Research on cyst segmentation and classification has revealed several shortcomings. A primary challenge is achieving precise segmentation of cysts in postmenopausal women due to their small size. Current methods, including Adaptive Thresholding, Adaptive K-means, and the Watershed algorithm, struggle with accurate diagnosis. Additionally, existing optimization algorithms like HHO and RSA are insufficient for precise cyst description and require extensive training time. Segmentation of the cyst image’s edges is difficult, leading to potential overfitting and incorrect size calculation due to improper weight updates. Classification techniques such as SVM, AI, and DLNN suffer from low accuracy, negatively impacting ultrasound image analysis. In contrast, the proposed algorithmic technique addresses these issues effectively, offering the highest accuracy for cyst detection in ultrasound images.
The female ovaries can develop cysts, which are sacs that contain fluid. These cysts can be identified at an early stage through the use of ultrasound imaging. This technique involves employing adaptive deep-learning methods and an optimization algorithm to classify ovarian cysts. The initial step involves pre-processing the images by applying a guided trilateral filter (GTF) to eliminate any noise present in the input image. Subsequently, the cysts are segmented based on their size. By utilizing an Adaptive Convolutional Neural Network (AdaResU-Net), they can predict whether the cysts are benign or malignant. To achieve the optimal accuracy of AdaResU-Net, the Wild Horse Optimizer (WHO) is employed to fine-tune hyperparameters such as the learning rate, batch size, and epoch count. The optimization algorithm addresses two metrics, namely Dice Loss Coefficient (DLC) and weighted Cross-Entropy (WCE), to evaluate the segmentation output without any loss. This approach has successfully classified different types of cysts with an impressive accuracy rate of 98.87%. Figure 1 explains the overall proposed diagram of ovarian cyst segmentation.
Advanced AI Techniques: The use of an Adaptive Convolutional Neural Network (AdaResU-net) represents a sophisticated application of deep learning specifically tailored for medical image segmentation. This network architecture adapts to the complexity and variability of ovarian cyst images, enhancing segmentation accuracy.
Optimization with WHO Algorithm: The integration of the Wild Horse Optimization (WHO) algorithm for hyperparameter tuning is novel in the context of medical image analysis. This optimization technique helps to fine-tune the segmentation model, potentially improving its robustness and performance.
Clinical Relevance: By focusing on ovarian cyst detection and classification (benign vs. malignant), the method addresses a critical need in women’s health. Early detection through accurate segmentation can significantly impact patient outcomes by enabling timely medical interventions.
The novelty of this work lies in its integration of advanced artificial intelligence techniques, specifically tailored for early disease detection through deep learning-based segmentation algorithms. As of the current stage (2024), the field has evolved significantly with the introduction of adaptive deep learning architectures like AdaResU-net, which can dynamically adjust to the complexities and variability of medical images such as ultrasound scans of ovarian cysts. This adaptability enhances accuracy in detecting and classifying diseases at early stages, surpassing traditional methods that may struggle with image noise and variability. The use of innovative optimization techniques like the Wild Horse Optimization (WHO) algorithm further enhances the precision of these algorithms, marking a significant advancement in medical imaging and diagnostic capabilities.
Background and detailed explanation
Ovarian cysts, fluid-filled sacs within the ovaries, often develop asymptomatically but can lead to serious health complications such as ovarian torsion, infertility, and ovarian cancer. Early detection and accurate characterization are crucial for timely treatment and preventing adverse outcomes. Ultrasonography is the primary imaging modality due to its non-invasiveness, real-time capability, and lack of ionizing radiation. However, interpreting ultrasound images of ovarian cysts presents challenges like weak contrast, speckle noise, and hazy boundaries. To address these, this study proposes an advanced deep learning-based segmentation technique. It employs a Guided Trilateral Filter (GTF) in pre-processing to reduce noise while preserving edge information for clearer images. The Adaptive Convolutional Neural Network (AdaResU-net) adapts to the variability in cyst images, accurately segmenting and classifying cysts as benign or malignant using learned features. The Wild Horse Optimization (WHO) Algorithm optimizes hyperparameters like Dice Loss Coefficient and Weighted Cross-Entropy to maximize segmentation accuracy across diverse cyst types. Furthermore, a Pyramidal Dilated Convolutional (PDC) Network enhances diagnostic utility by classifying ovarian cyst types, thus improving clinical decision-making beyond segmentation alone.
They picked input pictures to apply a pre-processing technique that enhances their quality by removing noise through GBF. This technique prepares them for further processing.
Let measured as the entire picture space is the confined spatial region centered at pixel region with sizes . D describes a noisy picture also it’s a guessed separated picture. Then, at that point, the likelihood of the leftover , among every local pixel .
where denotes scale boundary, is a differentiable kernel capability with monotonically expanding assets, is a sign of the closeness of the noiseless powers in pixels area .
where weight represents yields high qualities as well as advances smoothing. It has a low rate when the shape is available, therefore, protecting the boundaries. is the iteration integer, at the initialization . Stand for the spatial space weight.
The primary goal of the segmentation process is to precisely separate the cyst from the background image. The proposed method categorizes cysts based on their sizes and classifies them as benign or malignant using AdaResU-Net. The network’s hyperparameters, such as batch size, learning rate, and epoch count, were optimized by WHO through iterative algorithm enhancements.
AdaResU-Net’s expected design is shown in Fig. 2, which is dependent on the ResU-Net design. Like U-net, AdaResU-Net comprises a downsampling pass to the left and an upsampling pass to the right. Yet, the essential components of AdaResU-Net are the remaining knowledge systems, each consisting of three padded convolutional layers. The leftover building blocks within the reduction pass were utilized by the max-pooling activity of step 2, which logically reduces the size of the component map. This comprehensive approach employs upsampling, convolutional layers toward gradually expanding the dimensions of the element map up until it gets to the initial information dimension. By employing upsampling and downsampling techniques, the quality and level of detail in the shared image can be adjusted, fostering a strong residual association between blocks of similar dimensions. The final convolutional layer in the tissue possesses a suitable 1 × 1 dimensional channel and is activated using a sigmoid function.
The encoder, also known as the contracting path, plays a crucial role in extracting relevant spatial and feature information from the input data.
Convolutional Layers: Convolutional layers are often used as the initial component in the encoder. These layers utilize filters to identify features such as edges, textures, and patterns in the input data. Each convolutional layer builds upon the information obtained from the previous layer, allowing the model to learn progressively more complex attributes.
Activation Function (ReLU): After applying each convolutional layer, an activation function, typically ReLU (Rectified Linear Unit), is used. ReLU introduces non-linearity into the network, enabling it to understand complex data relationships.
Pooling Layers: Pooling layers, often max-pooling layers, typically reduce the spatial dimensions of the feature maps while preserving essential information. These layers help to reduce computational requirements and decrease the risk of overfitting.
Feature compression: As encoders evolve, the number of channels (depth) increases as the characteristic map’s spatial dimension shrinks. This feature compression allows the network to collect hierarchical and abstract input data representation.
The decoder, also known as the expansion path, works with the encoder to reconstruct the segmented output. The decoder gradually increases the spatial decision of the characteristic map. The process operates through the following steps:
Transposed Convolutional Layers (Up-convolution): The decoder uses transposed convolution layers known as up-convolution or deconvolution to increase its spatial resolution and upscale feature map. These layers recover the original input image.
Skip Connections: The incorporation of skip connections plays a pivotal role in U-Net and several other CNN architectures. Skip connections establish connections between layers in the encoder and decoder. By doing so, the decoder gains access to the high-resolution feature maps of the encoder, which serve as a crucial spatial data source. Skip connections are indispensable in preserving intricate details and ensuring accurate segmentation.
Output Layer: The final step involves generating the segmentation mask, typically performed by the decoder. Depending on the task, this layer may use either a SoftMax or sigmoid activation function to produce a probability for each pixel, indicating its association with different classes or segments.
where received neuron output, output level output number layer.
Horses in social organizations are separated into two categories: non-protective and protective. The public grouping, connection and grazing, mating performance, headship hierarchy, and control of these two sorts of organizations are different. However, the focus of the WHO approach is on non-territorial groups, which consist of group leaders (stallions), various mares, and their offspring, including young horses. Stallions are positioned near mares for contact, and mating can transpire whenever. Foals typically begin grazing within the initial week of existence and increase their grazing while reducing relaxes while they grow grown-up. Before reaching puberty, Male horses join solitary groups, while foals go from their parent groups to reach maturity for mating. In non-territorial horses that roam freely, the dominant mare usually assumes the role of the family group’s leader, with the remainder of the team following in a descending hierarchy of power, and typically, the stallion is positioned a little distance behind the assembly. In the study, they optimize many problems and lay out a WHO via group behaviors, grazing, mating, dominance, and leadership. Figure 3 explains the proposed flowchart of AdaResU-net with WHO.
Population Initialization: First, they divide this initial population into several groups. It is the number of members of the population, the number of groups is
is the proportion of stallions in all the people.
Behavior of grazing: To incorporate grazing behavior, the approach designates the stallion as the focal point of the pasture, while the group members explore the surroundings in search of food. To simulate this behavior, they introduced Eq. (1). Using Eq. (1), the group members are encouraged to explore the vicinity of the leader, each within a distinct radius.
is the group member’s present position (foal if mare otherwise). is the stallion’s (group head’s) position. is a computed adaptive mechanism using Eq. (3). is an even random number within the range . The pi value of 3.14 is equivalent to the horses grazing at various angles (360°) relative to the group leader,
is a vector with the problem’s lengths between 0 and 1., are uniformly distributed random vectors in the range is a uniformly distributed random digit in the range [0,1]. random vector index vector returns that assure the condition . is an adaptive parameter that has a starting value of and decreases following Eq. (4) as the algorithm is executed, concluding the algorithm’s execution 0.
Horse mating behavior: Horses exhibit a unique behavior compared to other animals, involving the separation of young horses from the herd to facilitate companionship. Before reaching puberty, foals leave the main group; male foals join a separate solitary horse group, while female foals join a second family group to mature and find a mate. This division is essential to prevent mating between fathers and siblings or offspring. To implement this behavior, a specific procedure is followed: a foal from one group leaves and joins a temporary group, and a foal from another group also joins this temporary group. When these two foals reach adolescence, they can mate if they are unrelated and of different genders. The resulting offspring then leave the temporary group and join another group. All different horse groups undergo this cycle of leaving, mating, and reproducing again. To simulate this departure and mating behavior, Eq. (4) has been proposed, which is equivalent to the Crossover operator of the mean type.
is the location of the horse within the group, and when it leaves the group, it is replaced by a horse whose breeder(s) also left the group and has reached maturity. These horses lack any familial ties, yet they have successfully mated and given birth to offspring. The foal’s position is determined by its departure from the group and its subsequent mating with the horse in position within the group, which then leaves the group .
The group leader is responsible for guiding the group to a suitable location, defined as the water hole. The group must move toward this water hole, and other groups also head in the same direction. The leaders compete to control this water hole, with only the dominant group allowed to use it while other groups are prohibited until the dominant group vacates. The group leaders must guide their groups to the water hole and use it if they dominate. However, if another group gains dominance, they must retreat. To achieve this, it is suggested to use Eq. (5) to estimate the gap and contact.
where the leader of the next position , represents the water hole position, denotes the present place of the leader of the group, and represents an adaptive mechanism by Eq. (6). is a uniform arbitrary integer in the range .
Exchange and selection leaders: Initially, leaders are chosen randomly to maintain the algorithm’s inherent randomness. With the algorithm’s advancement, the selection of leaders is based on their fitness. If a group member exhibits superior fitness compared to the present leader, both the role of the leader and the corresponding member’s position will be altered following Eq. (11).
The cyst in the ovarian images has been successfully fragmented through the implementation of the AdaResU-net design. The organization is finely tuned by the WHO algorithm through the acquisition of the optimal configuration. Additionally, the hyperparameters of AdaResU-net are optimized by solving the objective function DLC and WCE.
The hyperparameters of AdaResU-net are improved by addressing the beneath objective function
where addresses the genuine conveyance of the sample, addresses the circulation anticipated by the sample, and denotes the weight loss function.
The classification procedure employs the Pyramidal Dilated Convolutional (PDC) network to classify cysts into types such as Endometrioid cyst, mucinous cystadenoma, follicular, dermoid, corpus luteum, and hemorrhagic cyst. This network uses a reduced feature set to enhance the accuracy of input images and generate improved images with optimal features.
Dilated convolution is a technique that modifies the standard convolution kernel by introducing gaps to increase its receptive field. The dilation rate, which determines the size of these gaps, must be manually specified. For a 3 × 3 convolution kernel, different dilation rates affect the receptive field size. When the dilation rate is set to 1, the dilated convolution kernel is identical to the original basic convolutional layer. However, increasing the dilation rate to 2 expands the 3 × 3 convolution kernel to resemble a 5 × 5 convolution kernel. During this dilation process, the gaps are filled with zeros, except at the central position as depicted in Fig. 6 (b). This results in a receptive field of 5 × 5 for the convolution operation. By using dilated convolution, the receptive field is increased without changing the integer of parameters or using pooling. This guarantees that the amount of the output feature map remains the same, while also incorporating multi-scale information in the convolutional output. In Fig. 7 the symbols Dilate1 - Dilate3 represent the dilated convolution kernels, while Conv1 - Conv4 represent the common convolution kernels. The symbol {∙} represents concatenated algorithms, and F1 - F3 shows the various speeds at which the outputs from dilated convolution are produced. Based on these representations, the output Y can be expressed using the following formula, assuming X as the input:
The PDC structure utilizes dilated convolution by varying dilation rates to expand the receiving area devoid of the need for pooling. Moreover, the pyramid arrangement effectively combines information from diverse receptive fields, thereby enhancing the network’s performance. This dilated pyramid module emulates the functioning of the human eye, which amalgamates features at different scales when observing an object. Similarly, the component for pyramid dilated convolution merges the output from distinct dilated convolutional blocks with different degrees of dilation, mimicking the human eye’s process to some extent.
