Nodule detection method suitable for ultrasonic image

Abstract

The invention discloses a nodule detection method suitable for an ultrasonic image. The method comprises the following steps: carrying out feature extraction on an input ultrasonic image by utilizing a backbone network of an integrated space self-adaptive deformation enhancement module, and introducing deformable convolution to improve the perception capability on an irregular marginal region; constructing a feature enhancement module fusing global and local information, and performing multi-scale fusion on feature maps extracted by the backbone network based on a double-branch network of Fourier transform and large kernel convolution; and inputting the enhanced feature map into an encoder and an interactive learning decoder to realize global context modeling and regression decoding, and outputting a nodule detection result through a composite loss function training model of joint classification and regression loss. The method can be widely applied to ultrasonic image nodule detection tasks of various organs, the radiograph reading burden of ultrasonic doctors can be effectively reduced, objective and quantitative auxiliary diagnosis indexes are provided, and the accuracy and efficiency of clinical diagnosis are improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, to a deep learning-based method for nodule detection in ultrasound images of the thyroid gland, breast, etc. Background Technology

[0002] Ultrasound examination, due to its non-invasive, real-time, and portable characteristics, has become the preferred imaging method for screening and diagnosing diseases of superficial organs such as the thyroid and breast. In clinical practice, sonographers need to interpret images to detect and diagnose nodular lesions, a process that is time-consuming and easily influenced by the physician's subjective experience, potentially leading to inaccurate diagnostic results. Therefore, utilizing computer-aided diagnostic (CAD) technology to assist physicians in the rapid, objective, and accurate detection and classification of nodules is of crucial clinical significance.

[0003] However, due to the characteristics of nodules, such as blurred boundaries, irregular shapes, and large size variations, detection systems based on traditional manual methods or classic convolutional neural networks (CNNs) often have low accuracy in practical applications, making it difficult to meet the needs of high-precision clinical auxiliary diagnosis. In recent years, DETR (Detection Transformer) detection methods have attracted widespread attention from researchers due to their advantages such as end-to-end training and the elimination of the need for manual anchor box design. However, due to the long-range modeling characteristics of Transformers, their sensitivity to local geometric features is poor, making them unsuitable for detecting nodules with blurred edges and complex structures in ultrasound images. In addition, the traditional encoder-decoder structure is a hierarchical cascade, which limits feature interaction and information fusion between different layers, resulting in limited improvement in detection performance. Therefore, there is an urgent need for a novel detection model that integrates local details, global semantics, and interactive learning capabilities to adapt to the task of nodule detection in ultrasound images. Summary of the Invention

[0004] To address the shortcomings of existing technologies, a nodule detection method suitable for ultrasound images is proposed. A novel detection network integrating spatial adaptive modeling, multi-branch information enhancement, and multi-layer interactive decoding mechanisms is constructed to solve the problem of low detection accuracy caused by irregular nodule shapes and low contrast in existing technologies.

[0005] To achieve the above objectives, this invention provides the following technical solution: First, a backbone network integrating a spatial adaptive deformation enhancement module is used to extract features from input thyroid or breast ultrasound images, introducing deformable convolution to improve the perception of irregular edge regions; then, a feature enhancement module fusing global and local information is constructed to fuse the features to obtain enhanced multi-scale feature representations; finally, global context modeling and regression decoding are achieved through an encoder and an interactively learnable decoder. The process includes the following steps:

[0006] Step 1: A backbone network with a spatial adaptive deformation enhancement module is used to perform preliminary feature extraction on the input thyroid or breast ultrasound images. The spatial adaptive deformation enhancement module is designed as an FFN similar to the Transformer structure, with its core component being deformable convolution v4 (DCNv4). DCNv4 has adaptive receptive field, dynamic sampling capability, and spatial geometric deformation capability. It can dynamically learn the offset of sampling points based on input features, allowing the receptive field to adaptively fit irregularly shaped nodule regions, thereby enhancing the perception and feature extraction capabilities of irregular edges. The spatial adaptive deformation enhancement module utilizes the spatial geometric deformation learning capability of DCNv4, enabling it to more effectively adapt to the detection of irregularly shaped nodules.

[0007] Step 2: The multi-scale feature maps output by ResNet50 are processed by the Global-Local Feature Enhancement (GLFE) module, which fuses global and local information, to obtain more comprehensive multi-scale feature maps. Specifically, this module adopts a dual-branch parallel structure. The global information branch uses Fourier transform to convert the feature maps to the frequency domain. By processing the frequency domain information, the global context of the features is effectively captured. The local information branch uses large-kernel convolution to process the feature maps. Compared to traditional small-sized convolution kernels, large-kernel convolution has a larger receptive field, enabling it to capture richer local spatial information and texture details. The global and local enhanced features output from the dual-branch network are fused to generate a more comprehensive multi-scale feature map, which is then used as the input to the encoder.

[0008] Step 3: Input the enhanced feature map into a detector head containing an encoder with L layers of transformers and an interactively learnable decoder with L layers for global context modeling and regression decoding, outputting the nodule detection result. Specifically, the encoder consists of L standard Transformer Encoder Layers stacked together, responsible for global context modeling of the input feature map, outputting L levels of encoder features {E0, E1, ..., E...}. L-1 The interactive decoder consists of L multiple query decoder layers (MI Decoder Layers). Unlike traditional decoders where only the first layer interacts with the features of the last layer of the encoder, the MI decoder establishes a U-shaped cross-layer interaction mechanism. Specifically, for the i-th MI decoder layer (i = 1, 2, ..., L), its input object query Q... i-1 It will be combined with the j-th layer feature E from the encoder j Cross-attention is calculated, where j = L - i + 1. This interaction process is called U-shaped feature interaction.

[0009] Furthermore, the specific implementation process of integrating the spatial adaptive deformation enhancement module in step one is as follows:

[0010] Traditional convolution operations are limited by their fixed geometric sampling patterns, with their receptive fields confined to regular grid regions. In contrast, DCNv4 introduces adaptive receptive fields, dynamically aggregated weights, and unconstrained value ranges, enabling more flexible spatial modeling. Specifically, DCNv4 enhances its geometrically adaptive modeling in two key ways: first, by integrating learnable offsets into each sampling point of the convolution kernel; and second, by modulating the offsets and amplitudes of the sampled features to focus on the target region. The spatially adaptive deformation enhancement module not only retains the spatial adaptability of deformable convolution but also integrates an enhanced feedforward network (FFN) and a double normalization layer. This hybrid structure allows the spatially adaptive deformation enhancement module to effectively capture local deformations and global semantic dependencies. To improve the stability of the training process and the generalization ability of the model, this module also combines DropPath regularization with residual connection mechanisms to achieve more robust and efficient feature representation. The computation process of DCNv4 and the spatially adaptive deformation enhancement module is shown in Equation (1-4):

[0011]

[0012] Mix_FFN=Dropout(Conv(Dropout(GELU(DWConv(Conv(X))))))) (2)

[0013] X mid =Droppath(LN(DCNv4(X) in )))+X in (3)

[0014] X out =Droppath(LN(Mix_FFN(X) mid )))+X mid (4)

[0015] Where p0 is the input pixel, p k For the k-th predefined grid sampling point, Δp gk X is the offset corresponding to the k-th sampling point in the g-th group. in ∈R C×H×W The input feature map, m gk ∈R represents the spatial aggregation weight (also known as the modulation scalar) corresponding to the k-th sampling point in the g-th group, w g This represents the modulation weight of the g-th group. K is the total number of sampling points, and G is the number of spatial clusters.

[0016] Furthermore, the specific implementation process of the feature enhancement module that integrates global and local information between the backbone network and the encoder in step two is as follows:

[0017] The feature maps at different scales extracted by the backbone network are processed by a dedicated Global-Local Feature Enhancement (GLFE) module that fuses global and local information to obtain more comprehensive multi-scale feature maps. Specifically, this module is deployed between the backbone network and the subsequent encoder, employing a dual-branch parallel structure. The global information branch uses Fourier transform to convert the feature maps to the frequency domain. In the frequency domain, low-frequency components correspond to the global structural information of the image; by processing the frequency domain information, the global context of the features is effectively captured. The local information branch uses large-kernel convolution to process the feature maps. Compared to traditional small-sized convolution kernels, large-kernel convolution has a larger receptive field, enabling it to capture richer local spatial information and texture details. The goal is to aggregate the multi-scale feature maps generated by stages 2-4 of the backbone network, fusing high-resolution spatial details from the shallow layers of the backbone network with high-level semantic context from the deeper layers. First, channel-aligned feature maps P1, P2, and P3 are generated through channel projection processing. Simultaneously, the output of stage 4 is processed by 3×3 convolution and GroupNorm to generate P4. Then, these four feature maps are integrated in a top-down manner, where F1 comes directly from P1, and each subsequent Fi contains details from the previous fine-grained level. This design enriches high-level semantic features with fine-grained spatial details, while also incorporating multi-scale contextual representations. The specific formula is shown below:

[0018]

[0019] Among them, P i F is the feature map after channel projection processing. i The i-th level multi-scale feature map is generated after fusion, and GLFE is a feature enhancement module that fuses global and local information.

[0020] Furthermore, the specific implementation process of the encoder and interactively learnable decoder described in step three is as follows:

[0021] The enhanced feature map is input into an encoder containing L transformer layers and a detector head containing L interactively learnable decoder layers. Specifically, the encoder consists of L standard Transformer Encoder Layers stacked together, responsible for global context modeling of the input feature map and outputting L levels of encoder features {E0, E1, ..., E...}. L-1The interactive decoder consists of L multi-query decoder layers (MIDecoder Layer). Unlike traditional decoders where only the first layer interacts with the features of the last layer of the encoder, the MI decoder establishes a U-shaped cross-layer interaction mechanism. Specifically, for the i-th MI decoder layer (i = 1, 2, ..., L), its input object query Q... i-1 It will be combined with the j-th layer feature E from the encoder j Cross-attention is calculated, where j = L-i+1. This interaction process is called U-like Feature Interaction (ULFI). The specific formulas are shown in formulas (6) and (7):

[0022]

[0023] Q i =D i (Q i-1 E j j=L-i+1 (7)

[0024] Among them, E j For the j-th encoder layer, D i It is the decoder layer of the i-th layer.

[0025] The composite loss function used employs a combination of classification loss and regression loss to drive training. Focal loss is used to constrain the imbalance between the number of benign and malignant nodules to improve classification accuracy, while L1 loss and GIoU loss are used to constrain the detection task to further improve regression accuracy. The specific formula is shown below:

[0026]

[0027] L set =λ focal L focal +λ L1 L L1 +λ GIoU L GIoU (9)

[0028] Where P and G represent the prediction set and the truth set, respectively. λ represents the Hungarian matching algorithm used to match the prediction set with the truth set. focal ,λ L1 and λ GIoU The weights for focal loss, L1 loss, and GIoU loss are set to 2, 5, and 2, respectively.

[0029] The beneficial effects of this invention are as follows:

[0030] This invention applies deep learning methods to combine ultrasound image nodule detection and classification tasks. At the same time, it improves the network in terms of nodule detection accuracy and robustness. It uses a composite loss function that combines classification and regression for model training, which improves the model's detection performance for nodules of different types and sizes. This can assist doctors in making rapid, accurate and objective diagnoses, reduce the burden of manual image reading, and improve clinical diagnostic efficiency.

[0031] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic flowchart of a nodule detection method applicable to ultrasound images according to an embodiment of the present invention;

[0034] Figure 2 This is an example of the spatial adaptive deformation enhancement module of the present invention;

[0035] Figure 3 This is an example of the feature enhancement module that integrates global and local information in this invention;

[0036] Figure 4 This is an example of the encoder and interactively learnable decoder module of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are some embodiments of the present invention, but not all embodiments.

[0038] Figure 1 This is a schematic flowchart of a nodule detection method applicable to ultrasound images according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method consists of the following three parts.

[0039] A nodule detection method for ultrasound images first extracts primary features through a backbone network incorporating frequency domain channel attention; then, the features are fused by constructing a feature enhancement module that integrates global and local information to obtain enhanced multi-scale feature representations; finally, global context modeling and regression decoding are achieved through an encoder and an interactively learnable decoder.

[0040] A backbone network integrating a spatially adaptive deformation enhancement module is used to perform preliminary feature extraction on input thyroid or breast ultrasound images. This module is designed as a Transformer-like FFN, with its core component being a deformable convolutional network v4 (DCNv4). This module features an adaptive receptive field, dynamic sampling capability, and spatial geometric deformation capability. It can dynamically learn the offset of sampling points based on input features, allowing the receptive field to adaptively fit irregularly shaped nodule regions, thereby enhancing the perception and feature extraction capabilities of irregular edges. DCNv4 enhances its geometric adaptive modeling in two key ways: first, by integrating learnable offsets into each sampling point of the convolutional kernel; and second, by modulating the offset and amplitude of the sampling features to focus on the target region. The spatially adaptive deformation enhancement module utilizes the spatial geometric deformation learning capability of DCNv4, enabling it to more effectively adapt to the detection of irregularly shaped nodules. The Spatial Adaptive Deformation Enhancement Module not only retains the spatial adaptability of deformable convolutions but also integrates an enhanced feedforward network (FFN) and a double normalization layer. This hybrid structure allows the Spatial Adaptive Deformation Enhancement Module to effectively capture local deformations and global semantic dependencies. To improve the stability of the training process and the generalization ability of the model, this module also combines DropPath regularization and residual connection mechanisms to achieve more robust and efficient feature representation. This module performs standard 3×3 convolutional embeddings in each residual block of the ResNet50 backbone network, aiming to enhance the model's ability to extract discriminative features from irregularly shaped nodules. The calculation process of the Spatial Adaptive Deformation Enhancement Module is shown in Equation (1-4):

[0041]

[0042] Mix_FFN=Dropout(Conv(Dropout(GELU(DWConv(Conv(X))))))) (2)

[0043] X mid =Droppath(LN(DCNv4(X) in )))+X in (3)

[0044] X out=Droppath(LN(Mix_FFN(X) mid )))+X mid (4)

[0045] Where p0 is the input pixel, p k For the k-th predefined grid sampling point, Δp gk X is the offset corresponding to the k-th sampling point in the g-th group. in ∈R C×H×W The input feature map, m gk ∈R represents the spatial aggregation weight (also known as the modulation scalar) corresponding to the k-th sampling point in the g-th group, w g This represents the modulation weight of the g-th group. K is the total number of sampling points, and G is the number of spatial clusters.

[0046] A feature enhancement module that fuses global and local information is designed between the backbone network and the encoder. The multi-scale feature maps output by ResNet50 are processed by this module to obtain more comprehensive multi-scale feature maps. Specifically, this module is deployed between the backbone network and the subsequent encoder, employing a dual-branch parallel structure. The global information branch uses Fourier transform to convert the feature maps to the frequency domain. In the frequency domain, low-frequency components correspond to the global structural information of the image; processing this frequency domain information effectively captures the global context of the features. The local information branch uses large-kernel convolution to process the feature maps. Compared to traditional small-sized convolution kernels, large-kernel convolution has a larger receptive field, enabling the capture of richer local spatial information and texture details. The global and local enhanced features output from the dual-branch network are fused to generate a more comprehensive multi-scale feature map, which is then used as input to the encoder. The aim is to aggregate the multi-scale feature maps generated in stages 2-4 of the backbone network, fusing high-resolution spatial details from the shallow layers of the backbone network with high-level semantic context from the deeper layers. First, channel-aligned feature maps P1, P2, and P3 are generated through channel projection. Simultaneously, the output of stage 4 is processed by 3×3 convolution and GroupNorm to generate P4. Then, these four feature maps are integrated in a top-down manner, where F1 directly comes from P1, and each subsequent F1 contains details from the previous fine-grained level. This design enriches high-level semantic features with fine-grained spatial details while also incorporating multi-scale contextual representations. The specific formula is shown below:

[0047]

[0048] Among them, P i F is the feature map after channel projection processing. i The i-th level multi-scale feature map is generated after fusion, and GLFE is a feature enhancement module that fuses global and local information.

[0049] The encoder and interactively learnable decoder: The enhanced feature map is input into an encoder containing L transformer layers and a detector head containing an interactively learnable decoder of L layers. Specifically, the encoder consists of L standard Transformer Encoder Layers stacked together, responsible for global context modeling of the input feature map, and outputting L levels of encoder features {E0, F1, ..., E...}. L-1 The interactive decoder consists of L multi-query decoder layers (MI DecoderLayer). Unlike traditional decoders where only the first layer interacts with the features of the last layer of the encoder, the MI decoder establishes a U-shaped cross-layer interaction mechanism. Specifically, for the i-th MI decoder layer (i = 1, 2, ..., L), its input object query Q... i-1 It will be combined with the j-th layer feature E from the encoder j Cross-attention is calculated, where j = L-i+1. This interaction process is called U-like Feature Interaction (ULFI). The specific formulas are shown in formulas (6) and (7):

[0050]

[0051] Q i =D i (Q i-1 E j j=L-i+1 (7)

[0052] Among them, E j For the j-th encoder layer, D i It is the decoder layer of the i-th layer.

[0053] The composite loss function used employs a combination of classification loss and regression loss to drive training. Focal loss is used to constrain the imbalance between the number of benign and malignant nodules to improve classification accuracy, while L1 loss and GIoU loss are used to constrain the detection task to further improve regression accuracy. The specific formula is shown below:

[0054]

[0055] L set =λ focal L focal +λ L1 L L1 +λ GIoU L GIoU (9)

[0056] Where P and G represent the prediction set and the truth set, respectively. λ represents the Hungarian matching algorithm used to match the prediction set with the truth set. focal ,λ L1 and λ GIoU The weights for focal loss, L1 loss, and GIoU loss are set to 2, 5, and 2, respectively.

[0057] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for nodule detection using ultrasound images, comprising the following steps: Step 1: Extract features from the input ultrasound image to generate a multi-scale feature map; Step 2: Fuse the multi-scale feature maps to capture global contextual information and local spatial details, generating an enhanced multi-scale feature map; Step 3: Input the enhanced multi-scale feature map into the encoder and decoder for global context modeling and regression decoding, and output the nodule detection results.

2. The method according to claim 1, characterized in that, The feature extraction in step one is achieved through a backbone network integrating a spatial adaptive deformation enhancement module. The core component of this module is deformable convolution v4 (DCNv4), where: DCNv4 introduces adaptive receptive fields, dynamic aggregation weights, and unconstrained value ranges, enabling more flexible spatial modeling. The spatial adaptive deformation enhancement module leverages DCNv4's spatial geometric deformation learning capabilities to more effectively adapt to detecting irregularly shaped nodules. The calculation process of DCNv4 is shown in Equation 1: Where p0 is the input pixel, p k For the k-th predefined grid sampling point, Δp gk Let m be the offset corresponding to the k-th sampling point in the g-th group. gk ∈R represents the spatial aggregation weight (also known as the modulation scalar) corresponding to the k-th sampling point in the g-th group, w g This represents the modulation weight of the g-th group. K is the total number of sampling points, and G is the number of spatial clusters.

3. The method according to claim 2, characterized in that, The spatial adaptive deformation enhancement module is embedded in the residual block of the backbone network and is performed after a standard 3×3 convolution. Its calculation process includes: The spatial adaptive deformation enhancement module not only retains the spatial adaptability of deformable convolution v4, but also integrates an enhanced feedforward network (FFN) and a double normalization layer. This hybrid structure allows the spatial adaptive deformation enhancement module to effectively capture local deformations and global semantic dependencies. To improve the stability of the training process and the generalization ability of the model, this module also combines DropPath regularization and residual connection mechanisms to achieve a more robust and efficient feature representation. The calculation process of the spatial adaptive deformation enhancement module is shown in Equation (2-4): Mix_FFN=Dropout(Conv(Dropout(GELU(DWConv(Conv(X))))))) (2) X mid =Droppath(LN(DCNv4(X in )))+X in (3) X out =Droppath(LN(Mix_FFN(X mid )))+X mid (4) Among them, X in ∈R C×H×W X represents the input feature map. out ∈R C×H×W This represents the output feature map.

4. The method according to claim 1, characterized in that, The feature enhancement in step two is achieved through a feature enhancement module that fuses global and local information. This module employs a dual-branch parallel structure, wherein: The feature maps extracted from the backbone at different scales are processed by a dedicated Global-Local Feature Enhancement (GLFE) module that integrates global and local information to obtain more comprehensive multi-scale feature maps. Specifically, this module adopts a two-branch parallel structure. The global information branch uses Fourier transform to convert the feature maps to the frequency domain. By processing the frequency domain information, the global context of the features is effectively captured; the local information branch uses large-kernel convolution to process the feature maps. Compared with traditional small-sized convolution kernels, large-kernel convolution has a larger receptive field and can capture richer local spatial information and texture details.

5. The method according to claim 4, characterized in that, The feature enhancement module is used to aggregate multi-scale feature maps generated at different stages of the backbone network. The fusion process is as follows: This module is deployed between the backbone network and subsequent encoders. Its purpose is to aggregate the multi-scale feature maps generated by stages 2-4 of the backbone network, fusing high-resolution spatial details from the shallow layers of the backbone network with high-level semantic context from the deeper layers. First, channel-aligned feature maps P1, P2, and P3 are generated through channel projection processing. Simultaneously, the output of stage 4 is processed by 3×3 convolution and GroupNorm to generate P4. Then, these four feature maps are integrated in a top-down manner, where F1 comes directly from P1, and each subsequent F1 contains details from the previous fine-grained level. This design enriches high-level semantic features with fine-grained spatial details while also incorporating multi-scale contextual representations. The specific formula is shown below: Among them, P i F is the feature map after channel projection processing. i The i-th level multi-scale feature map is generated after fusion, and GLFE is a feature enhancement module that fuses global and local information.

6. The method according to claim 1, characterized in that, The decoder in step three is an interactively learnable decoder. It establishes a U-shaped cross-layer interaction mechanism, enabling cross-attention calculation between the i-th decoder layer and the j-th encoder layer features, where j = L - i + 1, and L is the total number of layers in the encoder and decoder. The enhanced feature map is input into an encoder containing L transformer layers and a detector head containing L interactively learnable decoder layers. Specifically, the encoder consists of L standard Transformer Encoder Layers stacked together, responsible for global context modeling of the input feature map and outputting L levels of encoder features {E0, E1, ..., E...}. L-1 The interactive decoder consists of L multi-query decoder layers (MIDecoder Layer). Unlike traditional decoders where only the first layer interacts with the last layer of the encoder, this decoder establishes a U-shaped cross-layer interaction mechanism. Specifically, for the i-th MI decoder layer (i = 1, 2, ..., L), its input object query Q... i-1 It will be combined with the j-th layer feature E from the encoder j Cross-attention is calculated, where j = L-i+1. This interaction process is called U-like Feature Interaction (ULFI). The specific formulas are shown in formulas (6) and (7): Q i =D i (Q i-1 ,E j )j=L-i+1 (7) Among them, E j For the j-th encoder layer, D i It is the decoder layer of the i-th layer.

7. The method according to claim 1, characterized in that, This method employs a composite loss function L, which combines classification and regression losses. loss The model is trained by using Focal loss to constrain the imbalance between the number of benign and malignant nodules to improve classification accuracy, and by using L1 loss and GIoU loss to constrain the detection task to further improve regression accuracy. L set =λ focal L focal +λ L1 L L1 +λ GIoU L GIoU (9) Where P and G represent the prediction set and the truth set, respectively. λ represents the Hungarian matching algorithm used to match the prediction set with the truth set. foal ,λ L1 and λ GIoU These are the weighting coefficients for focal loss, L1 loss, and GIoU loss.

Images