Precision Measurement and Feature Selection in Medical Diagnostics using Hybrid Genetic Algorithm and Support Vector Machine

Breast cancer [1] is one of the most prevalent and consequential diseases in the world today, affecting millions of people, especially women. Breast cancer is not just a health problem or a tumor [2], but a critical disease where early detection and necessary treatment are especially important. In the early stages, breast cancer does not show any significant symptoms [3], but only small lesions on the skin, swelling in the breast or armpit, which are highly unlikely to be detected. In modern times, advanced methods of treating breast cancer have developed, but these treatments require an adequate system for early detection. According to a report published by the World Health Organization (WHO), 2.3 million new cases of breast cancer were registered in 2024, of which 93 % of patients are not aware of their infection in the earlier stages, as the disease does not show any decisive symptoms.

The three conventional methods for detecting breast cancer are physical examination, mammography, and biopsy. Breast cancer is detected using imaging techniques such as mammography [4], magnetic resonance imaging (MRI) [5] or 3-D mammography. Mammography images are a type of X-ray images [6] of the human breast.

Advances in technology and the incorporation of machine learning (ML) algorithms are helping and contributing to the detection of breast cancer at an earlier stage. Some of the well-known ML algorithms are k-nearest neighbor (KNN) algorithm [7], decision tree (DT) algorithm [8], logistic regression (LR), Naive Bayes (NB) algorithm [9], and random forest (RF) classifier [10]. The biggest challenge with the existing methods for early detection of breast cancer is the imbalance dataset. The self-generated imbalance dataset consists of a large number of features that may be essential and redundant and interfere with the classification process.

On the other hand, it is not only painful but also expensive, invasive, and time-consuming [7]. It is imperative that improved methods are used to detect breast cancer [8]. It has been mentioned that the chances of successfully treating breast cancer are significantly higher if the disease is detected at an earlier stage [9]. The optimization of feature selection from medical data is achieved by using a hybrid GA, which ultimately leads to an improvement in the accuracy of detection and classification models [10]. The main contributions of this proposed work are as follows:

The novel framework performs an optimized feature selection and supports a robust classification process.

The manuscript is organized into an analysis of recent related work in Section 2 and a detailed description of the proposed work in Section 3. The performance analysis and comparative analysis are presented in Section 4 with conclusive remarks in Section 5.

Basaad et al. (2024) used the characteristics of graph neural networks (GNN) [11] for the detection of breast cancer. The accuracy of this model is reported to be 83.16 %. Supriya et al. (2024) proposed a traditional ML and deep learning (DL) model for breast cancer prediction using the federated learning (FL) method [12]. The framework used the Wisconsin diagnostic breast cancer (WDBC) dataset for training the model and for testing purposes. The accuracy of this model was measured to be 94.73 %. Chen et al. (2024) proposed a modality specific information disentanglement (MoSID) method [13] for the earlier prediction of breast cancer. The major drawback of this model is that the MRI image can only be used for certain women and not for all diseased women.

Furtney et al. (2023) have developed a model for breast cancer detection using the multi-relational directed graph method [14]. This method accepts the MRI image of the patient and evaluates the features using relational graph convolutional neural networks for detecting probabilities of molecular subtypes. Wang et al. (2023) proposed a novel method of dynamic contrast enhanced magnetic resonance (DCE-MR) imaging [15] for breast cancer detection using MRI images. This model suppressed the excessive false negative results of other modern methods and showed an accuracy of 89.61 %. Panigrahi et al. (2024) used the ML algorithm using the minimum redundancy maximum relevance (MRMR) method [16] for feature selection. The model includes four different classifiers namely support vector machine (SVM), decision tree, multilayer perceptron and RF to achieve higher accuracy. The accuracy of this model was measured to be 94.09 % and the computational time was improved.

Thakur et al. (2023) proposed a hybrid model of convolutional neural networks (CNN) and recurrent neural networks (RNN) [17] for detection of cancer in multiple body parts, namely breast, kidney, uterus, etc. This model used VGG-19 and VGG-16 models for training the model and achieved an accuracy of 85.31 %. The main challenge with this model is the large amount of data in the dataset, which leads to dataset imbalance and overfitting. Almaslukh et al. (2024) proposed a computer-aided diagnosis model that incorporates the DL method [18] for early detection of breast cancer. The DL approach used the random search algorithm together with the DenseNet-121 transfer model and achieved an accuracy of 96.42 %.

David et al. (2024) showed a lower effectiveness when using CNN) [19], which may be a consequence of problems related to overfitting. Duan et al. (2024) investigated breast cancer using LASSO regression analysis and classification was performed using a hybrid SVM, LR. This hybrid model showed an accuracy of 93.6 % and an F-score of 88.9 %. The objectives of the proposed model are:

To optimize feature selection in breast cancer detection and improve prediction performance.

The proposed work focuses on a multi-class classification problem for breast cancer detection, distinguishing between Normal, cancer stage I, II, and III based on mammogram images. The study proposes a hybrid model that integrates a genetic algorithm (GA) for optimal feature selection and a SVM as the primary classifier with a Bucket of Models (BoM) ensemble approach to improve classification accuracy.

The proposed detection of early stage breast cancer using GA and SVM is triple folded. In the proposed method, an ensemble strategy is used for breast cancer diagnosis. There are two ensembles. One is for model selection and the other is for attribute selection. An SVM classifier is used to implement this ensemble strategy. Another GA is specifically designed to select the attributes from the models selected by the previous genetic algorithm. The proposed model consists of three stages. The model selection is done in the first stage. This model generates ten models as output. These outputs are responsible for classification. The GA is used for selecting the BoM technique. The classification of the model is done based on the ML technique.

The best configuration can also be selected based on the combination of the models. One such technique is the BoM. The possible model configurations are given by different forms. They are the base classifier and the set of parameters for these base classifiers. The BoM selects the model that fits best from the other available models. Governing the BoM is done using the genetic algorithm. It generates new populations. These populations are nothing more than a group of individuals. Fig. 1 shows the individual steps of the BoM method.

Fig. 1.

Steps involved in the BoM method.

The best classifier is selected and then the parameters are defined. Based on the parameters, the possible models are selected and sent for evaluation. The reason for using the GA in the BoM methodology and feature selection is that it does not require much data for evaluation. The first process is the pre-processing stage which consists of several steps, namely image resizing, intensity normalization, noise reduction, grayscale conversion and data augmentation. In the proposed work, no image cropping or RoI extraction is performed but the entire image is used as image cropping would lead to loss of minute features.

Feature selection using genetic algorithm

The result of the previous steps would be the best models selected to fit the problem of classification. The proposed model does not spend much time searching for regions that are not certain, saving the resources and thus minimizing the time required for computation as presented in Fig. 2.

Fig. 2.

Proposed breast cancer detection using GA and SVM.

The GA generates a population of potential chromosomes representing a subset of features related to breast cancer. The chromosomes were selected by assigning a gene value of 1, and the non-selected chromosome is labeled with a gene value of 0. The features are defined as shown in (1). (1) p={f1,f2,f3,…fn} p = \{{f_1},{f_2},{f_3}, \ldots {f_n}\} where i represents the feature f_i with 1 as selected and f_i = 0 as non-selected chromosome. The fitness function is defined as in (2). (2) S=arg maxf f(p) S = arg\,\mathop {\max}\limits_f \,f(p) where S is the feature subset, and f(p) is the fitness function of the features. The fitness function f(p) depends on the error rate as defined in (3). (3) f(p)=11+E(x) f(p) = {1 \over {1 + E(x)}} where E(x) is the error rate and a minimum error rate is required for an improved fitness function. One of the most common methods of feature selection is the roulette wheel selection method, in which chromosomes are selected with a probability proportional to the fitness of the chromosome, as defined in (4). (4) p(f)=f(c)∑i=1Nfi(c) p(f) = {{f(c)} \over {\sum\nolimits_{i = 1}^N {{f_i}(c)}}} where p(f) is the probability for the selected features and f(c) is the fitness of chromosome i, while N is the total number of chromosomes. The selection of the parent chromosomes is followed by the crossover function as defined in (5) and (6). (5) p1(c)={p11(c);p21(c);p31(c);…pn1(c)} {p^1}(c) = \{p_1^1(c);p_2^1(c);p_3^1(c); \ldots p_n^1(c)\} (6) p2(c)={p12(c);p22(c);p32(c);…pn2(c)} {p^2}(c) = \{p_1^2(c);p_2^2(c);p_3^2(c); \ldots p_n^2(c)\}

The crossover leads to the production of offspring with new chromosomes as defined in (7) and (8). (7) O1(c)={p11(c);p21(c);p31(c);…pk1(c);pk+12(c);pk+22(c);pk+32(c);…pn2(c)} {O^1}(c) = \left\{{\matrix{{p_1^1(c);p_2^1(c);p_3^1(c); \ldots} \hfill \cr {p_k^1(c);p_{k + 1}^2(c);p_{k + 2}^2(c);p_{k + 3}^2(c); \ldots p_n^2(c)} \hfill \cr}} \right\} (8) O2(c)={p12(c);p22(c);p32(c);…pk2(c);pk+11(c);pk+21(c);pk+31(c);…pn1(c)} {O^2}(c) = \left\{{\matrix{{p_1^2(c);p_2^2(c);p_3^2(c); \ldots} \hfill \cr {p_k^2(c);p_{k + 1}^1(c);p_{k + 2}^1(c);p_{k + 3}^1(c); \ldots p_n^1(c)} \hfill \cr}} \right\}

The crossover function is followed by the core function of the mutation process. The mutation process randomly flips the bits of the chromosomes and is represented by the m_i defined in (9). (9) mi={1; if p(c)=0 and mutation occurs0; if p(c)=1 and mutation occurs {m_i} = \left\{{\matrix{{1;\,\,{\rm{if}}\,p(c) = 0\,{\rm{and}}\,{\rm{mutation}}\,{\rm{occurs}}} \hfill \cr {0;\,\,{\rm{if}}\,p(c) = 1\,{\rm{and}}\,{\rm{mutation}}\,{\rm{occurs}}} \hfill \cr}} \right.

This process is repeated until the condition converges and the fitness function is stabilized. The performance of the GA in selecting the feature is improved by implementing the adaptive capability in the genetic algorithm. The adaptive process in the GA is defined in (10). (10) Pbest={Pmax×(Pc+PiPc+Pavg); if Pi<PavgPmin×(Pi+PcPi+Pavg); if Pi<Pavg {P_{\rm best}} = \left\{{\matrix{{{P_{\max}} \times \left({{{{P_c} + {P_i}} \over {{P_c} + {P_{\rm avg}}}}} \right);\,\,\,if\,{P_i} < {P_{\rm avg}}} \hfill \cr {{P_{\min}} \times \left({{{{P_i} + {P_c}} \over {{P_i} + {P_{\rm avg}}}}} \right);\,\,\,if\,{P_i} < {P_{\rm avg}}} \hfill \cr}} \right. where, P_i is the initial fitness of the parent, P_c is the fitness of the current solution, and P_avg is the average fitness of the parent.

In contrast to the existing methods that rely on manual feature selection, the proposed work introduces a two-level ensemble strategy that combines a GA and a BoM for model selection. Fractal analysis is performed during the feature extraction to identify and measure the complexity and irregularities of breast tissue patterns.

The proposed work was trained and tested on the Kaggle dataset – breast ultrasound images dataset (BUID) [23], which consists of 10200 mammogram images with breast cancer. Of these 10200 images, 80 % (8160) were used for training, while the remaining 20 % (2040) were used for the testing process. The proposed method is intended for mammogram classification and the performance of the ML technique was improved by using the GA for feature selection. The performance of the proposed method was calculated by evaluating parameters such as accuracy, precision, recall and F1 score. All experiments were performed on a system with an Intel Core i7 processor (3.6 GHz), 32 GB RAM, and an NVIDIA GeForce RTX 3060 GPU (12 GB VRAM) running Windows 10 (64-bit). The implementation was carried out using Python 3.8 with key libraries such as Scikit-learn, NumPy, and OpenCV.

True positive is the number of positive instances that were correctly identified as positives. True negative is the number of negative instances that were correctly identified as negative instances. False negative is the number of negative instances that were incorrectly identified as positives. False positive is the number of negative instances that were incorrectly identified as negatives. Table 1 shows the selection models for cancer detection based on the area under the curve. The 8LTP and the 3LTP are the local ternary pattern (LTP) used for texture analysis in the image processing and classification application.

The 2LTP + Fractals + GLCM have a much lower inter-mediate and final selection values compared to the previous methods. Table 2 shows the performance evaluation of the different feature groups.

The GLCM technique has an F1 score of 95.36 %, an accuracy and sensitivity of 95.62 % and a specificity of 95.88 % and is depicted in Fig. 3.

Fig. 3.

Graphical representation of the performance evaluation of the different feature groups.

Table 3 contains the training dataset performance evaluation based on the number of genes. Different numbers of genes are considered for calculating the performance measures: 5502, 4096, 2048, 1024, 512, 256, 128, 64, 32 and 16.

When the number of genes is 5502, the accuracy value is 0.91 and when the number of genes is reduced to 16, the accuracy value is 0.75.

The highest recall value is achieved when the number of genes is 512. The values for the second highest recall value and the second lowest recall value are 0.89 and 0.65, respectively.

Table 4 shows the testing dataset performance evaluation as a function of the number of genes. The accuracy value is highest when the number of genes is higher. When the number of genes is 5502, the accuracy value is 0.83 and when the number of genes is reduced to 16, the accuracy value is 0.70. The highest recall value of 0.76 is reached when the number of genes is 2048, 1024 and 128. The second highest recall value of 0.68 is achieved when the number of genes is 5502, 256, 64 and 16. The F1 score values of 0.43, 0.49, 0.48 and 0.38 are obtained when the number of the genes is 512, 256, 128 and 64, respectively.

Fig. 4 shows the graphical representation of the training dataset performance evaluation as a function of the number of genes. Fig. 5 shows the graphical representation of the testing dataset performance evaluation as a function of the number of genes.

Fig. 4.

Graphical representation of the training dataset performance evaluation based on the number of genes.

Fig. 5.

Graphical representation of the testing dataset performance evaluation based on the number of genes.

Table 5 contains the 15 most important types of genes for differentiating breast cancer. The designations such as 6q26.2-q26.3 stand for the number of the chromosome and the region in which the gene is located on this chromosome.

Table 6 shows the performance comparison of the proposed GA in combination with information gain and information ratio for different classifiers. The harmonic mean of precision and recall is called F1 score.

For the SVM classifier, the accuracy has improved from 53.59 % for all features to 85.56 % for IG-GA. Recall has increased from 51.00 % to 85.35 % with IG-GA and precision has improved from 27.30 % to 85.70 %.

The accuracy of the NB method has increased from 49.46 % for all features to 63.90 % for IGR-GA. Recall increased from 47.94 % to 61.87 % and precision improved from 46.32 % to 80.32 %.

Accuracy [26] and recall [27] in the KNN classifier improved from 55.68 % to 86.62 % and from 55.44 % to 86.99 %, respectively.

Fig. 6 shows the graphical representation of the performance comparison of SVM [28] and NB classifiers. Fig. 7 shows the graphical representation of the performance comparison of KNN, DT and RF classifiers.

Fig. 6.

Graphical representation of the performance comparison of SVM and the NB classifiers.

Fig. 7.

Graphical representation of the performance comparison of KNN, DT, and RF classifiers.

Table 7 shows the number of selected features before and after applying the GA with different classifiers. The classifier performance is affected by the feature selection methods such as the binary particle swarm optimization (BPSO), information gain (IG), IG-GA, information gain ratio (IGR) and IGR-GA. The number of features after applying IGR-GA is 625, 605, 614, 624 and 619 in the case of SVM, NB, KNN, DT, and RF, respectively. The calculated accuracy values are around 617.

Table 8 shows the comparison of the percentage of accuracy values of the proposed method with the existing methods. The genetic improved SVM recursive feature elimination (GI-SVM-RFE) method has achieved an accuracy of 91 % using the RF classifier.

The performance of the proposed work is tested on different mammogram images of infected and normal breast tissue. The results are shown in Table 9.

From Table 9, it is clear that the proposed GA + SVM model is able to detect and classify breast cancer more effectively than the existing works. The main reason for the improved accuracy is the dataset balance provided by the genetic algorithm. GA efficiently reduces the number of features, which is crucial for high-dimensional datasets such as gene expressions and mammography images.

Sample images were used as a test image for the proposed model and the result was categorized into four classes, namely Normal cell, Cancer – stage I, II, and III. Training images 1 and 2 were assigned to stage II, while image 3 is assigned to stage I in the initial stage. Image 5 is widespread and is assigned to stage III, while image 4 is a normal cell.

The classification into stages I, II and III was based on the following information on the infected tissue.

Stage I: Tumor is small in size (< 2 cm) and no lymph node involvement.

Breast cancer is one of the main types of cancer affecting many women. It is the unwanted growth of cells in the breast that leads to a tumor. In the proposed method, a genetic algorithm-based feature selection technique has been used in combination with BoM methodology for feature selection. These methods are used for selecting the best features among the different features from the collected datasets. The proposed GA showed the best performance when it was tested with the information gain and information ratio gain filtering techniques. The effectiveness of the proposed model was evaluated based on various performance metrics. Different classifiers were tested with the proposed GA and the results were analyzed.