Disjoint Brain Tumor Segmentation Method for Uncertainty Reduction Using Fully Connected Learning Network

Rajendran et al. [16] developed a brain tumor MRI image segmentation automatically using deep learning. Convolutional Neural Networks, widely used in the field of biomedical image segmentation, demonstrated a considerable improvement in brain tumor segmentation accuracy when compared to the present state of the art. It is used to deconstruct images into their constituent elements. The developed method achieves the mean accuracy and sensitivity for the whole tumor.

Tejashwini et al. [17] designed an automatic brain tumor segmentation method using MRI images. A biomedical image analysis technique is employed in the method to analyze the brain features for segmentation. The method minimizes the computational cost and complexity ratio. The designed method elevates the sensitivity, specificity, and accuracy rate of the process.

Hernandez-Gutierrez et al. [18] introduced a lightweight U-Net-based brain tumor segmentation model. The model is used to locate the exact location and condition of a brain tumor. The model uses optimal MRI image slices to detect the types and classes of brain tumors for further disease diagnosis. The introduced model elevates the accuracy, precision, and sensitivity range.

Rabby et al. [19] developed a multi-task architecture model using MRI images for brain tumor segmentation and classification. The model uses a deep learning (DL) algorithm to localize and optimize the important features from MRI images. The model also reduces the latency and error rate of the classification process. Experimental results show that the developed model achieves high precision and accuracy.

Qin et al. [20] proposed a diffusion probability model-enabled brain tumor segmentation model. MRI images provide reliable information to detect and segment the tumor-infected regions. The important features of the tumors are extracted using the diffusion method, which minimizes the computational cost and latency rate. The proposed model elevates the accuracy rate of the brain tumor segmentation process.

Li et al. [21] developed an enhancement of the robustness of brain tumor segmentation with region-based evidential deep learning to measure uncertainty. The BraTS 2020 is used in the dataset for both quantitative and qualitative studies to assess our model's performance in segmentation and uncertainty estimation. The developed method was performed in terms of robustly segmenting tumors and assessing segmentation uncertainty. The developed method reduces the computing cost. A lightweight 3D attention U-Net model was developed by Alwadee et al. [22] as an improved version of Hernandez-Gutierrez et al. [18]. The model uses the attention mechanism to select the feasible features of the tumor from MRI images. The model minimizes the computational cost and latency by analyzing the features in the images. When compared with others, the developed model enhances the classification and segmentation services for disease diagnosis.

Sun et al. [23] designed a multi-view attention and multi-scale feature interaction method for brain tumor segmentation. The method analyzes global and local features of brain tumors from MRI images. The method selectively extracts the reliable features for the brain tumor segmentation process. It also eliminates noises and unwanted details from the dataset. The designed method improves the accuracy and precision rate of the process.

Chen et al. [24] introduced a U-Net-based Kolmogorov-Arnold network (KAN) model for brain tumor segmentation. The model uses a pyramid feature aggregation module to fuse the features from MRI images. The fused features are used as input, which decreases the latency rate of the process. The introduced U-KAN model enlarges the precision level of the segmentation process. Liu et al. [25] proposed a 3D auto-calibrated focus U-Net for segmenting brain tumors. SCAU-Net replaces the original convolution layers with multiple 3D self-calibrated convolution modules, which adaptively computes the receptive field of tumor images for efficient segmentation. It also embeds the external attention into the skip connection to better utilize encoding features for semantic up-sampling. The proposed model reaches exceptional performance.

Mostafa et al. [26] suggested a method for detecting brain tumors using MRI data that combines segmentation and feature fusion. Segmentation and feature fusion have been used to provide a novel and reliable automated brain tumor detector. Noninvasive MRI has been widely used for diagnosis without the need for ionizing radiation. The suggested method enhances the accuracy and precision. Zhang et al. [27] developed a deep fusion of multi-modal characteristics for the segmentation of brain tumor images. The developed method makes full use of the multi-modality information included in a deep convolutional neural network to improve brain tumor image segmentation by extracting and combining unique. It is tested on the BraTS2021 data set. The developed method enhances the diagnosis and treatment of brain tumors.

Qureshi et al. [28] proposed a robust multi-class brain tumor segmentation framework using a DL algorithm. MRI images are used here as input, which produce feasible data for detection and segmentation services. The proposed framework eliminates the noisy features from the dataset, which enhances the reliability of the process. The framework achieves a high precision and accuracy rate in the process.

Rutoh et al. [29] developed a 3D guided attention-based deep inception residual U-Net (GAIR-U-Net) model for brain tumor segmentation. The developed model uses MRI images as input, which gathers sufficient data for the segmentation process. The model identifies and analyzes the infected brain tumor regions from given images. Experimental results show that the developed model enhances the sensitivity, specificity, and precision level.

Liu et al. [30] introduced an enhanced feature-based vision patch transformer network for brain tumor segmentation. MRI images are employed here, which are used as input for segmentation. The model CNN algorithm is used to extract optimal features and factors from MRI images. The extracted features are used to segment the exact location of the brain tumor. The introduced model maximizes the precision level of tumor segmentation.

Conventional segmentation methods are stuck into uncertainties due to overlapping pixels and region misidentification. These issues are addressed in references [17, 22] using identified feature learning, with high time demands. The misidentification uncertainty is addressed using probabilistic methods as in references [19, 26, 30]. Distinguishable methods proposed in references [20, 25, 28] are useful in improving the efficiency regardless of the evidence and encoding-based segmentation. The reverse problem is the dissimilar feature extraction and disjoints region segregation from the complete pixel distribution preventing multiple errors. Therefore, this article introduced a disjoint segmentation method using fully connected learning network for sorting these issues.

In contrast to the current methodologies, which rely solely on probabilistic models or on feature-learning paradigms, the proposed method incorporates the two elements in its Fully Connected Layer Network (FCLN) architecture. This integration helps to have a stronger control of uncertainties and possible misidentifications. The distinguishing characteristic of the approach is that it focuses on the classification of the disjointed regions and finding a solution to the inverse task of dissimilar feature extraction. The algorithm (separating non-overlapping regions and those that form maximal disjoint regions) provides the algorithm with a greater level of accuracy in tumor segmentation. Also, the approach provides a unique way of managing the uncertainty through unanimous uncertainty detection. Once both parallel processes agree on a value of uncertainty, the system stops any additional recursive operations hence optimality in feature identification and segmentation performance. Overall, the presented methodology provides an all-encompassing approach to the problem of brain tumor segmentation through the combination of the disjoint region analysis and the uncertainty reduction along with the iterative learning in the form of FCLN architecture. Such a comprehensive view stands out as a distinctive feature within the given methods of analysis, which generally concentrate on one or two of these aspects.

Unlike existing uncertainty-aware segmentation techniques that rely primarily on probabilistic modelling or confidence propagation, the proposed DSM-FCLN reformulates uncertainty handling as a structural decision process based on disjoint region separation. The framework introduces a parallel consensus mechanism, where disjoint feature extraction and uncertainty estimation operate recursively until a unanimous uncertainty threshold is reachedan approach absent in prior Fully Connected Network-based segmentation or evidential learning architectures. This design allows uncertainty reduction to emerge from feature separation rather than post-processing estimation.

The fully connected learning network is used to detect uncertainty and disjoint feature detection. Here, similar and dissimilar regions are segmented which is based on different brain objects such as fluids, tissues, etc. The proposed work focuses on similar and dissimilar regions and the output is grouped for the concurrent training. In this state, the detection is performed from the determination of uncertainty. Here, the disjoint features and uncertainty are done by evaluating FCLN where the computation is performed for the disjoint region. The following equation is used to detect the disjoint and uncertainty using FCLN.

Unlike prior fully connected segmentation pipelines, the proposed FCLN incorporates a recursive unanimous-uncertainty stopping rule and a two-stream feature pathway, ensuring that segmentation refinement continues only when both uncertainty estimation and disjoint-region learning converge to an identical value.

$\begin{gathered}\emptyset=\mathrm{l}_0+\beta * \sum_\pi(\rho+\tau) * \Pi_{\mathrm{m}_0}^{\mathrm{g}^{\prime}} \mathrm{e}^{\prime}+\left(\partial+\mathrm{m}_{\mathrm{n}} / \mathrm{g}^{\prime}+\right.\nabla)-\left[\left(\mathrm{e}^{\prime}+\mathrm{U}\right) * \tau\right]+\mathrm{a}_0\end{gathered}$             (5)

The detection is done for the uncertainty and disjoint features for which the FCLN is used. To operationalize these equations during training, the computed uncertainty value UUU and the similarity grouping outputs $S_{\text {sim }}, S_{\text {dis }}$ are passed into the learning network as supervisory signals. The loss function incorporates disjoint-region error minimization and uncertainty reduction, enabling the perceptron to update weights www until the unanimous uncertainty criterion Φ(U)=0 is met. This mechanism ensures that the model does not only segment the tumor boundaries but also progressively suppresses ambiguous boundary behavior during optimization. The overlapping of pixels is detected and from that classification is performed based on the region. The classification of pixels is done for the independent and disjoint. Here, the detection is done for the disjoint region segmentation, the necessary features are extracted. The necessary feature is extracted and from that uncertainty is evaluated and it is denoted as $\left[\left(\mathrm{e}^{\prime}+\mathrm{U}\right) * \tau\right]+\mathrm{a}_0$. The independent and disjoint region is segmented and from that the splitting is done for the n-number of the image and it is represented as $\left(\partial+\mathrm{m}_{\mathrm{n}} / \mathrm{g}^{\prime}+\nabla\right)$. The detection is done for the disjoint feature that relies on the similar and dissimilar regions where the uncertainty is estimated from the overlapping of pixels. The detection is done for brain tumor segmentation using FCLN. Thus, the features are associated with the segmentation of similar and dissimilar regions. From this perceptron is used for weight assigning in FCLN. A perceptron is used in the FCLN to classify the number of neurons and based on the neuron the training set is improved. The following equation is used to evaluate the perceptron and weight is assigned.

$\mathrm{C}=\mathrm{f}\left(\mathrm{m}_{0, \ldots \mathrm{n}}, \mathrm{h}_{0, \ldots \mathrm{n}}\right) * \mathrm{n}_{0, \ldots \mathrm{~m}}+\left(\mathrm{l}_{\mathrm{n}} * \emptyset\right) * \omega^{\prime}+\omega_0$            (6)

The perceptron C is used to assign the weight $\mathrm{h}_0$ for the number of input images. In this perceptron, weight is assigned to improve the training set in the neural network. Here, the detection is done to find the disjoint features and it is determined from the similar and the dissimilar regions. The similar region with uncertainty is trained and the dissimilar is also trained on the number of neurons $n_{0, \ldots m}$. The function is defined as f based on the weight the perceptron training is used for better a segmentation phase. The following equation is used to monitor the similar and dissimilar regions and the output is grouped for concurrent learning.

$\begin{gathered}\mathrm{M}=\frac{\mathrm{g}^{\prime} * \pi}{\prod_{\nabla} \mathrm{l}_{\mathrm{n}}+\beta}+\prod \mathrm{C}+\left(\partial * \mathrm{e}^{\prime}+\mathrm{U} /(\rho+\tau)\right) * \omega_0-n_{\mathrm{m}}+\frac{\beta}{\left(\Omega_{\mathrm{n}} * \emptyset\right)}\end{gathered}$            (7)

The monitoring M is done for the similar and dissimilar regions and the output is grouped for concurrent learning. The learning network is used to train the uncertainty value in the proposed work. This part of monitoring is done recurrently for the identification of similar and dissimilar regions. These regions are detected based on splitting the number of pixels and finding the overlapping. This identification of overlapping is examined in Eq. (1), and from this periodic monitoring is done for similar and dissimilar regions for better training output.

Here, the features are extracted and the uncertainty is derived by the determination method $\left(\partial * \mathrm{e}^{\prime}+\mathrm{U} /(\rho+\tau)\right)$. From the monitoring phase, the grouping of similar and dissimilar region output is evaluated for concurrent learning. Post to this method the hidden layer is used for the computation and training of the uncertainty value. The following equation is used for the computation and improvement of the segmentation rate.

$\left.\begin{array}{c}m_0\left(e^{\prime}\right)=\prod_{\nabla}^{g^{\prime}}(U+\emptyset) * l_0+n_0 * \frac{\sum_\tau(C+\pi)}{\beta+\mathrm{s}_0} \\ m_1\left(e^{\prime}\right)=\prod_{\nabla}^{g^{\prime}}(U+\emptyset) * l_1+n_1 * \frac{\sum_\tau(C+\pi)}{\beta+\mathrm{s}_0} \\ \vdots \\ m_n\left(e^{\prime}\right)=\prod_{\nabla}^{g^{\prime}}(U+\emptyset) * l_{n-1}+n_{m-1} * \frac{\sum_\tau(C+\pi)}{\beta+\mathrm{s}_{\Omega}}\end{array}\right\}$             (8)

The hidden layer is used to train the uncertainty for the number of images; here the n-number of pixels is detected for the overlapping. The overlapping of pixels is denoted as the disjoint and it is estimated to find the similar and dissimilar region. The region-based detection is done for the independent and disjoint regions these are done by using FCLN. In the above equation, the perceptron is used to assign weight to the pixels in the MRI. Based on this processing the number of MRI is fed to the neuron by assigning weight. The weights are detected to the m-number of neurons in the network and improve the accuracy level. The training is used until the unanimous uncertainty value $S_0$ is identified. The identification is performed based on the perceptron where the disjoint region is done. In this work, every image feature is extracted and finds whether overlapping exists or not. If there is overlapping or uncertainty is detected the output from the first layer neuron is trained and forwarded to the second layer. The FCLN process for uncertainty detection is illustrated in Figure 4.

Figure 4. Uncertainty detection using FCLN

Figure 4 illustrates the uncertainty detection process that employs a Fully Connected Learning Network (FCLN). This network has three main layers namely, input layer, hidden layer and the output layer. Two kinds of data are fed into the input layer; disjoint features and uncertainty. The processing of these inputs in the hidden layer is then done where the mapping process takes place. Three outcomes are possible in the hidden layer; when $s_o=\tau$, the output will be $s_o$, when $s_o \neq \tau$, the output will be $\tau$, and when both $s_o$ and $\tau$ are present, both of them will be mapped. The processing of the hidden layer gives the final result of the output layer. When the $w^{\prime}(U) \forall \partial_o$ is less than $\tau$, it means that the area is highly disjointed as compared to when $s_o>\tau$. This is done in a recursive manner and the training process is repeated until the disjoint areas are reduced to the minimum. Reduction of the disjoint regions implies that the uncertainty that is witnessed has been tackled or minimized. The FCLN requires $(\rho, \tau)$ inputs for detecting $s_o$ and $\pi(\tau)$ through two different processes. In the first process (i.e.) the hidden layer, the $m_n$ mapping with $\phi$ or $M$ or both are performed. If $m_n$ matches $\phi$ then $\partial_o$ are the output else ∇ is the required output. Here, $\partial_o$ is the $l_o$ to $l_{n-1}$ mapping for which $\pi\left(s_o\right)$ is extracted. This is trained as $w^{\prime}(U) \forall \partial_o$ only such that new outputs are detected. If $\nabla$ the function is the process, then $\pi(\tau)$ is trained from the splitting function until $l_{n-1}$ is achieved. Finally, if $s_o>\tau$ then the disjoint regions are high otherwise, it is loss. The training is pursued until the disjoint regions are less (i.e.) the $\pi$ observed is less for either $s_o$ or $\tau$ or both (Figure 4). The processing is carried out until there is no uncertainty is identified in this process. To improve this perceptron is estimated for every neuron in the network to provide better identification. From this hidden layer, the training is performed and post to this uncertainty detection is done for the disjoint feature. From this training phase, similar and dissimilar training is carried out in the FCLN and it is equated in the below Eq. (9).

$\omega_0, \omega^{\prime}(U)=\emptyset * \frac{1}{n_{m-1}}+\sum_{\nabla}(M+\beta) * \partial-s_0+e^{\prime}$            (9)

The neurons with the respective weights are assigned and perform the detection of similar and dissimilar region detection. The analysis is done to detect whether there is uncertainty for every input image and from that the detected image which are overlapping is separated. The separated image is trained along with the weights of the neurons in the network and determines the uncertainty. The uncertainty detection is performed until there is no overlapping and thus, the segmentation rate is improved. The classification phase is done for the independent and disjoint in the MRI. Thus, the detection is performed for every fixed interval and finds the uncertainty; from this, the unanimous uncertainty value is detected to decrease uncertainty. The following equation is used to state the unanimous uncertainty in FCLN.

$\pi\left(\mathrm{s}_0\right)=\mathrm{g}^{\prime}(\sigma) * \mathrm{l}_{\mathrm{n}}+\left(\frac{\sum \mathrm{u} \nabla+\emptyset}{\mathrm{c} * \mathrm{~m}_0}\right)$            (10)

The unanimous uncertainty is detected for the input image and determines the better segmentation. The perceptron is used to examine the better pixel identification and from that brain tumor is detected. Brain tumor detection is done by evaluating the splitting of images and from that training is distributed. The training phase is used to estimate the better detection of brain tumors. By performing this less uncertainty is estimated for the disjoint region. From this derivation, less uncertainty is detected in Eq. (10). The proposed method to uncertainty identification and reduction of brain tumor segmentation can be used to identify and address uncertainty issues through the use of a few major equations. The uncertainty in Eq. (2) is determined in terms of pixel overlapping. The feature region is differentiated with the use of uncertainty based on feature regions of Eq. (3). The given integrated method enables to fully evaluate the uncertainty in the entire feature space. The uncertainty and disjoints features are identified in the Eq. (5) that combines pixel classification, feature extraction, and image splitting to offer a powerful uncertainty detection mechanism. In coming up with unanimous uncertainty in the FCLN, Eq. (10) employs the minimum of the values of various neurons. This is to guarantee that the most conservative measure of the uncertainty is taken into account. The suggested approach involves training the FCLN repeatedly until reaching an acceptable and minimum unanimous uncertainty level that guarantees the best performance. Also, parallel training on similar and different regions is imposed to improve the network to differentiate between the various types of tissues. Lastly, there is the separation of the non-overlapping regions and maximum disjoint regions which further narrows the segmentation procedure and minimizes the uncertainty. From this segmentation is carried out on two categories are derived one is identifying maximum disjoint region and the other is segregation. The following equations are used to derive the maximum disjoint region and segregation is equated.

$\begin{aligned} \pi(\tau)= & \frac{1}{\mathrm{~m}_{\mathrm{n}}} *\left[\sum_\beta \partial+\mathrm{g}^{\prime} *(\mathrm{U}+\mathrm{M})\right] *\left(\mathrm{e}^{\prime}+\sigma\right) * \\ & \varphi-\left[\left(\sum_{\mathrm{g}^{\prime}}^{\mathrm{l}_0} \omega_0+\omega^{\prime}\right)\right] * \mathrm{C}-\frac{\mathrm{m}_0}{\sigma+\mathrm{s}_0}\end{aligned}$              (11a)

$\mathrm{G}=\beta(\rho+\tau) * \mathrm{~g}^{\prime}+\frac{\bar{\nabla}}{\sum \emptyset\left(\sigma * \mathrm{~m}_n\right)} * \omega_0, \omega^{\prime}(U)$             (11b)

In the above equations, maximum disjoint region and segregation are done to improve the segmentation rate in the proposed work. Eq. (11a) states the maximum disjoint region identification that is performed by using segmentation and it is denoted as $\left[\Sigma_\beta \partial+\mathrm{g}^{\prime} *(\mathrm{U}+\mathrm{M})\right] *\left(\mathrm{e}^{\prime}+\sigma\right) * \varphi$. In this, the equation similar and dissimilar is evaluated based on the identification of disjoint in the region. The necessary features are extracted and from that, the segmentation is done. Eq. (11b) states the segregation G where the evaluation is carried out to analyze the important feature. Figure 5 presents the segregation process illustration.

Figure 5. Segregation process illustrations

The $s_o>\tau$ condition identifies multiple $l_o$ to $l_{n-1}$ regions across various $\nabla$ processes. In this case the $\partial_o$ based variations (i.e.) $l_n$ distinct from $I_o$ to $I_n$ are extracted for $M$. The odd case of $m_n\left(e^{\prime}\right)$ is another demand for $G$ process from $\phi$ detection. The case of $w^{\prime}$ and $w_o$ are independent of $\nabla$ (new) and $G$ process between $\rho$ and $\tau$ pixels. Therefore the regions are optimal for detecting $C$ (allocated) through the $\pi\left(s_o\right)$ and $\pi(\tau)$ classifications. This initiates $w^{\prime}$ output segregation from $I_o$ to $I_n$ regardless of $L_{n-1}$ for $G$ (Figure 5). The important features are extracted and segregated for better detection of the tumor region. This equation is derived from Eq. (9) includes similar and dissimilar features in the region and is represented as $\frac{\nabla}{\sum \emptyset\left(\sigma * \mathrm{~m}_{\mathrm{n}}\right)} * \omega_0, \omega^{\prime}(U)$. Thus, the uncertainty along with the similar and dissimilar features are split and segregated. By computing this precision is improved by validating Eq. (12).

$\eta=\frac{1}{\left(m_n+l_n\right)} *(\varphi+G) * \beta$            (12)

In the above Eq. (12) the precision is improved by determining segmentation and segregation for the number of images. From the number of images, the pixel identifies the better extraction of features. In this evaluation, the precision is improved when the segmentation results in better identification of disjoint and uncertainty. This segregation is followed up to provide better detection of brain tumors. This processing is done by FCLN along with the segmentation and detection.

The performance assessment is validated using the following metrics: precision, segmentation rate, uncertainty, detection time, and region detection. This assessment is performed as a comparative analysis by changing the number of regions (1 to 10) and feature extraction rates (0.1 to 1). The existing methods EDLF (Evidential Deep Learning Framework) [21], SCAU-Net (Self-Calibrated Attention U-Net) [25], and ASBTCNN (Automated Segmentation of Brain Tumor using CNN) [16] are paired with the proposed methods in this comparative performance assessment.

To ensure robustness, the proposed DSM-FCLN framework and baseline models were trained across five independent runs with varying initialization seeds (42, 77, 101, 128, and 256). All reported values are presented as mean ± standard deviation, and 95% confidence intervals were computed in accordance with model-to-model variation. This statistical reporting approach reflects stability across repeated trials rather than a single execution outcome. To further evaluate stability and sensitivity to initialization, the multi-run results were analyzed using variance-based sensitivity scoring. The proposed DSM-FCLN demonstrated low run-to-run fluctuation, with performance variation remaining within ±1.4% for segmentation rate and ±0.9% for precision. A paired t-test comparing the proposed model against the strongest baseline (SCAU-Net) confirmed that improvements were statistically significant (p<0.05). The narrow confidence intervals indicate that the observed performance gains are not incidental or seed-dependent but remain consistent across repeated training iterations. To ensure fair comparison, all baselines (EDLF, SCAU-Net, and ASBTCNN) were re-trained under identical experimental conditions. The same dataset split (70% training, 15% validation, 15% testing), preprocessing steps, and augmentation policies were applied consistently across all models. Training was standardized to 8 epochs, using the Adam optimizer with a learning rate of 0.001, batch size of 16, and controlled seed initialization (seed=42) to minimize stochastic variation. No model-specific tuning advantage was applied, and hyperparameter settings were aligned to prevent bias in model performance. This ensures that the reported improvements stem from methodological advantages rather than differences in training configuration as shown Table 4.

Table 4. Training configuration consistency across models

In Figure 7, the precision for the proposed work increases by identifying the similar and dissimilar regions. The feature extraction is performed based on the classification process that includes independent and disjoint. The precision is increased by determining the segmentation of images. The image segmentation is done by evaluating the pixels and decreasing the uncertainty and it is represented as $\sum_{\mathrm{a}_0}^{\mathrm{g}^{\prime}}\left[\phi * \mathrm{~m}_0\right]+\mathrm{l}_0-\tau$. In this number of pixels are detected from the respective regions. The region-based detection is done for the number of images and determining the disjoint. Thus, the proposed precision is improved in determining the segmentation. Eq. (3), states the uncertainty and decreases overlapping of pixels. The classification is carried out by splitting the independent and disjoint regions. The feature extraction is done from the MRI and the processing is done for the number of pixels and identifies the overlapping.

Figure 7. Precision analysis

In Figure 9, the uncertainty decreases by identifying the disjoint from the classification method. The maximum disjoint is identified by performing the segmentation method and it is represented as $\partial * \mathrm{e}^{\prime}+\mathrm{U} /(\rho+\tau)$. The computation is done for the detection of brain tumors in MRI. From this processing, the uncertainty is defined by extracting the necessary features from the input region. The desired features are extracted and split as independent and disjoint regions. Here, the uncertainty is defined by assigning the weights for the number of neurons in the connected network. The FCNL is proposed to decrease the uncertainty and detect the brain tumor by addressing the overlapping of pixels. The overlapping of pixels and uncertainty is estimated by equating Eq. (9). The processing is examined by improving the computation process by introducing the single hidden layer that is used to train similar and dissimilar regions.

Figure 9. Uncertainty analysis

In Figure 11, the region detection is high in the proposed work by determining the uncertainty. The uncertainty value for the proposed work shows better results that are based on similar and dissimilar identification. The classification phase is used to distinguish the independent and disjoint regions. Eq. (10) is used to identify the unanimous uncertainty value. The detection is performed for the region extraction and it is represented as $\left[\left(\sum_{\mathrm{g}^{\prime}}^{\mathrm{l}_0} \omega_0+\omega^{\prime}\right)\right] * \mathrm{C}-\frac{\mathrm{m}_0}{\sigma+\mathrm{s}_0}$. The segregation is done from the segmentation process and shows better region detection and it is represented as $\beta(\rho+\tau) * \mathrm{~g}^{\prime}$.

Figure 11. Region detection analysis

The appropriate detection of region is analyzed from the splitting of regions. The feature extraction is done by assigning the weights for the number of neurons by determining perceptron. The hidden layers train the error pixel from the first layer and perform better region detection. Thus, the region detection is performed for better feature extraction. In the below Tables 5 and 6 below, the above study is summarized with the improvements of the proposed method compared to the existing methods.

Table 5. Comparative study summary for regions

Table 6. Comparative study summary for feature extraction rate

The proposed method achieves the following: 7.09% more precision, 9.82% more segmentation rate, 9.89% more region detection, 6.86% less uncertainty, and 12.84% less detection time.

The proposed method achieves the following: 7.29% more precision, 10.39% more segmentation rate, 9.17% more region detection, 6.56% less uncertainty, and 9.62% less detection time.