Advanced Lung Cancer Segmentation Using Bilateral UNet and Hybrid Classification Using CT Images

Day by day, Lung cancer emerges as a significant global health concern in the world and also it represents a substantial burden on individual’s life and healthcare systems [1]. Based on up-to-date statistics, it is one of the leading causes of cancer-related deaths in the universe. Recognizing the symptoms early and identifying the lung cancer is crucial for timely diagnosis and treatment that significantly impacts patient outcomes.

Lung cancer is classified as non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC). Also, the NSCLC includes subtypes namely adenocarcinoma, squamous cell carcinoma and large cell carcinoma. These differences are based on the specific type of lung cells in which the cancer originated and their microscopic appearance [2, 3]. On the other hand, cancers are categorized as either benign or malignant by their behavior. Benign types are non-cancerous and usually grow slowly and do not spread. In contrast, malignant types grow aggressively and affect surrounding tissues. In medical imaging, computed tomography (CT) plays an essential role in evaluating lung cancer [4-6]. The imaging techniques are aided in extracting relevant data from a given image to diagnose and prognosis the cancer.

However, DL-based CT imaging is significant for an effective early prediction analysis of lung cancer effectively. It also distinguished the benign and malignant variants in lung cancer clearly and guided appropriate clinical management. A few popular DL models particularly neural networks have demonstrated remarkable capabilities in identifying patterns and abnormalities within medical images such as Convolutional Neural Networks (CNNs), ResNet, U-Net and so on [7-9].

Existing approaches for lung cancer detection and classification still face several challenges. One major limitation is the inability of many current models to effectively distinguish between benign and malignant tumors with high accuracy and in early stages. Additionally, most traditional methods are time-consuming and require extensive manual interpretation which leads to delays in diagnosis and treatment. Many current models for lung cancer detection struggle to effectively distinguish between benign and malignant tumors at early stages with high accuracy. Existing approaches require manual interpretation and are time-consuming. It leads to delays in diagnosis and treatment. Additionally, traditional models lack robust feature extraction methods to handle the complexity and variability in lung cancer appearances on CT images.

The primary objective of this study is to develop and assess a robust pipeline for lung cancer detection that integrates advanced imaging and classification techniques to enhance clinical diagnostic accuracy. The specific aims are:

1) Enhance Lung Cancer Segmentation: To refine lung cancer segmentation by using the Bilateral U-Net model, which incorporates Bilateral Attention Mechanisms (BAM) and Global Attention Mechanisms (GAM). This approach seeks to improve the precision of identifying and delineating cancerous regions in CT images.

2) Optimize Classification Performance: To achieve superior classification outcomes by integrating Lemur Optimization with Adaboost-Backpropagation Neural Network methods. This combination aims to optimize the detection of cancerous lesions and improve early diagnostic capabilities.

Ultimately, this study aims to advance early lung cancer detection, which could lead to better patient outcomes through more accurate and reliable diagnostic processes.

The LUNA16 dataset is used in this work and comprises CT-scanned images in DICOM format (https://luna16.grand-challenge.org/). Derived from the publicly accessible LIDC/IDRI database, this dataset excludes images with a slice thickness exceeding 2.5mm. It encompasses an 888-patient sample, where CT images have a resolution of 512×512×Z (Z ranging from 100 to 400 for each). Skilled radiologists meticulously annotated nodules and non-nodules. A nodule is considered a reference standard if at least 3 out of the 4 radiologists identified it as such, with a recorded size greater than 3mm. Each nodule's location in the lung is labelled to distinguish its cancerous nature, and the size of each nodule, with a total of 1186 annotations, is documented. The entire data set is divided into training and test sets with a 3:1 ratio.

In this preprocessing, the images are resized to a uniform resolution of 512×512 pixels. Additionally, the dataset is divided into 10 subsets for 10-fold cross-validation. Figure 1 shows the proposed segmentation model of Bilateral U-Net is characterized by encoding and decoding. The encoding uses a basic residual block that comprises two repeated convolutional layers with identical padding. The batch normalization (BN) and ReLU nonlinear layers are processed for each convolutional layer. A down-sampling is used as a 2×2 max-pooling operation in every basic residual block, with repetition of the feature channels at every step with 32 channels. Also, the Bilateral U-Net is integrated with two main modules namely Bilateral Attention Module (BAM) and Graph Attention Module (GAM) respectively. The BAM module integrates both the Channel Attention Layer and Positioning Attention Layer to capture global information that is used to serve as a key component in the encoder.

1.png

Figure 1. Proposed Bilateral U-Net

The GAM is situated at the bottom to review global data and generate the encoder output. Consistently, an up-sampling is executed in the decoding to reestablish the spatial size of the segmented output. The feature channels are processed by every up-sampling which involves 2×2 transposed convolutions. In the decoding process, features from an encoder are transferred using skip to the decoder after passing through the BAM. Here the residual blocks have two convolutional layers, BN and ReLU activation for feature extraction. That is benefitted from a reduction of training parameters and actual long-range residual connections for an enhanced training model. Furthermore, short skip connections adopt a residual structure to enhance network optimization, expedite model convergence and elevate accuracy by increasing the depth of the model. The balanced combination of long and short connections enables feature extraction at different levels to improve expressive capability while concurrently complementing high-level semantic data and refined segmentation frameworks at lower levels.

3.1 BAM

The BAM is a significant module in the Bilateral U-Net which integrates both Positioning Attention and Channel Attention modules as shown in Figure 2.

2a.png

(a) Position attention modules

2b.png

(b) Channel attention module

Figure 2. BAM architecture

3.2 Position attention module

This Module enabled the encoding of a broader range of relative data into local features to ensure a more robust representation. To combine a spatial framework, the local feature A $\epsilon$ ℝ^C×H×W is fed into convolution layers to generate feature maps B and C, where B, C $\epsilon$ ℝ^C×H×W and redesigned to ℝ^C×N where N indicates the total pixels (N=H×W). Subsequently, transpose of C and B performed a matrix multiplication that is followed by the SoftMax layer to evaluate spatial attention S $\epsilon$ ℝ^N×N.

Therefore, the process, the influence of the i-th position on the j-th position is measured by $S_{j i}$, reflecting the impact of feature representations' similarity between the two positions is expressed in Eq. (1):

$S_{j i}=\frac{\exp \left(B_i . C_j\right)}{\sum_{i=1}^N \exp \left(B_i . C_j\right)}$         (1)

A similarity based on a higher degree enhances the correlation among these positions. At the same time, the local feature A is processed through a convolution layer are used to generate another feature map D $\epsilon$ ℝ^C×H×W. The ℝ^C×N is followed by a matrix multiplication among D and the previously transposed evaluated spatial attention map S and reshaped to ℝ^C×H×W. At last, this result is scaled by a parameter α and experiences an element-wise summation with the original features A that yields a final output E $\epsilon$ ℝ^C×H×W.

$E_j=\propto \sum_{i=1}^N\left(S_{j i} D_i\right)+A_j$               (2)

where, α is set as 0 for gradual learning to allocate growing weight and E denotes a computed feature at every position. As training progresses, α increases to give more weight to the attention-adjusted features. This process aggregates contextual information across all spatial positions and provides a global view of the feature map.

3.3 Channel attention module

Each channel in the feature map is considered to represent a different class or semantic feature. The channel attention module focuses on these inter-channel relationships to refine the feature representation. Consider X $\epsilon$ ℝ^C×C of channel map is evaluated from a feature A $\epsilon$ ℝ^C×H×W. The matrix multiplication is performed among A and transposes A to provide features A to R C×N. Finally, a SoftMax layer is used to evaluate a feature X $\epsilon$ ℝ^C×C. Thus, the attention map X is influenced by i-th and j-th channels given in Eq. (3):

$X_{i j}=\frac{\exp \left(A_j \cdot A_i\right)}{\sum_{i=1}^C \exp \left(A_i \cdot A_j\right)}$         (3)

From the above Equation, the feature is reshaped as ℝ^C×H×W and it is scaled by a parameter β and provided a result of feature E $\epsilon$ ℝ^C×H×W given in the below Eq. (4):

$E_j=\beta \sum_{i=1}^c\left(x_{i j} A_i\right)+A_j$            (4)

Based on Eq. (4), the long series semantic improves the feature discriminability to enhance the performance.

It's worth noting that, in the channel relationships using global pooling or encoding layers, the convolution layers are refrained from implanting features in it. This decision allows for the preservation of relationships between different channel maps. By leveraging spatial data at all positions efficiently model channel correlations.

3.4 Graph attention module (GAM)

The GAM module is integrated into the proposed Bilateral U-Net to effectively use global structural details across a stack of medical images. It is transcending the utilization of non-local detail data within individual images. The vertices in the graph are derived a down sampling features by applying a fully convolutional layer. The for each vertex s_i are collectively regulated as considered Vertex (V) with character expressions (S_i) i.e., V={s₁, s₂, ..., s_N} with v_i $\epsilon$ ℝ^Q, where Q indicates the character expression dimension, N denotes slices quantity in the stack respectively.

The enhancement and diversification involve projecting the features into a higher-dimensional space which facilitates better differentiation and learning of the structural details across the image stack. To enhance and diversify, v_i $\epsilon$ ℝ^Q with a new dimensionality Q, a linear deformation W $\epsilon$ ℝ^Q×Q is applied to v_i and its neighboring vertices v_j. The attention coefficients (e_ij) measured the features vertex j to i by using the attention coefficient (α) with the function of $W_{s_i}$ and $W_{s_j}$. $W_{s_i}$ represents the linear transformation matrix applied to the feature vector of the i-th vertex in the graph. It maps the original feature space of vertex iii into a higher-dimensional space Q. Similarly, $W_{s_j}$ represents the linear transformation matrix applied to the feature vector of the j-th vertex in the graph. This module enables every vertex to use its structural data detail. To build a computation efficiency, the i-th node with the k-nearest neighbour (N_i) is evaluated with a significant portion e_ij. Whereas the i-th k slices are chosen from {s_i−k/2, ..., s_i−1, s_i+1, ..., s_i+k/2} with k set to 4. Next, the α undergoes normalization as given in Eq. (5):

$\alpha_{i j}=\frac{\exp \left(\sigma\left(a^T\left[W_{s_i}| | W_{s_j}\right]\right)\right)}{\sum_{k \in N_i} \exp \left(\sigma\left(a^T\left[W_{s_i} \| W_{s_k}\right]\right)\right)}$              (5)

where, || denotes concatenation, σ indicates LeakyReLU activation. By using Eq. (5), the new feature $\left(S_i^{\prime}\right)$ is derived as:

$S_i^{\prime}=\sigma\left(\sum_{j \in N_i} \alpha_{i j} W_{S_j}\right)$           (6)

where, $S_j$ as linear features with global structural data.

3.5 Feature optimization

Feature optimization or feature selection involves selecting the most effective subset of features from the original feature vector to enhance accuracy and decrease computational time.

This work uses the Lemur Optimization (LO) algorithm for feature selection. Compared to other algorithms, this optimizer achieves better results in terms of convergence rate and escaping from local minima issues [29, 30]. The LO model is grounded in the locomotor behaviors of lemurs specifically "leap up" and "dance hub" each representing a distinct stage: exploration and exploitation. In the exploration stage, a leap-up behavior is employed to pinpoint the optimal lemur location within the search space. In the exploitation stage, a dance-hub behavior is utilized to identify the best nearby location in a specified direction.

The LO model is then mathematically formulated based on these behavioral concepts. Each lemur solution is represented by an individual vector with specific coordinates for each lemur. The determination of the best location for each lemur is achieved through the fitness function evaluation. Accordingly, the lemurs adjust their position vectors. The best nearby lemur is identified through a dance hub, while the best global lemur is pinpointed through a leap-up. Initialize the random population matrix given in Eq. (7):

$T=\left[\begin{array}{cccc}l_1^1 & l_1^2 & \ldots & l_1^d \\ l_2^1 & l_2^2 & \ldots & l_2^d \\ \vdots & \vdots & \vdots & \vdots \\ l_S^1 & l_S^2 & \ldots & l_S^d\end{array}\right]$            (7)

The is randomly generated as follows:

$\begin{aligned} & \left.l_i^j=\operatorname{rand}() \times\left(u b_j-l b_j\right)+l b\right) \forall i\in(1,2, \ldots, n) \wedge \forall i \in(1,2, \ldots, d)\end{aligned}$           (8)

where, $l_i^j$ indicates the j-th dimension value of the ith lemur's position vector. The random distribution used for initialization is a uniform distribution. The discrete lower bound limits of j as lbj and the upper bound limit of j is ub_j. The best location of the lemur is expressed as the following equation:

$L_i^j=\left\{\begin{array}{c}l(i, j)+a b s(l(i, j)-l(b n l, j)) \\ *(\text { rand }-0.5) * 2 ; \text { rand }<F R R, \\ l(i, j)+a b s(l(i, j)-l(g n l, j)) \\ *(\text { rand }-0.5) * 2 ; \text { rand }>F R R,\end{array}\right.$        (9)

where, bnl is the Position of the best nearest lemur in the j-th dimension. gnl is the Position of the global best lemur in the j-th dimension. FRR is the Fitness Reduction Ratio. From Eq. (9), the direction variable $\left(L_i^j\right)$ performed for global best lemur selection and best nearest lemur selection.

The pseudocode for the LO model, presented in Algorithm 1, outlines the procedural steps involved.

The algorithm takes input parameters such as the number of iterations, dimensions, solutions, and bounds. In each iteration, the algorithm determines the objective function, evaluates the free risk rate, and generates the Global Best Lemur. For each lemur, the Best Nearest Lemur is assessed, and decision variables are updated based on randomization and the Jumping Rate parameter. This process iterates till the specified number of iterations is gotten by returning the Global Best (Best features). Following feature selection, the chosen features are applied in a classification task, showcasing the adaptability and efficacy of the Lemur Optimizer for optimizing feature sets and enhancing subsequent classification outcomes.

3.6 Classifier

In this work, the AdaBoost-BPNN model uses a BPNN as the base weak classifier. The outputs of multiple BPNN models are iteratively refined during the training process. AdaBoost is then applied to combine these weak classifiers into a stronger more robust classifier. Figure 3 illustrates the AdaBoost-BPNN classifier, which integrates multiple instances of the BPNN to achieve effective classification. The hidden layer uses the hyperbolic tangent sigmoid function ("tanh"). The output layer uses a linear function ("identity") to map predictions.

3.png

Figure 3. AdaBoost-BPNN classifier