Future Fusion+ UNet (R2U-Net) Deep Learning Architecture for Breast Mass Segmentation ^†

Abstract

R2U-Net, or Recurrent Residual U-Net, is a U-Net extension that includes both residual and recurrent connections for image segmentation tasks. R2U-Net is an image segmentation task-focused network that mixes residual and recurrent connections to boost performance and manage sequential data. Semantic segmentation algorithms based on deep learning (DL) have demonstrated state-of-the-art performance recently. Specifically, these methods have proven effective for tasks like medical image segmentation, classification, and detection. U-Net is one of the most prominent deep learning techniques for these applications. These proposed structures for segmentation problems have various advantages. In addition, better feature representation for segmentation tasks is provided by accumulating features using recurrent residual convolutional layers. Moreover allows us to design a more effective U-Net architecture for medical picture segmentation using the same amount of network parameters. The experimental results reveal that the model outperforms analogous models such as R2U-Net on segmentation tasks. The accuracy of the R2UNet model was 95.6%, while the FF + (AlexResNet + R2Unet) result was more than 97%, with an accuracy (%) of 97.4, AUC (%) of 97.35, precision (%) of 97.4, F1-score (%) of 95.26, and recall (%) of 97.16. The employment of these segmentation approaches in the identification and diagnosis of breast cancer produced outstanding results. Our proposed method could provide a more precise diagnosis of breast cancer, perhaps improving patient outcomes.

1. Introduction

In 2015, UNet—which originated from the traditional convolutional neural network—was created and put to use for the first time to process biological images. A general convolutional neural network has been designed for image classification, where an image is the input and one label is the output. However, in biological contexts, we need to identify the precise location of the abnormality in addition to determining whether the disease is present. UNet is dedicated to fixing this problem. Because it performs classification on each pixel, it can localize and distinguish borders; as a result, the input and output have the same size [1]. For example, consider the following for a 2 × 2 input image: It appears to have a “U” shape at first glance [2]. The architecture is symmetric and is made up of two major parts: the left part is called the contracting path and is made up of the general convolutional process; the right part is called the expansive path and is made up of transposed 2D convolutional layers (for now, think of it as an up-sampling technique). The feature extraction framework for breast cancer is shown in Figure 1.

UNet is dedicated to fixing this problem. Because it performs classification on each pixel, it can localize and distinguish borders; as a result, the input and output have the same size [1]. For example, consider the following for a 2 × 2 input image: It appears to have a “U” shape at first glance [2]. The architecture is symmetric and is made up of two major parts: the left part is called the contracting path and is made up of the general convolutional process; the right part is called the expansive path and is made up of transposed 2d convolutional layers (for now, think of it as an up sampling technique). Feature extraction framework for breast cancer shown in Figure 1.

2. Related Works

The Recurrent Residual U-Net (R2U-Net) is a U-Net design variant that combines residual and recurrent connections to increase picture segmentation performance, especially in applications using sequential or volumetric data [3]. The following is an explanation of the R2U-Net’s operation: Encoder for U-Net: The R2U-Net starts with a typical U-Net encoder, which is made up of multiple convolutional blocks that are organized in a contracting route. Each block is often composed of convolutional layers, activation functions (such as ReLU), and down-sampling processes (such as max-pooling) [4]. While extracting hierarchical features, the encoder gradually lowers the spatial dimensions of the input image [5]. Residual Connections: Residual connections are introduced within the U-Net blocks in the R2U-Net. These links allow the network to learn the residual (or difference) between a block’s output and its input [6]. The residual is then added back to the original input, assisting in gradient propagation during training and allowing the network to acquire richer representations without vanishing gradients. Recurring Connections: The addition of recurring connections is the R2U-Net’s fundamental innovation. Each U-Net block (both encoder and decoder) includes a recurrent layer, such as a Gated Recurrent Unit (GRU) [7]. By keeping an internal memory state that changes and transfers information through time, these recurrent connections allow the network to capture long-term dependencies in the data.

U-Net Decoder: Following the encoder, the R2U-Net includes a U-Net decoder, which is made up of multiple convolutional blocks arranged in a long path [8]. Each block is typically composed of transposed convolutions (also called up-sampling or deconvolution), concatenation with encoder skip connections, and convolutional layers with activation functions.

Skip connections, like the original U-Net, are used to connect the relevant encoder and decoder blocks [8]. By using these skip connections, the decoder can produce feature maps with high resolution from the encoder, assisting in the recovery of spatial in formation and the improvement of segmentation accuracy [9].

Final Segmentation: Map: The decoder output is routed via a final convolutional layer with an appropriate activation function [10] (for example, Sigmoid for binary segmentation or Softmax for multi-class segmentation). This layer creates the final segmentation map, which represents the likelihood of each pixel in the segmentation task belonging to a specific class.

3. Methodology

Several medical picture segmentation algorithms based on UNet have been developed for performance optimization [11]. U-Net has been enhanced in terms of improvements to the applicability range, feature enhancement, training speed optimisation, training accuracy, feature fusion, small sample training set, and generalisation. Various methodologies have been used in the design of various network architectures to meet various segmentation difficulties.

There are several advantages of these proposed architectures for segmentation tasks.

(1)

The residual unit is useful while training deep architecture;

(2)

Recurrent residual convolutional layers can be used to accumulate features offers;

(3)

Superior feature representation for segmentation tasks;

(4)

It enables us to create a better U-Net architecture with the same number of network; parameters but improved performance for medical picture segmentation;

Based on the FCN concept developed the U-Net segmentation model. its architecture, which is essentially separated into three components, is similar to the FCN encoder-decoder architecture. The first section, which comprises the down-sampling path, uses Resnet101 as its primary structural support [12]. In order to give local and global information during up-sampling, the model, as shown in Figure 1, bypasses connections between up-sampling and down-sampling paths. At output 1 1, the segmented image is provided by the convolutional layer [13]. In this encoder–decoder-based approach, a ResNet101 encoder trained on the ImageNet dataset was used.

As seen in Figure 2 the horizontal bottleneck is made up of two 3 × 3 convolutions followed by a 2 × 2 up-convolution. As illustrated in Figure 4, each stage essentially applies two 3 × 3 convolutions with batch normalisation, followed by 2 × 2 max-pooling [14]. The up-sampling path also consists of four stages, depicted as a decoder with two 3 × 3 convolutional layers and 2 × 2 up-sampling. The feature maps are divided in half at each stage [15]. Because of the use of a pretrained encoder, the model converges quickly [16]. The input image is fed into the pretrained ResNet101 encoder, which is made up of a series of residual blocks. These skip connections enable the decoder to build the appropriate feature maps by supporting the model in receiving all of the low-level semantic data from the encoder. The next steps include the addition of two 3 × 3 convolution layers, a batch normalization layer, and a ReLU non-linearity layer [17]. The output of the last decoder block is fed into a 1 × 1 convolution layer, which provides the required binary mask when it is fed into a sigmoid activation function. The term “R2U-Net” appears to combine “Residual Recurrent U-Net”. By fusing these ideas, a “R2U-Net” might be a design that incorporates both recurrent and residual connections into a U-Net framework [18]. Semantic segmentation has undergone tremendous advancements, particularly in the area of medical image analysis. Semantic segmentation divides an image into various regions based on the items or structures of interest by categorizing each pixel in the picture into a certain category. For activities like mammographic detection, organ delineation, tumour detection, disease diagnosis, treatment planning, and monitoring in medical image analysis, correct segmentation is essential. Residual units and recurrent residual convolutional layers are two advanced architectural components that have significantly enhanced the capabilities of convolutional neural networks (CNNs) and deep learning models in various tasks [19,20]. Residual units, also known as residual blocks, were introduced to address the challenge of training very deep neural networks.

4. Results and Discussions

Medical photographs depicting various states in the context of cancer or other disorders are referred to as benign and malignant image data. These images are frequently utilised in medical imaging and radiology for diagnostic purposes, research, and the training of machine learning algorithms for medical image analysis.

• Benign Images: Benign images are medical photographs that depict tissues or structures that are normal or non-cancerous. The term “benign” in the context of cancer diagnosis means that there is no evidence of malignant growth or tumour in the imaging region. Normal tissues, non-cancerous growths, or benign tumours that do not represent harm to the patient’s health are examples of benign pictures, as shown in Figure 3.
• Malignant Images: Malignant images, on the other hand, are medical photographs that depict malignant tissues or structures. These photos show the existence of a malignant growth or tumour in the imaged area. Malignant tumours have the ability to grow and spread to other regions of the body, necessitating immediate medical care and treatment.

Medical practitioners and researchers need both benign and malignant imaging data to accurately diagnose and manage diseases, including cancer. The process of analysing these photos entails the detection of patterns and traits that distinguish benign from malignant diseases. Because of the nuanced and complex nature of some diseases, this process can be difficult, and it frequently necessitates the expertise of radiologists, pathologists, and oncologists. Both benign and malignant images (Figure 3 and Figure 4) are shown above training efficiency is shown in Figure 5.

Feature Fusion: Selection Technique

In order to construct a single-dimensional feature vector, we integrated the features from ResNet 101 FeatResNet and AlexNet CNN and UNet module (i.e., FeatAlexResNet + UNet). Before performing the classification in our suggested model, both feature sets were concatenated.

Table 1. Comparison of the performance of different pre-trained models.

Deep Features	Performance Variables
	Accuracy (%)	Precision	Recall (Sensitivity)	F-Measure	Specificity
R2U-Net	94.8	95.6	95.6	81.8	77.8
Proposed AlexResNet + R2Unet	97.35	97.4	97.4	97.26	97.16

shows the comparison of performances of different pre-trained models.

Feat DeepHybrid = Conc [Feat_AlexNet, Feat _ResNet, Feat _Unet]

(1)

The proposed AlexResNet + R2Unet parametric analysis of the ResNet101 with classification models is shown in Table 1, above. There is a similar pipeline for AlexNet, extraction and classification, and area classification. Ultimately, the channel’s full detection and categorization tasks are carried out. The extensive AlexNet retraining with various classifiers was not necessary for the CAD that was presented. A round of training takes about 10 min with ten cycles, and AlexNet model testing takes less than 1 min. From the simulation, it has been proven that the FF of DNN with RestNet101 + Unet model demonstrated an performance AUC of 97.35 percent, precision of 97.4 percent, accuracy of 97.4 percent, an F1-range of 97.26 percent and a recall value of 97.16 percent (Figure 6). As a consequence, it is shown that the suggested strategy achieves superior performance compared to current techniques.

5. Conclusions

The results clearly show that FF: AlexResNet + UNet has far greater computational power than other classification algorithms, such as DNN + AlexNet, DNN + VGG, and DNN + ResNet50, for addressing the issue of early breast cancer diagnosis. Here, the pre-trained model with DNN, which is of the feed-forward type, has been used and the results of the study reveal that FF: AlexResNet + UNet performs better than other ML models by simulating the dataset and attaining an accuracy rate of 97.35%, In comparison to DNN:AlexNet, which achieves an accuracy rate of 89.4% and an error rate of 0.16%, the combination of FF + (AlexResNet + R2Unet) yields good results and an error rate of 0.12%. Therefore, we can conclude that the FF: AlexResNet + UNet approach will benefit different medical domain stakeholders by providing accurate and precise results in a shorter amount of time.