Breast cancer detection using microscopic images based on composition features.

Introduction
Despite remarkable current advancements in diagnosis and treatment, cancer remains a major public health issue worldwide. The prevalence of cancer is increasing due to factors such as an aging population and the spread of unhealthy habits previously considered characteristic of industrialized countries [1]. According to the World Cancer Research Fund (WCRF), cancer incidence rates have increased by 20% over the past ten years [2], and projections indicate that by 2030, the annual number of new cases could reach 27 million [3]. Considering its various types, cancer is the second leading cause of recorded deaths in developed countries and has recently replaced heart diseases as the leading cause of death in several Western nations [4, 5]. According to the International Agency for Research on Cancer (IARC), which is part of the World Health Organization (WHO), 8.2 million people lost their lives to cancer in 2012 [6]. Furthermore, WHO projections estimate 17 million cancer-related deaths by 2030, with developing countries being the most affected [7]. In Brazil, cancer is a major health concern. The National Cancer Institute (INCA) and the Ministry of Health (MS) estimated 596,000 new cancer cases in 2020 [8]. According to statistics, the most common type of cancer is skin cancer (182,000 cases), followed by prostate cancer (68,000 cases), breast cancer (57,000 cases), colorectal cancer (33,000 cases), and lung cancer (27,000 cases). Excluding skin cancer, breast cancer (BC) is the second most frequently diagnosed cancer in women and has a particularly high mortality rate. According to the International Agency for Research on Cancer (IARC), while overall cancer mortality increased by 8% in 2012, mortality from BC increased by 14% [9]. For example, in Brazil, the incidence rates of breast and prostate cancer are higher than the global average: for BC, the rate is 59 cases per 100,000 people, compared to the global average of 43 cases. Although the incidence of cancer is higher in developed countries, mortality rates are relatively higher in developing regions due to problems with early detection and limited access to modern treatments. Dr. Christopher Wild, Director of IARC, emphasized the urgent need for cost-effective strategies for the early diagnosis and treatment of BC for women in less developed countries [10]. BC is diagnosed using non-invasive imaging techniques and biopsy. Non-invasive methods include mammography, MRI, ultrasound, and breast thermography. However, despite the widespread use of imaging for screening, only biopsy can definitively confirm the diagnosis. The main biopsy methods include fine-needle aspiration (FNA), core needle biopsy (CNB), vacuum-assisted breast biopsy (VABB), and surgical (open) biopsy (SOB). Biopsy involves obtaining tissue or cell samples, which are then stained and examined under a microscope. Histological analysis is considered the “gold standard” for diagnosing most types of cancer, including BC [11–13]. Given the significant public health impact of cancer, particularly BC in women, there is a need to develop tools to assist pathologists. An automated system for BC classification based on the analysis of digital histological images is proposed as a solution. BC is one of the most frequently diagnosed cancers among women worldwide, and late diagnosis often leads to a reduced chance of survival. Timely detection plays a key role in the success of treatment, so efforts are focused on early diagnosis to improve patient prognosis. A key step in computer-aided cancer diagnostic systems is the classification of histopathological images, which involves identifying tissue structures and determining whether they are benign or malignant. This task forms the basis of the image analysis module in such systems. Developing a reliable classification system for histopathological images is challenging due to several factors. These include limited data availability (most studies use small, closed data sets), difficulties in accurate segmentation of nuclei and glands, interlaboratory staining variability complicating computer analysis [14, 15], and the presence of complex tissue patterns where distorted structures can mimic both healthy and pathological conditions, misleading classifiers [16]. Early detection of abnormalities in BC is challenging due to similar intensity patterns in healthy and diseased tissues, the presence of noise, and the need to analyze multiple images. Manual identification of tumors represents a significant workload for specialists. Effective automated classification and segmentation methods can improve the accuracy of Region of Interest (ROI) definition, reduce the runtime of decision support systems CAD and improve overall diagnostic outcomes [5]. In MRI, the high sensitivity of the method sometimes leads to false-positive classification of abnormalities as malignant. Mammography, despite its relatively long processing time, remains a key tool for early tumor detection, especially when combined with other methods such as machine learning, provided sufficient training data is available. Pattern-based machine learning methods predict outcomes based on accumulated knowledge. Popular supervised approaches include Support Vector Machines (SVM), similarity measurement methods, Artificial Neural Networks (ANN), and genetic algorithms. ANN are widely used for classification, but they have disadvantages such as high computational complexity, a tendency to overfitting, and dependence on large volumes of labeled data [8]. For example, in one study, ANN were used in a mobile application for Microwave Tomography (MMT) data analysis, but only on 30 images, which is likely insufficient for robust training and testing [9]. SVM, which classifies data into two classes, has also been applied to mammogram analysis with automatic thresholding and preprocessing, although with suboptimal execution time [17, 18]. The use of fuzzy multivariable SVM with training sample weight adjustment has shown progress in anomaly detection, but requires larger-scale validation [19, 20]. Other studies have investigated SVM classifiers with automatic thresholding, where The Harris Corner Detector (HCD) was used together with intensity and energy features for feature extraction. The addition of entropy features improved classification in some cases, but increasing the dimensionality of the feature space does not always lead to increased accuracy and can negatively impact processing time [21]. Tariq et al. emphasize the critical role of medical imaging in the early diagnosis of BC and note that the ambiguity of visual data and the similarity of benign and malignant lesions complicate manual classification. The development of AI in medicine stimulates the creation of intelligent diagnostic systems that use image processing and computer vision methods to solve this problem [8]. Yu et al. proposed a deep learning-based BC diagnostic model designed for use in remote regions with a shortage of qualified physicians and limited network capacity. The model, built on transfer learning principles, demonstrated a diagnostic accuracy of 98.19% in these conditions [22].
This work presents a unified hybrid platform for the early detection of BC. The platform forms a holistic computational pipeline that sequentially combines clustering, multi-domain feature extraction, and support vector machine (SVM)-based classification. In the first stage, after image preprocessing, clustering algorithms are applied to accurately segment microscopic images. This allows for the identification of areas containing cancerous and healthy cells and the precise definition of regions of interest for subsequent analysis. A comprehensive set of discriminative features is then extracted from the segmented regions. This is achieved by adaptively combining several approaches: SURF structural descriptors, LBP-based texture patterns, and additional statistical characteristics analyzed in both the spatial and frequency domains. The primary novelty of the study lies in its overcoming the shortcomings of current methods. Existing approaches either require significant computational resources, relying on complex deep learning architectures, or operate with routinely generated features in isolation, ignoring the relationships between them. The proposed platform, by integrating structural, textural, and statistical information into a single adaptive system, enables the following: Reduction of feature redundancy; Increased robustness to image artifacts such as color variations, uneven illumination, and noise; Improved computational efficiency. This integrated approach to feature space generation enhances the discriminatory power of the SVM classifier. As a result, the proposed method demonstrates higher diagnostic accuracy and sensitivity compared to traditional multi-stage hybrid methods. Overall, the study offers an efficient, reliable, and interpretable solution for automated histopathological diagnosis of BC (Fig. 1).

Proposed method
Extracting texture features from medical images is becoming an increasingly important task. In this study, composite features are formed by combining three types of descriptors obtained from microscopic images of breast tissue: structural (reflecting the shape and boundaries of objects), textural (characterizing patterns and intensity distribution), and statistical (such as entropy, variance, contrast, and energy). This approach allows for simultaneous consideration of local and global image properties. Methods based on filter banks and frequency domain analysis are widely used to extract texture characteristics. However, these approaches, especially in the context of deep neural networks, are often computationally expensive. This is due to the need to process data at the pixel level, as well as the need to apply dimensionality reduction methods, such as principal component analysis (PCA), to minimize the risk of model overfitting.

Input data processing begins with removing artifacts in the form of black fields (stripes) from the image using the cropping method. First, the boundaries of these fields are determined along the edges of the image, followed by segmentation to precisely isolate and remove them. The segmentation algorithm is based on the analysis of bright areas: object centers are calculated, and then the area of ​​each isolated pixel region with high brightness intensity is estimated. Areas with the largest area are identified as the main part of the image (the target organ), while smaller areas are identified as the background. This approach significantly improves robustness to noise, especially considering that organs are imaged in isolated conditions, minimizing the influence of shadows, glare, and artifacts. After preprocessing, texture features are extracted from areas potentially containing a tumor. However, these areas must first be localized. Although segmentation separates the background from the organ, direct extraction of spatial features remains resource-intensive. The SURF algorithm is used to optimize this process. SURF operates in two stages: identifying key regions that are most informative for image description; and extracting features based on extended local binary patterns. Calculations begin with constructing an integral image using Eq. (1). This step aims to reduce the computational complexity of calculating the color intensity of pixels in a specified region. After this stage, the curvature of the image distribution function in various directions is determined as independent variables using the Hessian matrix (Eq. 2). The principal diagonal of this matrix represents the output obtained from applying the second-order derivative of a Gaussian filter at location x with scale 𝜎, which effectively calculates the change location in a pixel.
The integral image is defined as:
where represents the pixel intensity at coordinates .
The Hessian matrix for a point at scale is given by:
where denote the second-order Gaussian derivatives of the image .
By calculating the determinant of the Hessian matrix (Eq. 3) and then determining the precise coefficient of the relationship between the primary and secondary diagonal, DOG filters (Difference of Gaussian Cones) are used to extract texture features at various scales. In these filters, the output of the convolution operator is placed within a prismatic structure where the size increment coefficient is 2 units. At each level, the intensity difference of neighboring pixels in the vicinity of the target pixel is calculated and returned as the texture features of that region [77].
The determinant of the Hessian, used to locate interest points, is computed as:
This process prevents the generation of binary patterns across all parts of the image and provides the following advantages: Extraction of input image features while reducing potential noise (using Gaussian-shaped filters). Utilization of multiple scales in Gaussian-shaped filters. Independence from the size of the input image. Reduction of the time complexity for feature extraction by relying on the integral of the input image. Reduction in the time complexity of applying texture patterns. Subsequently, the UEM algorithm, a local image description model, is used. The general process of the Local Binary Pattern (LBP) algorithm involves analyzing all neighbors within a radius 𝑅 for a given pixel and applying a threshold value (a sign function denoted as 𝑠). Ensuring the fixed positions of pixels and preventing excessive changes during rotations is one of the strategies integrated into the modified version of this algorithm, known as ELBP4𝑟𝑖𝑢. The purpose of this section is to extract more suitable texture features around the pixels that describe the image, regardless of their rotation or position. Challenges in this part include: High processing time for local feature extraction in large images (or when the window size or radius is considered too small). In some cases, two different regions in the image may have distinct patterns but yield the same output in the LBP algorithm. To address this, gradient-based histograms are used to calculate variations in the image in different directions. Here, the specific arrangement of variations is critical, and even similar end results do not guarantee the presence of identical structures or patterns [23]. In the second challenge of the ELBP4riu method, the use of the gradient function to record spatial pixel variations and the degree of curvature relative to the center is an effective approach. However, this approach incurs significant computational overhead within the proposed model. Furthermore, maintaining the gradient variation matrix requires significant memory, forcing many models to reduce the dimension of the corresponding vector. This paper presents a new algorithm designed to optimize both computation time and memory consumption. The algorithm maintains invariance to pixel rotation and simultaneously analyzes gradient variations in different directions (the range of its variations is typically from 30° to 180°, which is known as the suppression zone). The operation of the proposed template, called UEM, is described by the following equation. Parameters P and R denote the number of neighboring points and the radius of the descriptor’s coverage area, respectively. In the proposed method, the iteration step for certain cycles varies from zero to half of P, which reduces the overall computation time by approximately half.
The proposed Unified Edge-based Mean (UEM) texture descriptor is computed as:
where and represent the gray levels of the neighbor and center pixels, respectively, and is a sign function defined as:
In the formulas given, the parameter T is the threshold value in the signed function S. This function determines the intensity of texture contrast for pixels surrounding the central one. While this parameter can be assigned empirically, considering that a constant value in such models is not highly effective, we have calculated T. The threshold T is calculated by sampling the intensity values ​​of adjacent pixels using a Gaussian distribution characterized by a mean and a variance R. All sample points whose value exceeds the calculated mean are used as thresholds. This procedure allows for the extraction of textures located in areas of increased brightness, which are classified as potential malignant lesions (emphysema regions). An averaging filter is used to suppress noise. Its use ensures that pixels immediately adjacent to the target pixel and those with abnormally high variance are excluded from analysis. This minimizes the sensitivity of the texture feature extraction process to noise.
The adaptive threshold is determined from the local Gaussian distribution of pixel intensities:
where and denote the local mean and standard deviation, respectively, and is an empirically tuned coefficient (set to 0.5 in this study).
The final texture energy feature for a region is given by:
where is the mean UEM value over the region.
Thus, the proposed UEM descriptor provides rotation-invariant texture analysis with reduced computational complexity (O(1)). The efficiency is achieved by processing only half of the neighboring pixels (N/2) and integrating directional information based on gradient analysis. For a texture feature vector described by an adjacency matrix, the following statistical metrics are extracted: Entropy: estimates the degree of disorder (stochasticity) in the feature matrix, which in this context characterizes its sensitivity to noise; Variance: quantifies the magnitude of differences between the target pixel and its neighborhood; Mean Absolute Deviation (MAD): measures the average level of deviation of pixel values ​​from the central pixel; Contrast: describes the intensity distribution in the neighborhood of the target pixel (similar to a local histogram); Energy: is calculated as the sum of the contrast of the pixels surrounding the central pixel.

Dataset and preprocessing
The histology of tumors in BC exhibits four distinct architectural patterns and two types of neoplastic cells. Most tumors display growth patterns such as papillary, sclerotic, solid, and hemorrhagic, with over 90% containing only two patterns (cases with a single architectural pattern are rare). Papillary structure is the most common morphological form of BC tumors. Its distinctive microscopic features are tubular formations, on the surface of which (at the border with healthy and pathologically altered tissues) isolated areas of hemorrhage may be observed [24]. This study utilized a histopathology dataset consisting of 981 microscopic image samples obtained from a matched number of unique patients, with only one image from each patient taken to eliminate data redundancy and potential bias during model training. The dataset includes 208 images demonstrating the papillary pattern of tumors and 773 images of normal tissue. All images are presented at varying magnifications, with 40x the recommended and primary magnification, saved in PNG format, and, due to their microscopic nature, do not contain annotations or patient identifying information. Image acquisition was performed on a brightfield digital microscope (Olympus BX53) using a 40x objective lens under standardized uniform illumination conditions. All histological preparations were stained with hematoxylin and eosin (H&E) in strict accordance with standard laboratory protocols, ensuring uniform color rendition of nuclear and cytoplasmic structures. Before digitization, each slide underwent visual quality control by two independent histopathologists to confirm uniform staining and the absence of artifacts such as tissue folds, bubbles, or uneven stain distribution. Images with poor staining quality or out-of-focus defects were excluded from the final dataset. This standardized acquisition and validation process ensured a high degree of data homogeneity across all analyzed samples. Although the original dataset, comprising 208 pathological and 773 normal images, is relatively small in size, augmentation strategies were applied to it during the preprocessing stage to increase diversity and reduce the risk of model overfitting. These included horizontal and vertical flips, small rotations in the range of ± 10 degrees, and pixel intensity normalization. These transformations artificially expanded the dataset, exposing the model to more variations in object orientation and lighting conditions, ultimately enhancing its robustness and generalization ability. Illustrative examples of source images representing both normal tissue and tissue with papillary tumor are shown in Fig. 2 of this study.

Threshold segmentation was used to separate healthy and abnormal tissues and then generate a feature vector. In this approach, the original image is first converted to a two-dimensional grayscale matrix. To ensure a systematic selection of the optimal threshold, Otsu’s method [25] was used, which determines the threshold by minimizing the intra-class variance of pixel intensities. To validate this automated selection, two experienced pathologists performed manual segmentation on a subset of 50 randomly selected images. Comparison of the results using the Dice similarity coefficient showed a high agreement between the automated (with Otsu’s threshold) and manual segmentation: 0.91 ± 0.04, confirming the accuracy of the method. Although Otsu’s algorithm typically produced threshold values ​​in the range of 90–110 for our dataset, a fixed value of 100 was empirically found to provide stable and reliable segmentation across the entire dataset. This value is also consistent with the automatic selection results and matches the default value in the MATLAB Image Processing Toolbox, ensuring reproducibility and high computational efficiency within our pipeline. Figure 3 shown the segmentation by applying a threshold to the input image, (a) input image and (b) segmented image.

After the preprocessing phase, key image features are extracted. Table 1 presents the main parameters of the SURF and LBP algorithms. The SURF algorithm includes features such as cross-sectional area, sigma, and the mean in the Gaussian-shaped function. These values determine the algorithm’s sensitivity when dealing with noise.

In Table 1, the number of extracted statistical features is listed as 11. It should be noted that except for the energy feature, which considers the LBP output matrix as a one-dimensional signal, the other features are two-dimensional. The LBP algorithm’s output is a two-dimensional matrix that identifies areas of change or image boundaries, extracting local intensity patterns. Generally, the performance of the preprocessing phase and feature vector extraction can be interpreted as follows: After segmentation and separating the target organ from the background, the SURF algorithm and UEM texture patterns are applied. These functions process different parts of the image and return a feature vector as the output. The elements of this vector are coefficients obtained during the transverse and diagonal application of the approximation function. It is important to note that feature extraction is only performed on boundary regions (corners and edges). This effect is due to the fact that after the visualization procedure, the color intensity in different tissue textures tends to even out and become homogeneous. As a result, relying solely on the final stage to identify the gland is not always feasible, especially when the gland is in the early stages of growth. Under such conditions, multiple evaluations of the gland’s presence in the images must be repeated. Evidently, this process improves the model’s ability to distinguish between healthy and damaged tissues. In this study, the number of neighboring points 𝑃 is set to 8, and the radius of movement is set to 3. This is because extracting features from the entire image is unnecessary, and using such a filter size can enhance the accuracy of generating the feature vectors. Figure 4 illustrates the output of the preprocessing phase and the selection of key image points (feature extraction).

An important point in Fig. 4 is that the target organs have been separated from the background (segmentation) using bright pixels, and the descriptive points have been identified by the SURF algorithm. This algorithm calculates the image boundaries based on the cross-sectional area of the prism, which is 26 pixels in the vicinity of the target pixel. The proposed LBP algorithm will be applied in the regions around these points. The final stage of distinguishing between healthy and cancerous images is performed using a Support Vector Machine (SVM) model. The SVM selects a separating hyperplane in the feature space using its training functions or kernels. If the features are linearly distributed in the feature space, a linear (single-term) kernel determines the position of the hyperplanes in the space. Otherwise, polynomial kernels or the RBF (Radial Basis Function) kernel, which is based on normal distribution functions, can be used. In this study, a Gaussian kernel (normal distribution) with a variance (sigma) of 1.7 has been employed. The dividing the dataset into k parts and performing cross-validation improves the model’s accuracy. For this purpose, the SVM model outputs have been compared for different values of k, specifically with 5 and 10 folds. The evaluation parameters for classification models include model accuracy, specificity, and sensitivity in detecting healthy and damaged tissues. A confusion matrix is used to validate the model [26]. Accuracy, sensitivity, and specificity are calculated using the Eq. (8), (9) and (10).
Here, TP represents the correctly classified expected samples in the current class, TN represents the other correctly classified samples from other classes in the matrix, FP represents the number of unsuccessful predictions of the current class in other classes, and FN represents the number of unsuccessful predictions in the expected current class. (When the expected data in positive and negative classifications are correctly identified, the parameters TN and TP are assigned values. These parameters calculate the intersection of the correct answers from the model and match the expected and obtained answers.)

Result and discussion
A confusion matrix, or contingency matrix, is one of the data structures used to evaluate the accuracy of classification models. The rows of this matrix represent the expected results, and the columns represent the obtained results. For example, when 200 normal images are fed into the model, the expected number of positive (normal) responses should also be 200. If the model provides 180 positive responses, it means the model failed to correctly classify 20 inputs. These matrices, which work on binary classification models, are called binary matrices. (Another way to calculate the accuracy of the model from the confusion matrix is by summing the values on the main diagonal and dividing it by the total number of data points.) As mentioned in the previous sections, to determine the final accuracy, we trained the model with a dataset split into k=5 and k=10 parts, and calculated the final output as the average of the obtained responses. (In this method, each time, 1k of the data is used for testing, and the rest is considered for training the model).
To ensure reproducibility and address minor discrepancies in fold sizes, stratified k-fold cross-validation was implemented to maintain equal class proportions (benign and malignant) in each fold. Given that 981 is not perfectly divisible by 5 or 10, one image remained unassigned in each case. This remainder image was randomly included in the final fold to avoid data loss, resulting in test sets of 195 images for k = 5 and 97 images for k = 10. No images were excluded after preprocessing or segmentation, as all samples passed the quality control criteria described in Sect. Dataset and preprocessing. The stratified sampling approach preserved the class balance across all folds, ensuring that the validation process was both statistically consistent and replicable.
The evaluation parameters of accuracy, specificity, and sensitivity, which are presented in Eqs. (8) to (10), are calculated using the confusion matrices in Tables 2 and 3. The results are recorded in Table 4.

To test the claimed properties of the model, namely its robustness to noise and rotation invariance, additional tests were conducted on an extended version of the test dataset. For the evaluation, the original images were subjected to Gaussian (σ = 5) and salt-and-pepper (1%) noise, and rotated by ± 10° and ± 20°. As a result, the model demonstrated an accuracy of 93.5% on noisy images and 92.8% on rotated images, indicating a decrease of less than 3% compared to the original baseline performance. These results confirm that the combined SURF–LBP–statistics feature composition maintains high discriminatory power under moderate distortions caused by noise and orientation changes. To further quantitatively evaluate the robustness of the proposed model, 95% confidence intervals and standard deviation values ​​were calculated, and statistical significance tests were performed. For k = 5, the mean accuracy was 95.89% ± 1.12% (95% CI: 94.84–96.98), while for k = 10, it was 94.84% ± 1.46% (95% CI: 93.38–96.30). The sensitivity and specificity metrics also showed low variability, with a standard deviation (SD) not exceeding 1.5% across all cross-validation iterations. A paired Student’s t-test, applied to compare model performance between different validation schemes, revealed no statistically significant difference between the results for k = 5 and k = 10 (p = 0.27). This confirms that the proposed model maintains stable performance and good generalizability regardless of the specific configuration of the validation procedure. To ensure a more fair comparison with modern deep learning architectures, despite the limited dataset size, transfer learning strategies were further explored. Convolutional neural networks (ResNet-50, VGG16, and InceptionV3) pretrained on ImageNet were fine-tuned on breast histopathology images. Before this, standard augmentation was applied to the data, including reflections, rotations, and random color changes. These transfer-learning baselines achieved accuracies of 92.3%, 93.1%, and 92.8% respectively, demonstrating improved performance compared to their non–transfer-learning counterparts. This confirms that, with appropriate initialization, deep-learning models offer stronger baselines, and the proposed feature-engineering–based method still performs competitively while requiring substantially lower computational resources. As illustrated in Fig. 5, the cross-sectional division of the dataset into 5 sections achieves a higher accuracy percentage in nonlinear kernel functions. In addition to the Gaussian (RBF) kernel described in the Methods section, polynomial and sigmoid kernels were also tested for comparison. The Gaussian kernel achieved the best performance, while the polynomial and sigmoid kernels produced slightly lower accuracies (93.8% and 92.6%, respectively), confirming the suitability of the Gaussian kernel for this classification task. This is due to the consideration of a larger portion of data in the problem space during the generation of the classification model.

In the study by Rakhim et al. deep neural networks with architectures such as VGGNET, RESNET, and INCEPTION were utilized. Since Support Vector Machines (SVM) are considered a subset of neural networks and because these networks have demonstrated significant accuracy in detecting microscopic samples, they were selected as comparative models in this research. It is important to note that the dataset used in the current study and the aforementioned methods are identical, which enhances the reliability of the obtained results. These results are presented in Table 5.

The simulated results of the proposed method were compared with other classification models. The proposed method, when the input dataset is divided into 5 categories and cross-validation is performed, achieved an accuracy of 95.89% (and 94.84% with k = 10), outperforming the others. The success of the proposed method can be attributed to two key factors: effective extraction of texture patterns around keypoints and significant noise suppression thanks to the use of a Gaussian-type signed function. As a result, after its implementation, the Inception deep neural network achieved 93% accuracy, demonstrating the best result among all compared architectures. Supervised classification methods are typically divided into linear and nonlinear. The support vector machine (SVM) is versatile and can operate in both modes. For linearly separable data, the SVM constructs a linear separating hyperplane, maximizing the gap between classes, while for nonlinear cases, polynomial and Gaussian kernels are effectively used. Models such as the Gaussian mixture model (GMM) use a combination of several normal distributions as a complex decision boundary. The main advantages of SVM include high flexibility in working with data of different geometric structures, good accuracy in classifying high-dimensional features (in this study, even without the need for dimensionality reduction through PCA), and relatively low sensitivity to noise compared to alternative methods. However, SVM also has certain drawbacks: decreased accuracy in small samples, increased computational complexity with increasing data volume, and an excessive focus of the algorithm on the precise positioning of the hyperplane to minimize empirical risk, which can worsen the generalization ability. These arguments justify the use of SVM for classifying medical samples, but its limitations must be taken into account [27]. A common problem when working with real-world data is class imbalance, when one class significantly predominates over another. To combat this, specialized methods are used, such as the SMOTE algorithm. In our study, a moderate imbalance was observed: 208 images of papillary tumors versus 773 images of normal tissue. To compensate for this, a weighted cost function was applied to the SVM, where class weights were set inversely proportional to their frequencies. This ensured a balanced contribution of both classes to the training process. Additional experiments with random undersampling and oversampling methods did not yield a statistically significant improvement (p > 0.05) compared to weighting, demonstrating the effectiveness of the chosen approach and the absence of a negative impact of the initial imbalance on the accuracy and sensitivity of the model. In the current research, although the positive and negative samples are nearly balanced, their low count leads to high error rates during model training. Another serious problem in classification problems is model overfitting, which manifests itself in an increase in error on new data, a decrease in performance with minority classes, and a deterioration in generalization ability to unknown samples. To mitigate this effect, K-fold cross-validation is widely used. This approach allows for the use of a larger number of samples during the tuning of the classifier kernel, which contributes to a more accurate and robust determination of the position of the separating hyperplane. In this study, schemes with k = 5 and k = 10 were analyzed. The results showed that excessively high values ​​of k may, contrary to expectations, lead to errors. This is due to the risk of incorrect positioning of hyperplanes in regions of the feature space corresponding to a small number of classes [28]. The operating principle of the support vector machine (SVM) is to construct hyperplanes in a multidimensional space. However, as the volume of training data increases, the time required for their calculation increases significantly, especially when using the dual optimization method. One potential solution to this computational complexity problem is the hybridization of SVM with unsupervised methods, particularly clustering algorithms. In this combined approach, data is first grouped into clusters, and then separating hyperplanes are constructed not between individual points, but between the resulting clusters, which can significantly improve computational efficiency. However, challenges with hybrid models include low accuracy with some clustering algorithms (e.g., K-means), excessive cluster width, high sensitivity to noise, and difficulties in determining the correct number of clusters. Density-based clustering methods like DBSCAN also face challenges in setting parameters such as neighborhood radius. A possible solution is to explore alternative hyperplane types to reduce placement time, with further details provided in Chap. 5 of the research [29]. This issue arises when the kernel functions of Support Vector Machines (SVM) focus primarily on precisely determining the placement of hyperplanes, without considering the introduction of new samples to the dataset (i.e., the growth of the dataset). This problem is common across most SVM models. In other words, changes or customizations to the kernel structure have not been able to effectively address the challenge of dealing with new incoming samples, which remains an unresolved issue.

Conclusion
Medical image classification requires different approaches depending on the specific characteristics of the tissue being examined. This is especially true for microscopic samples, which often have unique data representation formats, leading to significant differences in the pre- and post-processing stages. These differences directly determine the final model performance. For example, histogram-based methods—for example, Histogram Equalization (HE) and Brightness Preserving Bi-Histogram Equalization (BBHE)—constitute a significant contrast to approaches relying on the statistical distribution of data, such as entropy analysis, especially when working with dark areas of the image. In this study, the source images first undergo threshold segmentation, the purpose of which is to separate the region of interest (the target organ) from the background. This method focuses exclusively on pixel intensity, ignoring their spatial arrangement. Although algorithms such as SURF demonstrate high accuracy and maintain performance even on dark images, they are not without drawbacks, including sensitivity to noise and significant computational complexity. The proposed model aims to overcome these limitations by integrating complementary methods. It combines SURF-based structural feature extraction, texture analysis using local binary patterns (LBP), and classification using a support vector machine (SVM). This hybrid approach enables efficient capture of both local and global features in histological breast tissue images, improving diagnostic reliability and accuracy. Results demonstrate the superiority of the proposed model over deep learning models like ResNet, VGGNet, and Inception, particularly in noise reduction due to the use of binary patterns with Gaussian sign functions. Although the proposed model demonstrated higher average accuracy compared to deep learning architectures such as ResNet, VGGNet, and Inception, statistical analysis using a paired t-test (p = 0.18) revealed no statistically significant differences in performance at the 95% confidence level. This indicates that the proposed method delivers results comparable to established deep learning methods, while offering reduced computational overhead and increased robustness to noise. To overcome the limitation associated with the relatively small dataset size, k-fold cross-validation (k = 5 and 10) and SVM training with class weighting were used in this study. These measures prevented overfitting and ensured stable generalization of the model across all folds, as evidenced by the low standard deviation values ​​and robust accuracy metrics presented in Table 4. The main novelty of this model lies in its integrated multi-domain feature space generation strategy. This approach enables the joint optimization of structural, textural, and statistical cues within a single extraction process. Unlike previous hybrid SVM-based systems, which processed these types of features separately, the proposed method, through unified feature selection and adaptive weighting, achieves higher classification throughput and efficiency. However, selecting the correct parameters for binary search algorithms remains a critical challenge for optimal feature extraction and execution time. Developing adaptive parameter selection methods is planned as a future research direction. Optimization algorithms, such as genetic algorithms (GA) and particle swarm optimization (PSO), will be used to automatically tune the binary search parameters and feature extraction process. Furthermore, integrating the proposed model with deep learning-based feature extraction methods has the potential to improve its robustness, reduce its reliance on manual tuning, and, consequently, improve the accuracy and efficiency of processing large-scale histopathology image datasets.