition along the channel dimension; it divides the features into Then, the features of the three branches are processed using ordinary convolution, dilated convolution with a dilation rate of 2, and dilated convolution with a dilation rate of 4, respectively, and top-down information interaction is achieved; the final output model feature result is... .

[0027] As a preferred option, step S5 specifically includes the following steps:

[0028] S5-1: Output of the cross-scale semantic feature information fusion module After passing through two different branches The convolutions are used to obtain classification and regression results;

[0029] S5-2: Patch with a size of 48*48*48 based on the predicted nodule center cutoff , Will be through by The convolutional block, consisting of a convolutional layer, a batch normalization layer, and a max pooling layer with a stride of 2, yields the final output feature map. Output feature map The input is processed by the average pooling layer, i.e. ;

[0030] in, This represents an average pooling layer with a step size of 2; then... The final output result is obtained by processing through a fully connected layer. ,in This is represented as a fully connected layer with 6 input channels;

[0031] S5-3: The final output result is... It consists of coordinates, diameter, and confidence level.

[0032] This invention, employing an innovative technical solution that deeply integrates a selection state-space model with a 3D convolutional neural network, offers the following advantages compared to existing technologies: Unlike traditional methods that primarily rely on 3DCNN structures or their variants for feature extraction, this invention innovatively integrates a highly efficient selection state-space model into the lung nodule detection framework. This model fully leverages the long-range spatial dependency modeling capabilities of the selection state-space mechanism, capturing the global contextual relationships between lung nodules and surrounding complex anatomical structures across multiple CT image slices. Simultaneously, its core dynamic selection mechanism adaptively filters and focuses on the most valuable feature information (such as edge features and special textures) for current nodule identification based on input data, effectively compensating for the shortcomings of 3DCNNs in complex scenarios, such as insufficient feature information or inadequate attention to key information. More importantly, the dynamic focusing characteristics of the selection state-space model closely align with the diagnostic cognitive process of radiologists, simulating their prioritization and focused examination of potential lesion areas during image review, making the model's learning and decision-making more relevant to real-world applications. This deep synergistic fusion significantly improves the model's ability to characterize the overall features of lung nodules and its robustness in discrimination. It can quickly and accurately assist doctors in locating and judging nodules in lung CT images, providing strong technical support for the early screening of lung cancer.

[0033] Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0034] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0035] Figure 1 This is a flowchart of the method of the present invention;

[0036] Figure 2 This is a network architecture diagram of the present invention;

[0037] Figure 3 This is a structural diagram of the selected state space module of the present invention;

[0038] Figure 4 This is a structural diagram of the cross-scale semantic feature information fusion module of the present invention. Detailed Implementation

[0039] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0041] The following is combined Figures 1 to 4 The lung nodule detection method based on 3DCNN and dual-branch structure technology according to embodiments of the present invention will be described in detail.

[0042] like Figure 1 As shown, this invention proposes a method for detecting lung CT nodules based on a selected state space and a convolution model, specifically including the following steps:

[0043] S1: Dataset preprocessing and grouping; The 10 lung CT image data contained in the luna16 dataset are integrated, the CT image data are preprocessed, converted into machine-readable storage type, and saved separately; The preprocessed image data is grouped and divided into training set, validation set and test set; 70% of the data is used as training dataset, 10% of the data is used as validation set, and 20% of the data is used as test set.

[0044] S2: Feature extraction from training data; first, a feature encoder based on ResNet50 is used to extract features from the previously preprocessed training set to obtain feature maps of lung nodule images; and the output feature maps of each layer ( ) to retain and pass on to the next; such as Figure 2 As shown.

[0045] S3: Design the state space selection module; the module consists of two parts: the selection state space block (SSM) and the frequency domain information enhancement block (FIE); the feature map extracted by the encoder... First, select the state space block ( ) Results Then via frequency domain information enhancement block ( The final output of the selected state-space module is as follows: ,like Figure 3 As shown; specifically including the following steps:

[0046] S3-1: Design the state space selection module; the module consists of two parts: the state space selection block (SSM) and the frequency domain information enhancement block (FIE);

[0047] S3-2: The selected state space block is composed of convolutional modules, employing a linear normalization layer + 2D selective scan block (SS2D) + Se block and dual residual connections. The normalization layer, the self-attention mechanism SS2D and SeBlock are sequentially input, and the target state space is selected by using self-attention and 2D selective scan blocks.

[0048] S3-3: The frequency domain information enhancement block also uses convolutional modules. A learnable wavelet decomposition layer simulated by convolutional blocks distinguishes between high-frequency (HF) and low-frequency (LF) information. This is then passed through a high-frequency attention enhancement module consisting of residual convolutional blocks, a 3D batch normalization layer, and an adaptive max-3D pooling layer, and a low-frequency attention enhancement module consisting of residual convolutional blocks, a 3D instance normalization layer, an adaptive average 3D pooling layer, and a pose normalization block. The enhanced high-frequency and low-frequency features are added together along the channel dimension using a cat expression, and then restored to the original channel size through several convolutional layers, resulting in the final output of the selected state space module. .

[0049] S4: Design a cross-scale semantic feature information fusion module; this module includes two parts: cross-scale feature fusion (CSF) and multi-channel feature concatenation (MCF); the results extracted by the selected state space module will be used to fuse the semantic feature information. First, cross-scale feature fusion is performed. )get Then through multi-channel feature cascade ( ), to obtain the output result ,like Figure 4 As shown; specifically including the following steps:

[0050] S4-1: Design a cross-scale semantic feature information fusion module; this module includes two parts: cross-scale feature fusion (CSF) and multi-channel feature concatenation (MCF);

[0051] S4-2: The cross-scale feature fusion part first selects the state space module. The above-ground processing is performed using bilinear interpolation to obtain the above-ground data. To make its size and Consistency; to more effectively integrate detailed information from high-level semantics with global information from low-level semantics, for Band matrix processing is performed to extract detailed information. ,right Diagonal matrix processing is performed to extract global information. Then and Perform dot product processing. and Perform a dot product, then use a learnable parameter. The processed output is then fused at the pi-level to obtain the result. The specific formula is expressed as follows:

[0052]

[0053] in These are learnable parameters. It is a dot product in dimensions. It involves multiplying element-wise. By integrating detailed features extracted from high-level semantics into low-level semantics, and simultaneously integrating global features extracted from low-level semantics into high-level semantics, and then using learnable parameters to achieve a more balanced fusion of the two, the output of the cross-scale feature fusion component is obtained. ;

[0054] S4-3: The Multi-Channel Feature Concatenation (MCF) section establishes the correlation between feature information channels; firstly, the cross-scale feature fusion section obtains... The feature information is processed by using the chunks function to divide it into three equal parts, obtaining... And then The output of the convolution process enters the second branch. The results are added together, then processed by dilated convolution with a dilation rate of 2, and the output enters the third branch. Add them together, then... The output is obtained by superimposing along the channel dimension. Then through convolution and The results are added together and then convolved to obtain the final output. The specific formula is expressed as follows:

[0055]

[0056] in, It's a regular convolution. It is a superposition along the channel dimension. The features are divided into... Then, the features of the three branches are processed using ordinary convolution, dilated convolution with a dilation rate of 2 and 4 respectively, and top-down information interaction is implemented to establish the correlation between channels and to obtain the channel spatial dependencies; the final output model feature result is .

[0057] S5: The final output result obtained from the cross-scale semantic feature information fusion module After passing through two different branches The convolution is used to obtain classification and regression results; then patches are truncated based on the predicted nodule center position. Then, using a fully connected layer extraction method, false positive detection is performed to obtain the final prediction result; specifically, the following steps are included:

[0058] S5-1: Output of the cross-scale semantic feature information fusion module After passing through two different branches The convolutions are used to obtain classification and regression results;

[0059] S5-2: Based on the predicted nodule center, a patch of size 48*48*48 is extracted. ), Will be through by The convolutional block, consisting of a convolutional layer, a batch normalization layer, and a max pooling layer with a stride of 2, yields the final output feature map. Output feature map The input is processed by the average pooling layer, i.e. ;

[0060] in, This represents an average pooling layer with a step size of 2; then... The final output result is obtained by processing through a fully connected layer. ,in This is represented as a fully connected layer with 6 input channels;

[0061] S5-3: The final output result is... It consists of coordinates, diameter, and confidence level.

[0062] In the description of this invention, the term "a plurality of" refers to two or more. Unless otherwise explicitly defined, the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. The terms "connection," "installation," "fixing," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0063] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0064] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lung CT image nodule detection method based on a selection state space and a convolution model, characterized by Specifically, it includes the following steps: S1: Dataset preprocessing and grouping; Integrate the 10 lung CT image data contained in the luna16 dataset, preprocess the CT image data, convert it into a machine-readable storage type, and save it separately; Group the preprocessed image data into training set, validation set and test set. S2: feature extraction is performed on the training data; a feature encoder with ResNet50 as the basic framework is used to perform feature extraction on the images of the training set divided after previous preprocessing, to obtain a feature map of the lung nodule image; and the output feature map of each layer is reserved and passed back ; S3: Design the state space selection module; the module consists of two parts: the state space selection block (SSM) and the frequency domain information enhancement block (FIE); it will process the feature map extracted by the encoder. First, select the state space block. Processing to obtain results Then via frequency domain information enhancement block After processing, the final output of the selected state space module is obtained as follows: Specifically, it includes the following steps: S3-1: Design the state space selection module; the module consists of two parts: the state space selection block SSM and the frequency domain information enhancement block FIE; S3-2: The selected state space block is composed of convolutional modules, employing a linear normalization layer, a 2D selective scan block (SS2D), a compression and excitation block, and dual residual connections. The normalization layer, the self-attention mechanism SS2D, and the compression and activation block are sequentially input, and the selection of the target state space is achieved by using self-attention and 2D selective scanning blocks. S3-3: The frequency domain information enhancement block also uses convolutional modules. A learnable wavelet decomposition layer simulated by convolutional blocks distinguishes between high-frequency (HF) and low-frequency (LF) information. This is then passed through a high-frequency attention enhancement module consisting of residual convolutional blocks, a 3D batch normalization layer, and an adaptive max-3D pooling layer, and a low-frequency attention enhancement module consisting of residual convolutional blocks, a 3D instance normalization layer, an adaptive average 3D pooling layer, and a pose normalization block. The enhanced high-frequency and low-frequency features are added together along the channel dimension using a cat expression, and then restored to the original channel size through several convolutional layers, resulting in the final output of the selected state space module. ; S4: Design a cross-scale semantic feature information fusion module; the module consists of two parts: cross-scale feature fusion (CSF) and multi-channel feature concatenation (MCF); the results extracted by the selected state space module are then processed. First, cross-scale feature fusion is performed. get Then through multi-channel feature cascade , obtain the output result ; S5: The final output result obtained from the cross-scale semantic feature information fusion module After passing through two different branches The convolution is used to obtain classification and regression results; then patches are truncated based on the predicted nodule center position. Then, the fully connected layer extraction method is used to judge false positives and obtain the final prediction result.

2. The method for detecting lung CT nodules based on a selected state space and a convolution model according to claim 1, characterized in that, In step S1, the preprocessed image data is divided into groups, with 70% of the data used as the training set, 10% as the validation set, and 20% as the test set.

3. The method for detecting lung CT nodules based on a selected state space and a convolution model according to claim 1, characterized in that, Its features are, Step S4 specifically includes the following steps: S4-1: Design a cross-scale semantic feature information fusion module; this module includes two parts: cross-scale feature fusion (CSF) and multi-channel feature concatenation (MCF); S4-2: The cross-scale feature fusion part first selects the state space module. The above-ground processing is performed using bilinear interpolation to obtain the above-ground data. To make its size and Consistent; To Band matrix processing is performed to extract detailed information. ,right Diagonal matrix processing is performed to extract global information. Then and Perform dot product processing. and Perform a dot product, then use a learnable parameter. The processed output is then fused at the pi-level to obtain the result. The specific formula is expressed as follows: ， in These are learnable parameters. It is a dot product in dimensions. It involves multiplying element-wise; by integrating detailed features extracted from high-level semantics into low-level semantics, and simultaneously integrating global features extracted from low-level semantics into high-level semantics, the two are then fused more evenly using learnable parameters to obtain the output of the cross-scale feature fusion part. ; S4-3: The multi-channel feature cascade MCF part establishes the correlation between feature information channels; firstly, the cross-scale feature fusion part obtains... The feature information is processed by using the chunks function to divide it into three equal parts, obtaining... And then The output of the convolution process enters the second branch. The results are added together, then processed by dilated convolution with a dilation rate of 2, and the output enters the third branch. Add them together, then... The output is obtained by superimposing along the channel dimension. Then through convolution and The results are added together and then convolved to obtain the final output. The specific formula is expressed as follows: ， in, It's a regular convolution. It is a superposition along the channel dimension; it divides the features into Then, the features of the three branches are processed using ordinary convolution, dilated convolution with a dilation rate of 2, and dilated convolution with a dilation rate of 4, respectively, and top-down information interaction is achieved; the final output model feature result is... .

4. The method for detecting lung CT nodules based on a selected state space and a convolution model according to claim 1, characterized in that, Its features are, Step S5 specifically includes the following steps: S5-1: Output of the cross-scale semantic feature information fusion module After passing through two different branches The convolutions are used to obtain classification and regression results; S5-2: Patch with a size of 48*48*48 based on the predicted nodule center cutoff , Will be through by The convolutional block, consisting of a convolutional layer, a batch normalization layer, and a max pooling layer with a stride of 2, yields the final output feature map. Output feature map The input is processed by the average pooling layer, i.e. ; in, This represents an average pooling layer with a step size of 2; then... The final output result is obtained by processing through a fully connected layer. ,in This is represented as a fully connected layer with 6 input channels; S5-3: The final output result is... It consists of coordinates, diameter, and confidence level.

Images