Classification of benign and malignant pulmonary nodule based on local-global hybrid network

Abstract

BACKGROUND:

The accurate classification of pulmonary nodules has great application value in assisting doctors in diagnosing conditions and meeting clinical needs. However, the complexity and heterogeneity of pulmonary nodules make it difficult to extract valuable characteristics of pulmonary nodules, so it is still challenging to achieve high-accuracy classification of pulmonary nodules.

OBJECTIVE:

In this paper, we propose a local-global hybrid network (LGHNet) to jointly model local and global information to improve the classification ability of benign and malignant pulmonary nodules.

METHODS:

First, we introduce the multi-scale local (MSL) block, which splits the input tensor into multiple channel groups, utilizing dilated convolutions with different dilation rates and efficient channel attention to extract fine-grained local information at different scales. Secondly, we design the hybrid attention (HA) block to capture long-range dependencies in spatial and channel dimensions to enhance the representation of global features.

RESULTS:

Experiments are carried out on the publicly available LIDC-IDRI and LUNGx datasets, and the accuracy, sensitivity, precision, specificity, and area under the curve (AUC) of the LIDC-IDRI dataset are 94.42%, 94.25%, 93.05%, 92.87%, and 97.26%, respectively. The AUC on the LUNGx dataset was 79.26%.

CONCLUSION:

The above classification results are superior to the state-of-the-art methods, indicating that the network has better classification performance and generalization ability.

Keywords

Pulmonary nodule classification Chest CT Hybrid network Attention mechanism

1 Introduction

According to the "Global Cancer Statistics 2020", the global incidence of lung cancer is 11.4%, ranking second among all cancers. Lung cancer causes 1.8 million deaths worldwide, accounting for 18% of the mortality rate, ranking first among all cancers [1]. Especially in China, lung cancer incidence and mortality rank first among all malignant tumors [2]. Lung cancer has become the leading cause of cancer-related deaths, posing a significant threat to human health. Relevant studies show that early lung cancer screening and treatment are crucial for improving patient survival rates and disease prognosis [3]. Low-dose computed tomography (LDCT) has emerged as an essential tool for early lung cancer screening due to its non-invasive nature, rapid imaging capabilities, and minimal radiation exposure. Early lung cancer diagnosis is primarily based on the analysis of lung nodules in CT images [4]. A pulmonary nodule is an abnormal area of the lung measuring less than 3 cm that appears denser than the surrounding tissue. While most pulmonary nodules are benign (non-cancerous), those with characteristics such as increased solidity, larger size, and irregular borders may have the potential to develop into malignant nodules and lead to lung cancer. In clinical, radiologists need to interpret many lung CT images daily, relying on their personal experience to diagnose lung nodules. However, this process is not only highly subjective but also time-consuming, which is prone to misdiagnosis or missed diagnosis. To improve radiologists’ diagnostic efficiency and accuracy, a computer-aided diagnostic (CAD) system has been developed for the diagnosis of early lung cancer screening [5]. The classification of benign and malignant pulmonary nodules is the last step in diagnosing pulmonary nodules and one of the critical components of the CAD system. Currently, two primary methods are employed for pulmonary nodule identification: handcraft-based methods and convolutional neural networks(CNNs)-based methods. The first method [6 –10] involves extracting features such as size, shape, density, texture, and others from pulmonary nodules using various feature descriptors like Gray-level Co-occurrence Matrix(GLCM)[11], Histogram of Oriented Gradient(HOG) [12], and Local Binary Pattern(LBP) [13]. Subsequently, it selects the most relevant features based on medical knowledge and experience, feeding these extracted nodule features into different algorithms [14 –16] of machine learning for classification. While handcraft-based methods have been widely employed in the past and have yielded favorable results in some cases, this approach heavily relies on individual domain knowledge and experience, making it susceptible to subjectivity. Additionally, handcrafted features may fail to capture certain complex structural or textural information, potentially limiting classifier performance.

In recent years, with the development of deep learning technology, CNNs have become the mainstream framework for classifying benign and malignant pulmonary nodules. Compared with handcraft-based methods, CNNs learn features and classification rules from original image data through end-to-end learning, which can automatically extract richer features from images. To enable the network to learn pulmonary nodule features at different scales, Shen et al. [17] and Xu et al. [18] cropped regions of interest in three different sizes from the original CT image and trained them into three networks to extract multi-scale and diverse pulmonary nodule features. Moreover, Sakshiwala et al. [19] proposed a new multi-scale (64 × 64, 32 × 32 and 16 × 16) CNN architecture and initialized the weights of the multi-scale architecture using transfer learning, achieving a classification accuracy of 93.88%. To fully utilize the spatial information of nodules, Xie et al.[20] further extracted CT image slices from nine fixed view angles of 3D pulmonary nodule CT images as input and combined multiple submodels based on knowledge collaboration to classify pulmonary nodules, obtaining richer nodule feature information. El-Regaily et al. [21] designed a multi-view CNN by extracting CT image slices from the axial, coronal, and sagittal of 3D pulmonary nodule CT images as input images. Wu et al. [22] proposed a multi-scale multi-view model based on ensemble attention for capturing more comprehensive discriminative nodule representations of nodules. Although the above multi-scale or multi-view methods achieve good classification performance, they require building multiple parallel networks for training, increasing the network’s complexity. The attention mechanism can enable networks to emphasize regions of interest while filtering out irrelevant information, making the network’s inference process interpretable. This mechanism has been widely applied in the field of pulmonary nodule classification. Jiang et al. [23] constructed a 3D dual-path network that combines contextual attention and spatial attention mechanisms to improve the representation of deep features and the robustness of predictions. Meanwhile, a multi-level feature fusion network with ResNeXt as the backbone and embedded channel attention mechanism was also used to classify benign and malignant pulmonary nodules [24].

CNNs can capture features through the convolution operation with a local receptive field, but they cannot directly model global information. To address this limitation, Wu et al. [25] and Sun et al. [26] used Swin Transformer [27], which can extract global information for pulmonary nodule classification. Since input pulmonary CT images contain both local and global features, relying solely on CNNs or transformers may result in the loss of valuable information and a reduction in network performance [28]. Given the complex morphology and varying sizes of pulmonary nodules, it is not only necessary to model local information to describe detailed features such as texture and edges of pulmonary nodules but also to capture global features to describe the shape or structure of pulmonary nodules [29]. This approach aligns with the clinical practice of doctors in diagnosing benign and malignant pulmonary nodules. Therefore, accurately capturing the local and global features of pulmonary nodules is crucial for the benign and malignant diagnosis of pulmonary nodules. In this paper, we construct an MSL block for extracting local spatial features at different scales to better represent the detailed features of pulmonary nodules. Additionally, an HA block can capture key information about the image globally, achieving better adaptability in both spatial and channel dimensions. By integrating local and global information, the classification performance of the network for pulmonary nodules is improved.

2 Methods

2.1 Overall architecture

To effectively capture the local and global features of pulmonary nodules, we propose a pulmonary nodule classification network with an architecture designed using a standard "four-stage" pyramid paradigm, as illustrated in Fig. 1. In this paper, we utilize 32×32 CT images of pulmonary nodules as input for LGHNet. The input image is initially processed through a stem module, comprising a 3×3 convolutional layer and layer normalization. This process yields feature maps with a resolution of 32×32 and 16 channels. Subsequently, the output feature maps enter the first stage of the network, which includes two Local blocks for extracting low-level semantic information. The structure of Local block is similar to ConvNeXt block [30], where a 3 × 3 depth-wise convolution with the number of groups equal to the number of channels and two 1 × 1 convolutions for channel mixing. In addition, there are Layer Normalization (LN) [31] and Gaussian Error Linear Unit (GELU) [32] also incorporated into the Local block. The network’s second, third, and fourth stages have the same configuration, consisting of an MSL block and an HA block. At the same time, we have designed an effective downsampling module composed of a max pooling layer and a 1 × 1 convolutional layer. The max pooling layer employs a 2 × 2 kernel with a stride of 2, which is used to reduce the spatial dimension of the feature map and preserve the most important features. The 1 × 1 convolutional layer increases the number of channels. This combination effectively enhances the network’s representation ability. Following the encoding of input images with semantic information across four stages, 4 × 4 ×128 feature maps are generated. Finally, the category probability is obtained through a sequence of global average pooling, a fully connected layer, and Softmax activation.

Fig. 1

The overall architecture of LGHNet for lung nodule classification. The detailed structures of MSL block and HA block are described in Sections 2.2 and 2.3.

2.2 MSL block

The MSL block enhances the network’s ability to perceive different scales and details by encoding multi-scale and local information, which helps process pulmonary nodules with different sizes and complexities, whose structure is shown in Fig. 2. The MSL block has three primary components: channel splitting, dilated convolution, and efficient channel attention (ECA).

Fig. 2

The diagram of our proposed MSL block.

Channel Splitting. Initially, the input image X with dimensions H × W × C is divided into r groups through channel splitting. Each group contains C/r channels while maintaining the spatial resolution at H × W. Notably, in the second and third stages, we divided C channels into three groups with r = 3. The first two groups perform convolution operations with receptive fields of 3 × 3 and 5 × 5, respectively, and the number of channels in these two groups matches $⌈ \frac{C}{3} ⌉$ . The third group preserves the original feature information of the input image, and its channel number is $⌊ \frac{C}{3} ⌋$ . In the fourth stage, we divide C channels into four groups. The feature maps from the first three groups are inputted into the depth-wise convolution with dilation rates of 1, 2, and 3, respectively, while the channel number of all four groups is C/r. Channel splitting allows each channel group to concentrate on capturing characteristics at different scales of pulmonary nodules. This process reduces redundant information and ultimately enhances the network’s representational capabilities.

Dilated Convolution. The diameter of pulmonary nodules varies from 3mm to 30mm, and the fixed-size convolution kernel is used to extract the features of pulmonary nodules, which easily cause information loss or redundancy. Therefore, in the MSL block, we use dilated convolution [33] with different dilation rates to extract local contextual semantic information of pulmonary nodules at various scales. Fig. 3 illustrates the variation of receptive field for 3x3 convolution with different dilation rates. It is worth noting that we combine dilated convolution with depth-wise convolution to further reduce the number of parameters and minimize memory requirements.

Fig. 3

Schematic diagram of dilated convolution with different dilation rates based on 3 × 3 convolution. (a) A 3 × 3 convolution. (b) A 3 × 3 convolution with a dilation rate of 2. (c) A 3 × 3 convolution with a dilation rate of 3.

ECA. Building upon the foundation of channel splitting and multi-scale feature extraction, various groups contain rich, local, and detailed information about pulmonary nodules. To capture the correlation between channels within each group and improve the network’s attention to important information on channels, we introduce ECA [34] into each branch, whose structure is shown in Fig. 4. The ECA module avoids the side effects of dimension reduction on network prediction and effectively captures inter-channel interactions. Input feature maps of size H × W × C after global average pooling, resulting in an aggregated feature vector of size 1 × 1 × C. Then, channel weights are generated by performing a 1D convolution of size k followed by a Sigmoid activation function, where k is adaptively determined based on the channel dimension C. Here, k represents the coverage range for local inter-channel interactions, indicating how many neurons are involved in the attention prediction for a channel. The 1 × 1 × C attention weight vector is element-wise multiplied with the input image to obtain the final feature map. The relationship between kernel size K and channel C is expressed as follows: $K = ψ (C) = {| \frac{{log}_{2} (C)}{γ} + \frac{b}{γ} |}_{odd} .$ (1) where |t|_odd represents the nearest odd number to t, γ and b are set to 2 and 1 respectively.

Fig. 4

The structure of ECA, taking a one-dimensional convolution with a kernel size of 3 (K = 3) as an example.

After the above operations, feature maps from different groups are concatenated along the channel dimension. Subsequently, Softmax and GELU activation functions, LN, and two 1x1 convolutions are applied to enhance the network’s non-linear transformation and feature representation ability.

2.3 HA block

Compared to CNN, the key component of the Transformer [35] is self-attention, which allows the network to automatically focus on information from different positions when processing input sequences, helping the network to model long-range dependencies and capture the global structure of the input sequence. We leverage the advantages of self-attention to construct an HA block for extracting global features, such as the shape and size of pulmonary nodules, with the aim of capturing obvious visual differences among pulmonary nodules. The structure of the HA block is shown in Fig. 5.

Fig. 5

The diagram of our proposed HA block.

Self-attention typically calculates the similarity score between the current pixel and other pixels in the spatial dimension, overlooking global modeling in the channel dimension. To address this limitation, the proposed HA block can simultaneously consider the relationships between channels and between different locations in spatial dimensions on a global scale. The input image $X \in ℝ^{H \times W \times C}$ is reshaped as $X \in ℝ^{HW \times C}$ after LN. Then X is multiplied with three weight matrices ( $W^{q}, W^{k}, W^{v} \in ℝ^{C \times C}$ ) to generate $Query (Q) \in ℝ^{HW \times C}$ , $Key (K) \in ℝ^{HW \times C}$ , and $Value (V) \in ℝ^{HW \times C}$ , where W^q, W^k, and W^v are learnable parameters. Particularly, the self-attention across the two dimensions shares the parameters of Q, K, and V to improve the network’s inference speed. In addition, in order to improve the concurrency and computational efficiency of the network, we use multi-head self-attention (MSA) with a setting of 4 heads.

Self-attention in spatial dimension. Q and K^T are multiplied to calculate similarity scores between pixels, and these similarity scores are then normalized using Softmax to obtain the attention weight matrix $A_{s} \in ℝ^{HW \times HW}$ . Pixels with higher weights are more influenced by other pixels. Subsequently, A_s is multiplied with V to perform a weighted average across all pixels, generating a new representation for each pixel. This new representation includes information about the relationships between pixels and the global contextual information. The above process can be expressed as follows: $O_{s} = Attention (Q, K, V) = Softmax (\frac{{QK}^{T}}{\sqrt{d_{k}}}) V$ (2) where $O_{s} \in ℝ^{HW \times C}$ represents the output result of self-attention in the spatial dimension, and d_k is the dimension of K.

Self-attention in channel dimension. This process is similar to the previous one, where Q^T and K are first multiplied to obtain the attention weight matrix $A_{c} \in ℝ^{C \times C}$ , and then V is multiplied by A_c to obtain the final feature map. This process can be expressed as follows: $O_{c} = Attention (Q, K, V) = VSoftmax (\frac{Q^{T} K}{\sqrt{d_{k}}})$ (3) where $O_{c} \in ℝ^{HW \times C}$ represents the output result of self-attention in channel dimension, and d_k is the dimension of K.

After feature mapping through 1 × 1 convolution, O_s and O_c are element-wise multiplied to encode more comprehensive global semantic information.

3 Datasets and Experimental settings

3.1 LIDC-IDRI and LUNGx dataset

The LIDC-IDRI dataset [36] was created through a collaboration between the Lung Image Database Consortium (LIDC) and the Image Database Resource Initiative (IDRI). This dataset contains 1,018 CT scans from 1,018 patients, each of which has been annotated by 1 to 4 professional radiologists. The annotation information is stored in extensible markup language (XML) files and includes features such as the location, diameter, shape, density, and malignancy level of nodules. Similar to the method [37], we only use nodules annotated by at least three radiologists and calculate the median of the malignancy levels annotated by all radiologists to categorize pulmonary nodules as benign or malignant. When the median malignancy level of a pulmonary nodule is less than 3, it is categorized as benign; when the median is greater than 3, the nodule is categorized as malignant. If the median malignancy level is equal to 3, it indicates that the radiologist is not sure whether the nodule is benign or malignant, and the nodule is excluded. In the end, we obtained a total of 848 CT images of pulmonary nodules, comprising 442 benign pulmonary nodules and 406 malignant pulmonary nodules.

The LUNGx [38] dataset was provided for the benign and malignant pulmonary nodule classification challenge at the SPIE Medical Imaging Conference in 2015. The dataset contains 10 CT scans for calibration and 60 CT scans for testing. The case names, coordinates of the approximate nodule centroid, and diagnosis results are stored in an associated Excel file. In this paper, since this dataset provides a small number of CT scans, the calibration and test data are mixed to form a larger test dataset containing a total of 83 pulmonary nodules (42 benign and 41 malignant).

3.2 Data processing

We used LIDC-IDRI as training and test data, while LUNGx was only used as external test data to evaluate the generalization ability of the network. Since the CT images come from different institutions and devices, it is necessary to linearly interpolate and resample CT images to ensure a uniform spacing of 1mm for each dimension. Subsequently, we apply a window range of [-1000, 400] to filter out air and bone regions and normalize the data using Z-score standardization, as follows: $x = \min (\max (x, - 1000), 400)$ (4) $x = \frac{x - μ}{σ}$ (5) where μ is the mean of sample data, σ is the standard deviation of sample data. Both datasets adopt the preprocessing methods mentioned above.

The following preprocessing steps are applied to LIDC-IDRI dataset: (1) Crop CT images volume of size 32mm × 32mm × 32mm centered around the annotated position of the pulmonary nodule to remove a large amount of interference information. (2) Extract 2D slices from axial, coronal, and sagittal perspectives to obtain rich spatial information from CT images. (3) Apply data augmentation operations for each 2D slice to prevent overfitting of the network and improve its generalization ability. We use two augmentation methods. The schematic diagram of the two data augmentation methods is shown in Fig. 6. One is the conventional data augmentation strategy, which includes operations such as horizontal flipping, rotations (90⁰, 180⁰, 270⁰), and the addition of Gaussian blur; the other is the MedAugment [39] method, which uses a sampling strategy to select data augmentation operations from an augmentation space (pixel augmentation space, spatial augmentation space). The selected operations are then randomly ordered and executed sequentially. Through these two augmentation methods, the training dataset is increased by ten times. (4) Divide the dataset according to the ten-fold cross-validation method, where nine folds are used for training, and the remaining one fold is used for testing.

Fig. 6

Different data augmentation methods. (a) Conventional data augmentation method. (b) MedAugment method. The green rectangular box represents operations in the pixel augmentation space. The light orange rectangular box represents operations in the spatial augmentation space. ’Pass’ represents no augmentation in the augmentation space.

For the LUNGx dataset, we only need to crop CT image patches of size 32mm × 32mm based on the nodule coordinates provided in the Excel file.

3.3 Evaluation metrics

We evaluate the network’s classification performance using five criteria: accuracy(Acc), sensitivity (Sen), precision(Pre), specificity(Spe), and AUC. Here are the expressions for some of the metrics: $Acc = \frac{N_{TP} + N_{TN}}{N_{TP} + N_{TN} + N_{FP} + N_{FN}}$ (6) $Sen = \frac{N_{TP}}{N_{TP} + N_{FN}}$ (7) $Pre = \frac{N_{TP}}{N_{TP} + N_{FP}}$ (8) $Spe = \frac{N_{TN}}{N_{TN} + N_{FP}}$ (9) where N_TP represents the number of true positive nodules, N_TN represents the number of true negative nodules, N_FP represents the number of false positive nodules, and N_FN represents the number of false negative nodules.

3.4 Experimental settings

We used Python3.9 and PyTorch1.12 framework for programming. The network was trained and tested on 32GB memory, NVIDIA GeForce RTX 3090Ti, Intel Core i5-12600KF 3.70GHz CPU processor, and Windows 10 system. We have chosen the cross-entropy loss function as our loss function. Additionally, experiments indicated that the network reached a converged state after 50 training epochs. Therefore, in this paper, we set the training epochs to 50. For other detailed parameter settings, please refer to Table 1.

Table 1
Training parameter setting

Training parameter Parameter value

Batch size 64

Epochs 50

Optimizer AdamW

Learning rate 0.006

Learning rate schedule Cosine annealing

Weight decay 0.05

Training parameter	Parameter value
Batch size	64
Epochs	50
Optimizer	AdamW
Learning rate	0.006
Learning rate schedule	Cosine annealing
Weight decay	0.05

4 Results and discussion

4.1 Comparison with some state-of-the-art methods

Table 2 shows the comparison results of the proposed method with other state-of-the-art methods. To ensure fairness, all compared methods utilized the LIDC-IDRI dataset for both training and testing. We categorized pulmonary nodule classification methods into three groups: CNN-based classification methods, Transformer-based classification methods, and hybrid classification methods combining CNN and Transformer architectures. Liu et al. [29] designed a pulmonary nodules classification network that incorporated residual and transformer blocks. However, the designed network is too shallow and cannot effectively capture the hierarchical structure and high-level features of pulmonary nodules. Consequently, it has limitations in representing complex features, leading to lower classification performance compared to our method. Similarly, Al-Shabi et al. [40] used residual blocks and Non-Local [41] blocks to construct classification networks, but maintaining consistent feature map resolution can lead to feature duplication and redundancy, thereby negatively affecting classification performance. Additionally, Cao et al. [42] employed PixelShuffle [43] to first reconstruct the 32 × 32 pulmonary nodule CT images into 64 × 64 before feature extraction and classification. However, during the pixel shuffle process, some fine details may be lost, potentially increasing computational resources and time usage. While this method achieves slightly higher precision than our proposed approach (by 0.73%), other metrics in [42] are lower than those of our method. Wang et al. [44] proposed an improved Vision Transformer (ViT) [45] pulmonary nodule classification network, but due to the lack of CNN’s inherent inductive bias, its performance on classification tasks is reduced. In contrast, the Local block and MSL block in our method can extract rotation and translation-invariant local features, compensating for the shortcomings of the Transformer structure in modeling local spatial structures. Jiang et al. [46] used neural architecture search to construct a pulmonary nodule classification network automatically. Although the specificity is 2.17% higher than that of the proposed method, the sensitivity is much lower than our method’s 94.25%, indicating that the network is more likely to identify malignant nodules as benign, which increases the risk of patients missing the best treatment opportunity in clinical practice. At the same time, our proposed method outperforms other methods across various metrics, reflecting that LGHNet has stronger feature extraction abilities and can better classify benign and malignant pulmonary nodules.

Table 2
Comparison with state-of-the-art methods on the LIDC-IDRI dataset–label tab2

Categories Methods Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

CNN Zhang et al. [47] 87.5 94.1 - 81.7 90.1

Al-Shabi et al. [40] 88.46 88.66 87.38 - 95.62

Lyu et al. [48] 92.19 92.10 - 91.50 97.05

Jiang et al. [46] 90.77 85.37 - 95.04 -

Al-Shabi et al. [49] 92.81 92.36 92.59 - 96.17

Wu et al. [22] 91.71 90.98 - 92.40 96.42

Huang et al. [50] 91.07 90.93 91.18 91.22 95.84

Sakshiwala et al. [19] 93.88 93.36 - 93.26 93.31

Wu et al. [51] 92.36 91.62 92.99 93.08 97.17

Transformer Wang et al. [44] 93.33 - - - 96.04

Hybrid Liu et al. [29] 92.92 93.84 91.62 - 96.28

Cao et al. [42] 94.04 92.50 93.78 - 96.36

LGHNet (Ours) 94.42 94.25 93.05 92.87 97.26

Categories	Methods	Acc(%)	Sen(%)	Pre(%)	Spe(%)	AUC(%)
CNN	Zhang et al. [47]	87.5	94.1	-	81.7	90.1
	Al-Shabi et al. [40]	88.46	88.66	87.38	-	95.62
	Lyu et al. [48]	92.19	92.10	-	91.50	97.05
	Jiang et al. [46]	90.77	85.37	-	95.04	-
	Al-Shabi et al. [49]	92.81	92.36	92.59	-	96.17
	Wu et al. [22]	91.71	90.98	-	92.40	96.42
	Huang et al. [50]	91.07	90.93	91.18	91.22	95.84
	Sakshiwala et al. [19]	93.88	93.36	-	93.26	93.31
	Wu et al. [51]	92.36	91.62	92.99	93.08	97.17
Transformer	Wang et al. [44]	93.33	-	-	-	96.04
Hybrid	Liu et al. [29]	92.92	93.84	91.62	-	96.28
	Cao et al. [42]	94.04	92.50	93.78	-	96.36
	LGHNet (Ours)	94.42	94.25	93.05	92.87	97.26

We used the LUNGx dataset to evaluate the generalization capability of LGHNet. The best weight for each fold trained on the LIDC-IDRI dataset is selected and then tested using the LUNGx dataset. The results compared with state-of-the-art methods are shown in Table 3. Except for a 1.7% lower specificity than reference [20], other evaluation metrics of LGHNet on the LUNGx dataset are better than those of reference [20] and [52], reflecting that the proposed method has stronger generalization ability. In particular, LGHNet achieved a sensitivity of 90.30% on the LUNGx dataset, which is 3.08% higher than reference [20]. This indicates that LGHNet has significant advantage in accurately identifying malignant nodules, thereby reducing the risk of missed diagnosis.

Table 3

Comparison with state-of-the-art methods on the LUNGx dataset–label tab3

Methods	Acc(%)	Sen(%)	Pre(%)	Spe(%)	AUC(%)
Xie et al. [20]	75.62	87.22	70.63	64.32	76.85
Liao et al. [52]	-	-	-	-	78.08
LGHNet (Ours)	78.31	90.39	71.22	62.62	79.26

4.2 Comparison with different models

To validate the superiority of LGHNet in the pulmonary nodule classification task, we compared it to six classical pre-trained models commonly used in natural image classification. These models include ResNet18, ResNet34, ResNet50, ResNet101, DenseNet121, and DenseNet161. Subsequently, we fine-tune these six models on the LIDC-IDRI dataset for training and testing. Since these six models require three-channel images as input, we replicate the single-channel pulmonary nodule CT image three times and concatenate them in the channel dimension, resulting in 32 × 32 × 3 size images. Furthermore, we freeze the convolutional layers of the models and only modify the output feature dimension of the final fully connected layer to 2 for the binary classification task of pulmonary nodules. The comparison results of the classification metrics of each model are shown in Fig. 7. Table 4 shows the comparison results of the number of parameters, floating-point operations (FLOPs), and inference time of different models.

Fig. 7

Comparison of classification metrics of different models.

Table 4

Comparison of parameter number, FLOPs and inference time of different models–label tab4

Model	Parameters/M	FLOPs/G	Inference time/ms
ResNet18	11.18	0.04	8.80
ResNet34	21.29	0.08	10.03
ResNet50	23.51	0.08	10.38
ResNet101	42.50	0.16	13.43
DenseNet121	6.96	0.05	11.48
DenseNet161	26.48	0.16	19.65
LGHNet (Ours)	0.45	0.02	9.40

From the comprehensive results shown in Fig. 7 and Table 4, it is evident that, compared to the six models, the proposed method achieves higher classification performance with fewer parameters in the task of benign and malignant pulmonary nodule classification. Notably, it demonstrates significant advantages in terms of accuracy, sensitivity, and precision. At the same time, the FLOPs of LGHNet is only 0.02G, which is much lower than the FLOPs of other models, indicating that the proposed method is lightweight and requires fewer computing resources in actual deployment applications. Although the inference time of LGHNet is longer than that of ResNet18, LGHNet is more competitive than ResNet18 in terms of the number of parameters and FLOPs of the model.

4.3 Different data augmentation methods

The classification results of different data augmentation methods is presented in Table 5. As can be seen from Table 5, compared with the traditional data augmentation method, the accuracy on the test dataset increased by 0.74% when using data augmented by the MedAugment [39] method for network training. The MedAugment method excludes operations that are not suitable for medical images, such as invert, equalize, and solarize operations that may destroy details and features in medical images. Furthermore, using sampling strategy and hyperparameter mapping to make the augmented image closer to the real pulmonary nodule CT image. This may be one of the reasons why the MedAugment method outperforms the conventional data augmentation method. In this paper, we adopt two data augmentation methods and merge the data after the two data augmentation methods to create a larger training dataset, which enables the network to learn more diverse and rich pulmonary nodule features and have better adaptability and generalization ability for complex pulmonary nodule CT images. As a result, the network’s classification performance has significantly improved.

Table 5
Classification results with the different data augmentation methods–label tab5

Augmentation methods Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

Conventional 91.51 91.84 90.79 91.2 96.02

MedAugment 92.25 91.97 91.61 92.09 96.24

Conventional and MedAugment 94.42 94.25 93.05 92.87 97.26

Augmentation methods	Acc(%)	Sen(%)	Pre(%)	Spe(%)	AUC(%)
Conventional	91.51	91.84	90.79	91.2	96.02
MedAugment	92.25	91.97	91.61	92.09	96.24
Conventional and MedAugment	94.42	94.25	93.05	92.87	97.26

4.4 Ablation experiment

The impact of each block of the network and the structure of different networks on classification performance were discussed. The results of the ablation experiment are shown in Table 6. We compare the classification performance of the following networks:

N1: The four stages of the network are composed of Local blocks, which serve as the baseline network for comparison.

N2: Based on N1, replace the second block of each stage with an HA block.

N3: Replace the Local block of N2 with the MSL block. The MSL blocks in the first, second, and third stages all divide the channel into three groups.

N4: Based on N1, only add HA blocks in the second, third, and fourth stages.

The baseline model N1 achieved classification results with an accuracy of 91.51%, sensitivity of 91.12%, precision of 90.06%, specificity of 90.51%, and an AUC of 95.95%. Due to the addition of HA block in N2, the global features of pulmonary nodules can be captured in both spatial and channel dimensions, which improves the recognition ability of the network. Compared with N1, all classification metrics of N2 have improved. Based on N2, N3 adds the MSL block, and the classification ability of the network has declined, especially the sensitivity, which is only 89.66%. In addition, N4 removes the HA block in the first stage, and its classification metrics are better than N2. The reason may be that the shallow network is more suitable for modeling local features to capture detailed information on pulmonary nodules. Finally, we replaced the Local block in the second, third, and fourth stages of N4 with the MSL block to form the overall structure of LGHNet. Compared with N4, the classification accuracy of LGHNet increased by 1.9%.

Table 6
The results of ablation experiment

Networks Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

N1 91.51 91.12 90.06 90.51 95.95

N2 91.85 92.45 91.43 91.86 96.18

N3 91.39 89.66 92.29 92.98 96.02

N4 92.45 92.6 91.79 92.09 96.56

LGHNet (Ours) 94.42 94.25 93.05 92.87 97.26

Networks	Acc(%)	Sen(%)	Pre(%)	Spe(%)	AUC(%)
N1	91.51	91.12	90.06	90.51	95.95
N2	91.85	92.45	91.43	91.86	96.18
N3	91.39	89.66	92.29	92.98	96.02
N4	92.45	92.6	91.79	92.09	96.56
LGHNet (Ours)	94.42	94.25	93.05	92.87	97.26

4.5 Self-attention selection and fusion methods in HA block

In LGHNet, we introduce the HA block that performs self-attention in the spatial dimension and channel dimension to capture the global contextual information of pulmonary nodules from multiple dimensions. In this section, we explore the effects of different self-attention choices and fusion methods on classification performance. There are mainly the following methods:

M1: Select only self-attention in the spatial dimension.

M2: Select only self-attention in the channel dimension.

M3: Perform self-attention in the spatial dimension first and in the channel dimension.

M4: Perform self-attention in the channel dimension first and in the spatial dimension.

M5: Self-attention in the spatial dimension and self-attention in the channel dimension share Q, K, and V, but they are fused by element-wise addition.

M6: Self-attention in the spatial dimension and self-attention in the channel dimension share Q, K, and V, but they are fused by element-wise multiplication, which is the method used in the HA block.

The comparison results of the above different methods are shown in Fig. 8. Firstly, it can be observed that compared to the method using hybrid attention, the classification performance is worse when self-attention is applied separately in either the spatial or channel dimension, with a classification accuracy of only 91.99% and 91.75%, respectively. M3 and M4 sequentially execute self-attention in two dimensions. Although each metric is better than M1 and M2, this method cannot achieve shared parameters of Q, K, and V, which, to some extent, increases the number of parameters of the network and consumes more memory resources. M5 fuses the results of self-attention in two dimensions by element-wise addition. Compared with M3 and M4, the accuracy and sensitivity of M5 are both improved, reaching 93.4% and 93.2%, respectively. However, the precision is only 91.72%, indicating an imbalance between sensitivity and accuracy, where the network tends to predict most samples as malignant. M6 shows a significant improvement in various metrics. Fusion through element-wise multiplication helps capture more complex feature relationships, reduces conflicts between unrelated features, and takes into account both spatial relationships between pixels and feature correlations between channels, thus enhancing the network’s classification performance.

Fig. 8

Comparison of different self-attention selection and fusion methods.

4.6 The setting for the number of heads in MSA

MSA maps input features into multiple heads, each of which learns to focus on a different part of the input. However, setting too many heads may make it difficult for the network to concentrate on learning key features and reduce network performance. Therefore, it is necessary to select an appropriate number of heads for each HA block. In addition, according to the overall structure of LGHNet, we only consider the setting of the number of heads in the second, third, and fourth stages. The results are shown in Table 7. When the number of heads in all three stages is set to 2, the classification accuracy, sensitivity, precision, specificity, and AUC are 91.87%, 91.85%, 91.5%, 91.88%, and 96.10%, respectively. When the number of heads in each stage increases to 4, the network’s classification performance reaches its optimum.

Table 7
Classification results with the different number of head settings

Number of heads configuration Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

{0, 2, 2, 2} 91.87 91.85 91.5 91.88 96.10

{0, 2, 2, 4} 91.98 92.1 91.47 91.87 96.12

{0, 2, 4, 4} 92.75 92.87 91.2 91.64 96.34

{0, 4, 4, 4} 94.42 94.25 93.05 92.87 97.26

Number of heads configuration	Acc(%)	Sen(%)	Pre(%)	Spe(%)	AUC(%)
{0, 2, 2, 2}	91.87	91.85	91.5	91.88	96.10
{0, 2, 2, 4}	91.98	92.1	91.47	91.87	96.12
{0, 2, 4, 4}	92.75	92.87	91.2	91.64	96.34
{0, 4, 4, 4}	94.42	94.25	93.05	92.87	97.26

4.7 The setting for the number of branches in MSL block

In this section, we discuss the impact of different numbers of branches in the MSL block on classification performance. The results are presented in Table 8. When the number of branches of the MSL block in the second, third, and fourth stages is set to 3, 3, and 4, respectively, LGHNet has the best classification performance. However, when the number of branches in the second stage is set to 3, and the number of branches in the third and fourth stages is set to 4, accuracy and precision are the lowest, which are 91.16% and 90.6%, respectively. The other two configurations also performed poorly in the classification of the LIDC-IDRI dataset.

Table 8
Classification results with the different number of branch settings

Number of branches configuration Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

{0, 3, 3, 3} 91.86 92.35 91.00 91.41 96.12

{0, 3, 4, 4} 91.16 91.14 90.6 91.19 95.85

{0, 4, 4, 4} 91.98 90.87 92.45 92.99 96.13

{0, 3, 3, 4} 94.42 94.25 93.05 92.87 97.26

Number of branches configuration	Acc(%)	Sen(%)	Pre(%)	Spe(%)	AUC(%)
{0, 3, 3, 3}	91.86	92.35	91.00	91.41	96.12
{0, 3, 4, 4}	91.16	91.14	90.6	91.19	95.85
{0, 4, 4, 4}	91.98	90.87	92.45	92.99	96.13
{0, 3, 3, 4}	94.42	94.25	93.05	92.87	97.26

4.8 Analysis of visualization results

To gain insights into the network’s decision-making process, we employ the Grad-CAM algorithm [53] to generate heat maps that reveal the regions of interest for the network. The more interested the network is in certain areas, the darker the color in the heat map. Fig. 9 shows the Grad-CAM visualization and classification results of benign and malignant pulmonary nodules.

Fig. 9

Grad-CAM visualization and classification results of benign and malignant nodules.

As depicted in Fig. 9, the high heat values are predominantly concentrated at the center of pulmonary nodules and gradually extend towards the boundaries, which indicates that LGHNet can accurately locate the position of pulmonary nodules in CT images and extract local and global features of pulmonary nodules to determine the benign and malignant nodules. The diameters of the nodules in the figure ranged from 4mm to 13mm. However, LGHNet can correctly identify the categories of nodules with high accuracy, which, to a certain extent, shows that our proposed MSL block effectively addresses the multi-scale issue of pulmonary nodules, demonstrating adaptability to nodules of different sizes. At the same time, LGHNet still has a robust discriminative ability for some nodules with complex morphology, such as nodules ’C,’ ’E,’ ’F,’ as well as nodules ’B,’ ’G,’ and ’L,’ which are located around the trunk, reflecting the high generalization and robustness of the proposed method for nodules with different shapes and locations, which can further improve the reliability of LGHNet for practical clinical applications.

5 Conclusion

The LGHNet we proposed effectively captures local and global information of lung nodule CT images through the cooperation of the MSL block and the HA block, improving the classification accuracy and robustness of the network, thereby strengthening its diagnostic capability for potential lung diseases. However, due to the limited availability of CT images of pulmonary nodules, overfitting issues may arise during network training. Therefore, in future work, we will explore methods related to image generation for producing high-quality images to enrich training datasets.

References

Sung

, Ferlay

, Siegel

R.L.

et al. Global cancer statistics: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J], CA: A Cancer Journal for Clinicians 71(3) (2021), 209–249.

Zheng

, Zhang

, Zeng

et al. Cancer incidence and mortality in China, [J], Journal of the National Cancer Center 2(1) (2022), 1–9.

Tanoue

L.T.

, Tanner

N.T.

, Gould

M.K.

et al. Lung cancer screening[J], American Journal of Respiratory and Critical Care Medicine 191(1) (2015), 19–33.

Guo

, Yang

, Zhao

et al. 3D SAACNet with GBM for the classification of benign and malignant lung nodules[J], Computers in Biology and Medicine 153 (2023), 106532.

Yuan

, Fan

, Wu

et al. An efficient multi-path 3D convolutional neural network for false-positive reduction of pulmonary nodule detection[J], International Journal of Computer Assisted Radiology and Surgery 16(12) (2021), 2269–2277.

Chen

C.H.

, Chang

C.K.

, Tu

C.Y.

et al. Radiomic features analysis in computed tomography images of lung nodule classification[J], PloS one 13(2) (2018), e0192002.

Farag

A.A.

, Ali

, Elshazly

et al. Feature fusion for lung nodule classification[J], International Journal of Computer Assisted Radiology and Surgery 12 (2017), 1809–1818.

Ferreira

J.R.

, Oliveira

M.C.

, de Azevedo-Marques

P.M.

, Characterization of pulmonary nodules based on features of margin sharpness and texture[J], Journal of Digital Imaging 31 (2018), 451–463.

X.X.

, Li

, Tian

L.F.

et al. Automatic benign and malignant classification of pulmonary nodules in thoracic computed tomography based on RF algorithm[J], IET Image Processing 12(7) (2018), 1253–1264.

10.

Calheiros

J.L.L.

, de Amorim

L.B.V.

, de Lima

L.L.

et al. The effects of perinodular features on solid lung nodule classification[J], Journal of Digital Imaging (2021), 1–13.

11.

Haralick

R.M.

, Shanmugam

, Dinstein

I.H.

, Textural features for image classification[J], IEEE Transactions on Systems, Man, and Cybernetics 1973(6), 610–621.

12.

Dalal

, Triggs

, Histograms of oriented gradients for human detection[C]//, 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05). Ieee 1 (2005), 886–893.

13.

Ojala

, Pietikainen

, Maenpaa

, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns[J], IEEE Transactions on Pattern Analysis and Machine Intelligence 24(7) (2002), 971–987.

14.

Guo

, Wang

, Bell

et al. KNN model-based approach in classification[C]//On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE: OTM Confederated International Conferences, CoopIS, DOA, and ODBASE 2003, Catania, Sicily, Italy, November 3-7, 2003. Proceedings. Springer Berlin Heidelberg, (2003), 986–996.

15.

Quinlan

J.R.

, Learning decision tree classifiers[J], ACM Computing Surveys (CSUR) 28(1) (1996), 71–72.

16.

Dietterich

T.G.

, Ensemble learning[J], The Handbook of Brain Theory and Neural Networks 2(1) (2002), 110–125.

17.

Shen

, Zhou

, Yang

et al. Multi-scale convolutional neural networks for lung nodule classification[C]//Information Processing in Medical Imaging: 24th International Conference, IPMI 2015, Sabhal Mor Ostaig, Isle of Skye, UK, June 28-July 3, 2015, Proceedings 24. Springer International Publishing, (2015), 588–599.

18.

, Wang

, Guo

et al. MSCS-DeepLN: Evaluating lung nodule malignancy using multi-scale cost-sensitive neural networks[J], Medical Image Analysis 65 (2020), 101772.

19.

Singh Sakshiwala

M.P.

, A new framework for multi-scale CNN-based malignancy classification of pulmonary lung nodules[J], Journal of Ambient Intelligence and Humanized Computing 14(5) (2023), 4675–4683.

20.

Xie

, Xia

, Zhang

et al. Knowledge-based collaborative deep learning for benignmalignant lung nodule classification on chest CT[J], IEEE Transactions on Medical Imaging 38(4) (2018), 991–1004.

21.

El-Regaily

S.A.

, Salem

M.A.M.

, Aziz

M.H.A.

et al. Multi-view Convolutional Neural Network for lung nodule false positive reduction[J], Expert Systems with Applications 162 (2020), 113017.

22.

, Huang

Multi-Scale Multi-View Model Based on Ensemble Attention for Benign-Malignant Lung Nodule Classification on Chest CT[C]//2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI). IEEE, (2022), 1–6.

23.

Jiang

, Gao

, Xu

et al. Attentive and ensemble 3D dual path networks for pulmonary nodules classification[J], Neurocomputing 398 (2020), 422–430.

24.

Huang

Y.S.

, Wang

T.C.

, Huang

S.Z.

et al. An improved 3-D attention CNN with hybrid loss and feature fusion for pulmonary nodule classification[J], Computer Methods and Programs in Biomedicine 229 (2023), 107278.

25.

, Chen

, Wu

, Swin transformer based benign and malignant pulmonary nodule classification[C]//5th International Conference on Computer Information Science and Application Technology (CISAT 2022), SPIE 12451 (2022), 552–558.

26.

Sun

, Pang

, Li

, Efficient Lung Cancer Image Classification and Segmentation Algorithm Based on an Improved Swin Transformer[J], Electronics 12(4) (2023), 1024.

27.

Liu

, Lin

, Cao

et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]//Proceedings of the IEEE/CVF international conference on computer vision (2021), 10012–10022.

28.

Liu

, Shen

CECT: Controllable Ensemble CNN and Transformer for COVID-19 image classification by capturing both local and global image features[J]. arXiv preprint arXiv:2302.02314, 2023.

29.

Liu

, Liu

, Tie

et al. Res-trans networks for lung nodule classification[J], International Journal of Computer Assisted Radiology and Surgery 17(6) (2022), 1059–1068.

30.

Liu

, Mao

, Wu

C.Y.

et al. A convnet for the 2020s[C]// Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022), 11976–11986.

31.

J.L.

, Kiros

J.R.

, Hinton

G.E.

Layer normalization[J], arXiv preprint arXiv:1607.06450, 2016.

32.

Hendrycks

, Gimpel

Gaussian error linear units (gelus)[J], arXiv preprint arXiv:1606.08415, 2016.

33.

, Koltun

Multi-scale context aggregation by dilated convolutions[J], arXiv preprint arXiv:1511.07122, 2015.

34.

Wang

, Wu

, Zhu

et al. ECA-Net: Efficient channel attention for deep convolutional neural networks[C]// Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020), 11534–11542.

35.

Vaswani

, Shazeer

, Parmar

et al. Attention is all you need[J], Advances in neural information processing systems (2017), 30.

36.

Armato

S.G.

III McLennan

, Bidaut

et al. The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans[J], Medical Physics 38(2) (2011), 915–931.

37.

Al-Shabi

, Shak

, Tan

, ProCAN: Progressive growing channel attentive non-local network for lung nodule classification[J], Pattern Recognition 122 (2022), 108309.

38.

Armato

S.G.

III Drukker

, Li

et al. LUNGx Challenge for computerized lung nodule classification[J], Journal of Medical Imaging 3(4) (2016), 044506–044506.

39.

Liu

, Lv

, Li

et al. Medaugment: universal automatic data augmentation plug-in for medical image analysis[J], arXiv preprint arXiv:2306.17466, 2023.

40.

Al-Shabi

, Lan

B.L.

, Chan

W.Y.

et al. Lung nodule classification using deep local-global networks[J], International Journal of Computer Assisted Radiology and Surgery 14 (2019), 1815–1819.

41.

Wang

, Girshick

, Gupta

et al. Non-local neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition (2018), 7794–7803.

42.

Cao

, Tao

, Wang

et al. MSM-ViT: A multi-scale MobileViT for pulmonary nodule classification using CT images[J], Journal of X-Ray Science and Technology (2023), (Preprint): 1–14.

43.

Shi

, Caballero

, Husz?ar

et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network[C]// Proceedings of the IEEE conference on computer vision and pattern recognition (2016), 1874–1883.

44.

Wang

, Zhang

, Yang

TransPND: A Transformer Based Pulmonary Nodule DiagnosisMethod on CT Image[C]// Chinese Conference on Pattern Recognition and Computer Vision (PRCV). Cham: Springer Nature Switzerland (2022), 348–360.

45.

Dosovitskiy

, Beyer

, Kolesnikov

et al. An image is worth 16x16 words: Transformers for image recognition at scale[J], arXiv preprint arXiv:2010.11929, 2020.

46.

Jiang

, Shen

, Gao

et al. Learning efficient, explainable and discriminative representations for pulmonary nodules classification[J], Pattern Recognition 113 (2021), 107825.

47.

Zhang

, Zhu

, Liu

et al. Multi-scale pulmonary nodule classification with deep feature fusion via residual network[J], Journal of Ambient Intelligence and Humanized Computing (2018), 1–12.

48.

Lyu

, Bi

, Ling

S.H.

, Multi-level cross residual network for lung nodule classification[J], Sensors 20(10) (2020), 2837.

49.

Al-Shabi

, Shak

, Tan

, 3D axial-attention for lung nodule classification[J], International Journal of Computer Assisted Radiology and Surgery 16(8) (2021), 1319–1324.

50.

Huang

, Wu

, Li

et al. Self-supervised transfer learning based on domain adaptation for benign-malignant lung nodule classification on thoracic CT[J], IEEE Journal of Biomedical and Health Informatics 26(8) (2022), 3860–3871.

51.

, Liang

, Li

et al. Self-supervised transfer learning framework driven by visual attention for benign-malignant lung nodule classification on chest CT[J], Expert Systems with Applications 215 (2023), 119339.

52.

Liao

, Xie

, Hu

et al. Learning from ambiguous labels for lung nodule malignancy prediction[J], IEEE Transactions on Medical Imaging 41(7) (2022), 1874–1884.

53.

Selvaraju

R.R.

, Cogswell

, Das

et al. Grad-cam: Visual explanations from deep networks via gradient-based localization[C]// Proceedings of the IEEE International Conference on Computer Vision (2017), 618–626.

Classification of benign and malignant pulmonary nodule based on local-global hybrid network

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1 Introduction

2 Methods

2.1 Overall architecture

3.1 LIDC-IDRI and LUNGx dataset

3.2 Data processing

Table 1 Training parameter setting Training parameter Parameter value Batch size 64 Epochs 50 Optimizer AdamW Learning rate 0.006 Learning rate schedule Cosine annealing Weight decay 0.05

4.1 Comparison with some state-of-the-art methods

Table 5 Classification results with the different data augmentation methods–label tab5 Augmentation methods Acc(%) Sen(%) Pre(%) Spe(%) AUC(%) Conventional 91.51 91.84 90.79 91.2 96.02 MedAugment 92.25 91.97 91.61 92.09 96.24 Conventional and MedAugment 94.42 94.25 93.05 92.87 97.26

Table 6 The results of ablation experiment Networks Acc(%) Sen(%) Pre(%) Spe(%) AUC(%) N1 91.51 91.12 90.06 90.51 95.95 N2 91.85 92.45 91.43 91.86 96.18 N3 91.39 89.66 92.29 92.98 96.02 N4 92.45 92.6 91.79 92.09 96.56 LGHNet (Ours) 94.42 94.25 93.05 92.87 97.26

References

Table 1
Training parameter setting

Training parameter Parameter value

Batch size 64

Epochs 50

Optimizer AdamW

Learning rate 0.006

Learning rate schedule Cosine annealing

Weight decay 0.05

Table 5
Classification results with the different data augmentation methods–label tab5

Augmentation methods Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

Conventional 91.51 91.84 90.79 91.2 96.02

MedAugment 92.25 91.97 91.61 92.09 96.24

Conventional and MedAugment 94.42 94.25 93.05 92.87 97.26

Table 6
The results of ablation experiment

Networks Acc(%) Sen(%) Pre(%) Spe(%) AUC(%)

N1 91.51 91.12 90.06 90.51 95.95

N2 91.85 92.45 91.43 91.86 96.18

N3 91.39 89.66 92.29 92.98 96.02

N4 92.45 92.6 91.79 92.09 96.56

LGHNet (Ours) 94.42 94.25 93.05 92.87 97.26