AI-based automated breast cancer segmentation in ultrasound imaging based on Attention Gated Multi ResU-Net

Ultrasound imaging

Ragab et al. (2022) proposed the Ensemble Deep-Learning-Enabled Clinical Decision Support System for Breast Cancer Diagnosis and Classification. They utilized ultrasound scans for breast cancer detection, employing Kapur’s entropy and the Chaotic Krill Herd Algorithm (CKHA) for segmentation, followed by feature extraction using a combination of VGG-16, VGG-19, and SqueezeNet models. Classification was performed using Multilayer Perceptron (MLP) and Cat Swarm Optimization (CSO), demonstrating superiority over existing techniques. Another group (Yang & Yang, 2023) introduced a novel approach for breast lesion segmentation using breast ultrasound images. Their technique combined the strengths of basic CNNs and Swin transformers, utilizing a hybrid model. The method incorporated feature localization with CNNs, global feature extraction with Residual Swin Transformer Blocks (RSTB), and fusion of multi-scale features with Supplemental Feature Fusion (SFF) and Interactive Channel Attention (ICA). The boundary detection (BD) module was then employed for accurate lesion characterization, showing superior performance over other methods.

Cho, Baek & Park (2022) proposed a multistage-based breast cancer segmentation method for ultrasound images. Their approach, employing Breast Tumor Ensemble Classification Network (BTEC-Net) and Residual Feature Selection UNet (RFS-UNet), demonstrated improved accuracy in tumor segmentation compared to conventional methods. They utilized datasets like BUSI and UDIAT for validation and employed the Grad-CAM approach for precise feature extraction. A research group led by Jabeen et al. (2022) introduced a method for precise ultrasound-based breast cancer detection utilizing transfer learning techniques. They retrained the DarkNet-53 model with data augmentation and optimized features using reformed grey wolf (RGW) and differential evolution (RDE) techniques. Fusion techniques and machine learning algorithms were employed for accurate categorization, achieving a high accuracy rate of 99.1%.

Mammography and MRI

Mammography and MRI have long been foundational in the screening and diagnostic processes for breast cancer, providing detailed and contrast-rich images that are crucial for identifying early signs of malignancy. Iqbal & Sharif (2022) proposed the Multiscale Dual Attention-Based Network (MDA-Net) for early and accurate breast lesion detection. Their method, focusing on mammography and MRI images, utilized an encoder–decoder architecture with multiscale fusion (MF) and dual attention (dA) blocks for precise segmentation. Experimental results showcased the superiority of MDA-Net over alternative segmentation techniques. This method successfully combines information at many dimensions. It applies focused attention mechanisms to areas most indicative of pathology, addressing major obstacles in breast imaging, such as the diversity of tumor appearance and the overlap of tissues. The outcomes of the experiments showed that MDA-Net outperforms traditional techniques in segmentation precision, potentially leading to a notable increase in diagnosis accuracy. The technique improves the identification of tiny lesions that conventional techniques could miss and lowers false positives, supporting more trustworthy and self-assured clinical decision-making.

Thermography-based diagnosis

Thermography has emerged as a promising non-invasive diagnostic tool for early breast cancer detection, utilizing the subtle variations in temperature distribution across breast tissues that can indicate the presence of malignancies (Roslidar et al., 2020) reviewed deep learning methods for thermography-based early breast cancer diagnosis. They emphasized the utilization of temperature distribution in breast images for cancer detection. While acknowledging the effectiveness of thermography, they suggested improvements such as representative datasets, appropriate regions of interest (ROIs), kernel optimization, and lightweight CNN models for enhanced performance. Similarly, Amethiya et al. (2022) conducted a comparative analysis of machine learning and biosensor approaches for breast cancer identification. They evaluated various machine learning algorithms including artificial neural networks (ANN), Random Forests (RF), support vector machines (SVM), and others. Additionally, they explored the effectiveness of biosensors, highlighting electrochemical biosensors as particularly precise and efficient compared to piezoelectric and optical biosensors.

Deep learning system for breast tumor classification

Advancements in deep learning have significantly propelled the field of medical imaging, particularly in the classification and segmentation of breast tumors. Abunasser et al. (2022) proposed the Xception deep learning system for breast cancer identification and classification. Utilizing convolutional neural networks (CNNs) and generative adversarial networks (GANs), they demonstrated the effectiveness of the Xception algorithm in accurately detecting and categorizing breast tumors. Chen et al. (2022) introduces a new method for Automatic Breast Ultrasound (ABUS) tumor segmentation called the Region Aware Transformer Network (RAT-Net). To improve segmentation accuracy, RAT-Net incorporates region previous knowledge into a UNet architecture using a Transformer encoder4. Two major elements are presented: the Region Aware Transformer Block (RATB) and the Region Aware Self-Attention Block (RASAB), which concentrate on tumor locations with a high risk of being suspicious1. Experiments on a dataset of 256 subjects show that the system beats state-of-the-art algorithms, indicating its usefulness in medical image segmentation with strong region priors.

Despite substantial advances in machine learning and deep learning applications in this industry, some issues persist. Enhancing tumor classification accuracy, managing heterogeneous and non-uniform datasets, and lowering false positives in segmentation tasks are a few of these requirements. In order to address the issues of image variability and data sensitivity, our study presents an improved model that not only increases the resilience and precision of tumor identification of imaging modalities but also integrates cutting-edge computational techniques.

Architecture of Attention Gated Multi ResUnet

The Attention Gated Multi-Resolution UNet (Attention Gated Multi Res U-Net) employing multi-resolution blocks (MRBs) can be described as follows (also illustrated in Table 1):

1. Input: The encoder receives a breast ultrasound scan as input.

2. Feature extraction: The first stage involves a sequence of MRBs to extract characteristics from the image. An MRB consists of batch normalization layers followed by convolutional layers with a kernel size of (3 × 3) and rectified linear unit (ReLU) activation. Max pooling procedures may be applied after the MRBs to enhance the feature maps and reduce spatial resolution.

3. Deep encoding: To create a deeper encoder system, the above processes can be repeated multiple times, with the number of repetitions varying based on the desired depth of the model. After encoding, local features are upsampled using bilinear interpolation or transpose convolutional layers to improve spatial resolution. The upsampled feature maps are then concatenated with matching feature maps from the encoder.

4. Multi-resolution feature enhancement: The concatenated feature maps undergo further processing by several MRBs to generate more unique features.

5. Attention mechanism: An attention mechanism is introduced at this stage, utilizing attention gates to weigh features from encoder and decoder branches, prioritizing informative areas of the image.

6. Deep multi-resolution UNet: The above stages are repeated multiple times to create a deep multi-resolution UNet, with each repetition operating at a different spatial resolution.

The algorithm’s second feature map is multiplied with an attention mask to implement the learning algorithm. This procedure can be described as follows: (1)${\displaystyle \mathcal{I}\left(x,y\right)=\left\{\begin{array}{cc}\mathcal{F}\left(x,y\right)\ast 1.0\hfill & \text{if}\left(x,y\right)\in \mathcal{M}\hfill \\ \mathcal{F}\left(x,y\right)\ast 0.0\hfill & \text{if}\left(x,y\right)⁄\in \mathcal{M}\hfill \end{array}\right.,}$

where $\mathcal{I}\left(x,y\right)$ represents the features at position (x, y), and $\mathcal{F}\left(x,y\right)$ denotes the corresponding feature map.

Following the multi-resolution UNet, a series of MRBs process the final feature maps to create a segmented image of the breast tissue. The final layer employs a sigmoid activation function to produce a binary segmentation map, classifying each pixel as tissue or background. By utilizing MRBs in the Attention Gated Multi ResU-Net, the network achieves a modular and effective implementation while maintaining the ability to extract and enhance features at various resolutions. The attention mechanism enhances segmentation task performance by focusing more on informative image areas.

Multi-Res block

In the design of a deep neural network, a multiRes unit serves as a specialized type of residual block. Unlike a standard residual block, which operates solely at a single resolution, the multiRes block functions across various resolutions, hence the term “MultiRes”, derived from “Multiple Resolution”.

A MultiRes block consists of multiple sub-blocks operating at different resolutions (e.g., low-resolution, mid-resolution, high-resolution). By connecting sub-blocks in series, each sub-block passes its output as input to the next sub-block, allowing content to be distributed across multiple resolutions.

Each sub-block within a MultiRes Block typically incorporates various operations, including convolution operations, batch normalization layers, activation functions, and possibly pooling layers. This design enables the network to extract more semantic or fine-grained features from the input data.

Utilizing MultiRes Blocks enhances the accuracy and stability of deep neural networks, particularly in tasks such as image classification and segmentation, where accurately representing input data across multiple scales is essential. This skip connection method is implemented for each convolutional block using the following equation: (2)${\displaystyle y=F\left(x,\left\{w_i\right\}\right)+H\left(x\right),}$

Where:

y: Represents the output of the MultiRes block.

F(x,  {w_i}): Denotes the operations performed within the block, which may include convolution operations, max-pooling operations, or up-sampling operations, with corresponding weights {w_i}.

H(x): Represents the identity mapping or another convolutional operation to maintain the input’s feature dimensions as F.

In Fig. 1, the MultiRes Block (MRB) is a fundamental component of the Attention Gated MultiRes U-net architecture. It is embedded within the encoding and decoding paths of the network, facilitating multiscale feature extraction. Specifically, in Fig. 1, the MRB modules are represented within each block of the encoding and decoding paths. These MRB modules, as depicted in Fig. 2, are responsible for capturing features at multiple resolutions, enhancing the model’s ability to learn and extract meaningful information from the input data. Therefore, the MRB modules play a pivotal role in the overall architecture of the Attention Gated MultiRes U-net, contributing to its effectiveness in tasks such as cancer segmentation.

Attention gate

The attention gate is a pivotal mechanism within the UNet architecture, designed to enhance the robustness and accuracy of image segmentation models. By integrating elements from both U-Net and attention mechanisms, the attention gate offers a practical approach to semantic segmentation. Its core objective is to amplify activations of relevant feature maps in the encoding path while suppressing activations of non-relevant feature maps. Figure 3 shows the architecture of attention gate block.

The attention gate comprises two essential components: a gating process and a skip connection. The gating mechanism generates a weight map, representing the significance of each feature space in the encoding path. This weight map is created by concatenating the image features with a learnable weight tensor, followed by a sigmoid activation function. The skip connection then combines the encoded path activations with the weighted map to produce the final attention-gated image features.

By selectively focusing on input image areas containing the most relevant information, the attention gate enhances the performance of UNet-based segmentation models, improving accuracy and robustness. This allows the framework to remain unaffected by minor or distracting details, making it adaptable to challenges such as variations in object size, structure, and visual appearance. (3)${\displaystyle \begin{array}{c}\hfill z=\text{Conv}1\times 1\left(x\right)a=\text{Sigmoid}\left(z\right)y=a\odot x+\left(1-a\right)\odot xup\end{array}}$

Where:

x: Input feature map.

x_up: Upsampled feature map from the corresponding decoder layer.

Conv_1×1: 1x1 convolutional layer.

Sigmoid: Sigmoid activation function.

⊙: Element-wise multiplication.

Loss function

Binary cross entropy (BCE) loss (Bahdanau, Cho & Bengio, 2014), is a commonly used loss function for binary classification tasks such as breast cancer segmentation. It assesses the discrepancy between the ground truth label and the predicted probability that each pixel represents breast tissue.

The BCE loss penalizes the model more severely for predictions farther from the actual label, thereby encouraging the model to focus on the most crucial areas of the image for segmentation. It pushes the model to deliver high-confidence predictions for pixels correctly identified as breast tissue or background while penalizing it more severely for incorrectly categorized pixels. (4)${\displaystyle L=-\frac{1}{N}\sum _{i=1}^N\left(y_{i}log\left(\widehat{y_i}\right)+\left(1-y_i\right)log\left(1-\widehat{y_i}\right)\right)}$

Where:

N: Number of pixels in the image.

y_i: Ground truth label for pixel i.

$\widehat{y_i}$: Predicted probability for pixel i.

Hyperparameters

Hyperparameters greatly influence the effectiveness of the Attention Gated ResUNet for cancer segmentation. By carefully adjusting these parameters, the model’s accuracy may be increased. These parameters regulate many aspects of the model’s design and training procedure. Adam optimization techniques are utilized as optimization for the proposed network (Kingma & Ba, 2014). The learning rate, the number of epochs, batch size, activation function, and the number of multi-resolution blocks are some of the critical hyperparameters that might affect the performance of the Attention Gated ResUNet (MRBs). For instance, a low learning rate can prevent overfitting but may cause the model to converge slowly or not at all, whereas a high learning rate can allow the model to converge fast but may result in overfitting. The performance of the Attention Gated ResUNet for breast cancer segmentation may be significantly affected by tweaking the hyperparameters. These parameters can be adjusted carefully to maximize the model’s accuracy and resilience. Table 2 presents the hyperparameters selected for the experimentation.

The performance of the Attention Gated ResUNet for cancer segmentation is greatly influenced by several critical hyperparameters, among which the parameters related to the multi-resolution blocks (MRBs) play a significant role. The MRBs are fundamental components of the network architecture responsible for capturing multi-scale information from the input images, which is crucial for accurately segmenting cancerous regions. Tuning the parameters associated with MRBs can profoundly impact the model’s performance.

One of the most critical parameters within the MRBs is the number of convolutional layers and filters. Increasing the number of layers allows the model to capture more complex features and hierarchies in the input images, potentially leading to improved segmentation accuracy. However, adding too many layers can also increase the risk of overfitting, mainly if the dataset is limited. Conversely, reducing the number of layers may lead to underfitting, where the model fails to capture essential characteristics of the input data.

Similarly, the size of the filters used in the convolutional layers within the MRBs is another crucial parameter. Larger filter sizes enable the model to capture more global features and contextual information, which can be beneficial for segmenting cancerous regions that may span across multiple scales. On the other hand, smaller filter sizes allow the model to capture finer details and local features, enhancing the segmentation accuracy for smaller lesions or abnormalities.

Additionally, the choice of activation function within the MRBs can significantly impact the model’s performance. Activation functions such as ReLU (Rectified Linear Unit), Leaky ReLU, or ELU (Exponential Linear Unit) govern the non-linear transformations applied to the feature maps, enabling the network to learn complex patterns and relationships within the data. Selecting the appropriate activation function is crucial for ensuring that the model can effectively capture the non-linearities present in the input images, thereby improving segmentation accuracy.

Moreover, regularization techniques such as dropout or batch normalization within the MRBs can help prevent overfitting and improve the model’s generalization ability. Dropout randomly deactivates a fraction of neurons during training, forcing the network to learn more robust and generalizable features. Batch normalization normalizes the activations of each layer, reducing the internal covariate shift and accelerating the training process.

Tuning the parameters associated with MRBs in the Attention Gated ResUNet is essential for optimizing the model’s performance in cancer segmentation tasks. By carefully adjusting the number of convolutional layers and filters, filter sizes, activation functions, and regularization techniques within the MRBs, researchers can enhance the model’s ability to capture multi-scale information and accurately segment cancerous regions in medical images.

The learning rate, batch size, and number of epochs are key hyperparameters influencing the performance of the Attention Gated ResUNet. A suitable learning rate, such as 0.001, balances convergence speed and stability. The batch size, typically set at 8, affects convergence and computational efficiency. The number of epochs, often around 100, balances model complexity and prevents overfitting. These parameters play a crucial role in optimizing model training for accurate cancer segmentation while managing computational resources effectively.

Data augmentation

Several transformations, including rotation, flipping, scaling, and cropping, are applied to the existing images to create new, slightly modified versions of the original images. Data augmentation can be applied to breast cancer ultrasound pictures to expand the dataset and improve the model’s ability to generalize to new data. The model becomes more resistant to fluctuations in the dataset and less prone to overfitting particular photos in the training set by being exposed to various image variations. Randomly changing the images’ orientation helps the model recognize certain aspects of the scans. The model can learn to recognize photo features independent of their position by flipping the images horizontally or vertically. Changing the size of the images can train the model to recognize features in photos of any size. By focusing on different areas of the photos, the model can learn to recognize features in those areas. A considerably bigger dataset can be produced by applying these alterations to the already-existing photos, which can enhance the model’s performance and make it more resistant to changes in the data. However, it is crucial to remember that data augmentation should be done cautiously to prevent adding irrational changes to the data that could impair the model’s performance.

Experimental setup

The deep learning experiment uses a machine with an Intel Core i5 11th generation CPU and an NVIDIA RTX A5000 GPU. The experiment is built with the TensorFlow deep learning framework and runs on the Windows operating system. The experiment’s dataset is gathered, pre-processed, and divided into training, validation, and testing sets for which our deep learning model layout is designed and deployed using the deep learning framework of choice. The model is trained on the training data set, and its performance is measured on the validation data set. The model’s performance is optimized by fine-tuning the hyperparameters. The outcomes are studied and interpreted to learn more about the model’s strengths and weaknesses.

Evaluation metrices

The confusion matrix may be used to assess a classifier’s performance (CM). Additionally, CM determines assessment criteria like precision, sensitivity, specificity, and accuracy. It shows the classification’s outcomes based on actual results. In addition, four scores are computed: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). The following categories list the reference records incorrectly or correctly identified as successes or failures. The accuracy shows the number of sequences that have been appropriately detected based on how many batches were utilized in the cross-validation iteration, which is the most frequently utilized parameter for evaluating the performance of a classifier. (5)${\displaystyle ACC=\left(\sum _{i=1}^N\frac{TP+TN}{TP+FP+TN+FN}\right)\cdot 100\text{\%}/N.}$ The classification algorithm’s sensitivity determines its success in diagnosing all ill persons; specificity refers to the fraction of negative samples that the classification model properly classified as negative instances. The ROC curve is another crucial quality indicator used to assess the suggested solution’s effectiveness visually. (6)${\displaystyle SEN=\left(\sum _{i=1}^N\frac{TP}{TP+FN}\right)\cdot 100\text{\%}/N}$ (7)${\displaystyle SPE=\left(\sum _{i=1}^N\frac{TN}{FP+TN}\right)\cdot 100\text{\%}/N}$ (8)${\displaystyle PRE=\left(\sum _{i=1}^N\frac{TP}{FP+TP}\right)\cdot 100\text{\%}/N}$ (9)${\displaystyle F1-Score=\left(\sum _{i=1}^N\frac{2\left(TP\right)}{2\left(TP\right)+FP+FN}\right)\cdot 100\text{\%}/N.}$

Segmentation performance

The segmentation data shown in Fig. 6 can be used to assess the performance of the proposed model. The anticipated masks show that the proposed model can accurately segment disease tissues. The Attention Gated Multires U-Net can be used to segment breast tissue and spot abnormalities like masses, cysts, or calcifications in the context of breast cancer ultrasound pictures. The model may be trained to separate particular objects, such as the pectoral muscle or the ducts. For the early identification and diagnosis of breast cancer, accurate segmentation of breast ultrasound images is essential, and the Attention Gated Multi ResU-Net has demonstrated encouraging results in this area.

Results from a quantitative study of the Attention Gated Multi ResU-Net for segmenting breast cancer ultrasound images have been encouraging. In recent research, the model segmented breast tissue with an average Dice coefficient of 0.86, demonstrating strong agreement between the model’s segmentation and expert manual annotations. The Dice coefficient, which runs from 0 to 1, assesses the overlap between two sets of segmentation masks, with a value of 1 completely denoting agreement. Also, the model’s effectiveness in locating and classifying masses in breast ultrasound images was assessed. The mass segmentation average Dice coefficient was 0.74, suggesting a moderate to high level of agreement between the segmentation produced by the model and the manual annotations. The model’s mass detection sensitivity and specificity were 88.3% and 93.7%, respectively. The specificity assesses the percentage of genuine negative detections, whereas the sensitivity assesses the percentage of genuine positive detections.

The model was also tested for its ability to identify and separate calcifications in breast ultrasound scans. The segmentation of calcifications using the model and hand annotations had an average Dice coefficient of 0.75, indicating moderate to high agreement. For calcification identification, the model’s sensitivity and specificity were 75.0% and 97.7%, respectively. The Attention Gated Multires U-Net can precisely recognize and segment areas of abnormalities in the breast tissue, according to the quantitative analysis of the breast cancer ultrasound picture segmentation model. Although the model’s performance varies based on the particular job, the results show its promise for enhancing early detection and diagnosis of breast cancer.

Feature extraction capability

The performance of the proposed network in segmenting breast cancer ultrasound images is primarily due to its feature extraction capacity as shown in Fig. 7. The model uses a multiscale feature extraction method to collect data at various scales and resolutions. Convolutional neural network (CNN) layers with various filter sizes and pooling procedures are used to accomplish this task. The model’s capacity to extract multiscale features enables it to recognize structures in the breast tissue of different sizes. For instance, the model can precisely separate smaller and bigger lumps, crucial signs of breast cancer. For a precise diagnosis and treatment planning, this is essential.

The model can also capture both fine- and coarse-grained aspects of the breast tissue due to the multi-scale methodology. Coarse-grained features are more enormous structures like ducts, masses, and cysts, and fine-grained features are more minor details like texture, edges, and corners. The model can precisely segment the breast tissue and spot abnormalities by including both characteristics.

The proposed attention-gating method improves its feature extraction capabilities by enabling it to focus only on crucial areas of the image. This guarantees that the segmentation model catches the most pertinent properties while avoiding unimportant data, such as noise or distortions. Consequently, a critical factor in the Attention Gated Multi ResU-Net’s performance in the segmentation of breast cancer ultrasound images is its feature extraction capacity. The model’s attention-gating mechanism and multiscale approach allow it to capture the breast tissue’s fine- and coarse-grained characteristics, producing accurate and exact segmentation results.

Effect of resolution

The performance of deep learning models, such as our proposed Attention Gated Multi ResU-Net for breast cancer ultrasound image segmentation, can be significantly impacted by the resolution of the training data. The accuracy of the model’s segmentation outputs can be increased by using higher-resolution scans, which can give more detailed information and finer-grained breast tissue characteristics. Nevertheless, employing higher-quality images can lengthen the model’s training period and increase computing complexity.

In the next series of tests, we enlarge the scans to 500 × 500, 256 × 256, and 128 × 128 pixels to examine the proposed model’s resilience for various image sizes. The performance suffers due to increased visual noise when the images are scaled up to 500 × 500 pixels. Despite the drop in size, noise reduction makes the model performance on the image of 256 × 256 better than that on images of 500 × 500. The model performance starts to suffer when the size is further reduced to 128 × 128 pixels due to the smaller nucleus size. The model trained on 256 × 256 pixel pictures has the best F1-score, Dice, accuracy, and recall values, as shown in Table 3.

On the other hand, lower-resolution scans may suffer a loss of information and detail, which may have a detrimental effect on the segmentation findings’ accuracy. Nevertheless, employing lower-quality photos can make the model more efficient for large-scale applications by reducing computational complexity and training time. While collecting imagery for training, it is crucial to carefully weigh the trade-offs between image resolution, computational complexity, and training time. Overall, it is advised to utilize the highest resolution photographs that are computationally possible since this can increase the model’s performance. But it’s also crucial to balance this and factors like the accessibility of computational resources and the overall effectiveness of the training process.

Table 4 demonstrates that the proposed model outperforms the other models in terms of Dice coefficient, sensitivity, and specificity, demonstrating that it is the most accurate model for segmenting ultrasound images of breast cancer. Regarding performance among the examined models, the UNet++ and Mask R-CNN architectures come in second and third, respectively. The DeepLabV3+, PSPNet, RefineNet, and LinkNet designs exhibit reasonably high accuracy shown in Fig. 8 while being just marginally less accurate than the top-performing models, demonstrating their suitability for applications requiring the segmentation of breast cancer ultrasound images. Overall, the results illustrate the usefulness of Attention Gated Multi ResU-Net architecture and show the promise of deep learning models for effectively segmenting breast cancer ultrasound scans.

Considerations

Dataset selection, clinical relevance, methodology, and considerations are as follows:

Dataset selection rationale: The dataset was chosen based on availability, relevance to breast cancer ultrasound imaging, and the need for diverse image representation. While clinical data would offer real-world applicability, the use of publicly available datasets allowed for standardized evaluation.

Consideration of clinical data: Although clinical data was absent, the use of simulated data facilitated controlled experimentation. Future research should integrate clinical data to validate the model’s performance in real-world scenarios.

Interpretation of confusion matrix: The confusion matrix offers insights into the model’s performance metrics, including accuracy, precision, recall, and F1-score, crucial for identifying segmentation strengths and weaknesses.

Feature extraction methodology: The chosen methodology was selected for its effectiveness in capturing multiscale features from ultrasound images. While alternatives exist, the selected approach showed promising results in initial experiments. Future research could explore state-of-the-art feature extraction techniques for further improvements.

Comparison with prior works and novelty of the proposed method

Previous studies have explored automated breast cancer segmentation using various methodologies, providing valuable insights into the effectiveness of different approaches. For instance, Srinivasan et al. (2024) employed a CNN architecture for breast cancer segmentation, achieving a Dice coefficient of 0.85 ± 0.03 and a sensitivity of 0.88 ± 0.05. Similarly, Chen et al. (2022) utilized a AU-Net-based model and reported a Dice coefficient of 0.88 ± 0.04 and a sensitivity of 0.89 ± 0.06. Furthermore, Josan (2023) proposed a modified Mask R-CNN framework, yielding a Dice coefficient of 0.89 ± 0.03 and a sensitivity of 0.91 ± 0.04.

In comparison, our study utilizing the Attention Gated MultiRes U-Net architecture achieved superior segmentation performance. The proposed model exhibited an average Dice coefficient of 0.93 ± 0.03 and a sensitivity of 0.93 ± 0.04. These results surpass those reported in the aforementioned studies, indicating the efficacy of our approach in accurately delineating breast cancer tissue in ultrasound images. Additionally, our model demonstrated a specificity of 0.99 ± 0.01, highlighting its ability to accurately identify true negative instances.

By surpassing the performance of existing methodologies, our research underscores the effectiveness and promise of the Attention Gated MultiRes U-Net architecture for automated breast cancer segmentation. Furthermore, our study contributes to advancing the state-of-the-art in this field, offering a robust and reliable tool for early detection and diagnosis of breast cancer.

The proposed model demonstrates superior performance compared to other segmentation architectures primarily due to its effective utilization of attention mechanisms and MRB structures tailored specifically for breast ultrasound imaging datasets. The attention mechanism enhances the model’s segmentation accuracy by allowing it to selectively focus on relevant features while suppressing noise and irrelevant information. Mathematically, the attention mechanism can be represented as y= a⊙x + (1 − a)⊙x_up, where x is the input feature map, x_up is the upsampled feature map, a is the weight map derived from a sigmoid activation function applied to a 1x1 convolutional layer, and ⊙ represents element-wise multiplication. This mechanism enables the model to assign higher importance to informative regions in the ultrasound images, leading to more accurate segmentation outcomes. Additionally, the incorporation of MRBs facilitates the extraction of multi-resolution features, capturing both fine-grained and coarse-grained details present in breast ultrasound images. This is mathematically represented as y= F(x,  {w_i}) + H(x), where F comprises convolution operations, max-pooling, or upsampling, and H represents identical mapping or another convolution operation to maintain feature dimensions. By leveraging attention and MRB mechanisms optimized for the characteristics of breast ultrasound data, the proposed model achieves enhanced segmentation performance, making it particularly well-suited for breast cancer diagnosis and treatment planning.