Comparative analysis of features and classification techniques in breast cancer detection for Biglycan biomarker images

1.Introduction

Breast cancer (BC) is considered the world’s most prevalent cancer [1]. BC begins to form in the ducts’ lining cells (epithelium) and 15% in the lobules of breast glandular tissue [2]. The malignant lesion is initially localised to the duct or lobule, where it often exhibits no symptoms and has a minimal chance of spreading (metastasis) [3]. If left untreated, these malignancies have the potential to spread to nearby lymph nodes (regional metastasis) and eventually to other body organs (distant metastasis); in the latter case, metastatic BC is the leading cause of death for women [3]. The World Health Organization (WHO) reported that 2.3 million women were diagnosed with BC and 685,000 women worldwide lost their lives through BC in 2020 [1].

Early diagnosis of BC enables patients to receive better care and treatment, hence -lowering patient mortality rates [3, 4, 5, 6, 7]. As a result, specialists have recommended screening methods to support early diagnosis through the evaluation of medical images [8, 9]. The most often used screening techniques are Breast Magnetic Resonance Imaging (BMRI), Positron Emission Tomography (PET) scan, Breast Ultrasound (BUS), Computed Tomography (CT) scan, Digital Mammogram (DM), Thermography, and Histopathological (HP) images [3]. Screening techniques produce medical images that help radiologists and doctors diagnose diseases, thereby lowering mortality risk by 30–70% [10]. However, early diagnosis of BC is highly challenging due to the huge amount of data and the poor imaging features of early breast cancer [7]. Moreover, masses are typically encompassed or surrounded by other structures including muscle, blood arteries, and healthy tissue [11]. Additionally, breast lesion identification and classification are challenging even for experienced radiologists since the lesions tend to have: (a) different shapes and distributions, (b) a small size, which differs from 0.1 to 1 mm, and (c) low contrast compared to normal breast tissue [12]. In this sense, digital technologies including image processing and machine learning methods are being developed to help radiologists in early and accurately diagnosing/classifying BC [4]. Different algorithms have been developed specifically for the detection and automatic classification of breast masses [13, 14, 15, 16, 17, 18]. Despite that the main steps of processing are the same, i.e., pre-processing, detection, and classification, each step can be implemented using several approaches [12].

A successful detection algorithm requires accurate segmentation where a breast abnormality can be fully identified and segmented [12]. Subsequently, the tumour and healthy tissue are categorised according to the extracted features [7]. Many features can be extracted such as statistical, intensity, texture, morphological, and shape features [12, 19]. Additionally, Loizidou et al. [12] pointed out that a single feature change can significantly improve or worsen the accuracy of classification methods. Therefore, they recommended that choosing the best feature combination is typically necessary and can be performed in a variety of techniques, each of which results in the selection of a distinct subset of characteristics.

Numerous advanced techniques have been developed for classifying data on breast cancer; some of these techniques involve feature selection, while others carry out the classification process without feature selection [20]. Modi and Ghanchi [21] compared different feature selection methods and the associated machine learning algorithms to classify data of the WBCD, WDBC, and WPBC datasets. Their study concluded that combining classification algorithms performs better on WBCD than the other two datasets. Hazra et al. [22] investigated finding the minimum number of features that guarantee a highly accurate classification of breast cancer. Following that, the study compared classification approaches including Naïve Bayes (NB), Support Vector Machine (SVM), and Ensemble classifiers. Their study concluded the highest accuracy is yielded when 19 features are fed into SVM (98.51%) while using five features and a NB classifier results in an accuracy of 97.39%. Asri et al. [23] compared the performance of different machine learning algorithms: k Nearest Neighbours (k-NN), Decision Tree (C4.5), SVM, and NB on the Wisconsin Breast Cancer (original) datasets. They pointed out that, SVM provided the best accuracy (97.13%) with the lowest error rate. Similarly, Bazazeh and Shubair [24] evaluated the performance of SVM, RF, and Bayesian Networks (BN) on Wisconsin original breast cancer dataset to evaluate and compare the performance of the three ML classifiers in terms of key parameters such as accuracy, recall, precision and area of ROC. Their results showed that BN has the best performance in terms of recall and precision while RF has the optimum ROC performance. Thus, RF had a higher chance of discriminating between malignant and benign cases. Silva Neto [18] developed a Convolutional Neural Network (CNN) model that distinguishes between non-cancerous and cancerous tissues based on Biglycan expression. The proposed model demonstrated an average classification accuracy exceeding 93% on histology images.

The early detection of breast cancer (BC) is essential in enhancing treatment quality and reducing mortality rates while also minimising global and national health risks. This in return contributes to fostering healthy lives and promotes well-being for individuals of all ages, aligning with the objectives of the Sustainable Development Goals, i.e., SDG3 “Good Health and Well-being” [25]. Thus, this research aims to investigate and compare various feature and classifier selections applied to histological breast images, aiming to enhance the detection and classification of BC images. Histopathology, which is considered the “gold standard” for cancer diagnosis and clinical decision-making, involves observing cellular morphological changes under a microscope in biopsy or surgical specimens made into tissue slides for disease diagnosis [26]. More specifically this work focuses on comparing and analysing the accuracy of classifying Biglycan histological images of breast tissues as cancerous and healthy (cancer-free) using:

Figure 1.

Schematic diagram of the proposed framework.

2.Materials and methods

This section explains the procedures employed for dataset acquisition, feature extraction techniques, classification techniques, and evaluation metrics utilised for assessing the performance. Figure 1 shows the schematic diagram of the proposed framework. Data processing was conducted on a computer equipped with an Intel Core i7 10th Gen processor, 6-Cores with NVIDIA^® GeForce^® GTX 1650 with 6GB, and running on the Windows 11 operating system. MATLAB (R2020b) software was used to extract features and to implement the classification algorithms.

Figure 2.

Samples of the Biglycan biomarker images for healthy and cancerous tissues [27].

2.4Classification measures

Classification metrics are used to assess the effectiveness of the breast cancer prediction model. These metrics are accuracy, sensitivity, specificity, and positive predictive value [6]. Providing only the accuracy can be misleading as high accuracy can still be obtained by combining low sensitivity and high specificity [12]. However, a high rate of false negative detections is linked to low sensitivity, which is unacceptable in clinical applications [12]. These metrics are explained as follows [6, 35].

2.4.1Accuracy

Accuracy (Ac) is defined as the percentage difference between the predicted synergy scores and the observed results within the permitted error range. It is defined as:

Ac=(𝑇𝑃+𝑇𝑁)(𝑇𝑃+𝑇𝑁+𝐹𝑁+𝐹𝑃),

where TP is the number of true positives, TN is the number of true negatives, FN is the number of false negatives, and FP is the number of false positives, respectively.

Figure 3.

Classification accuracy of the six classifiers using the histogram of the different colour spaces of the images, i.e., the greyscale, the three layers of the H, S, and V dimensions, and the three layers of the L*a*b.

2.4.2Sensitivity

Sensitivity (Sn) is a metric that evaluates the efficacy of breast cancer detection prediction models. It is also known as the rate of recognition. Sensitivity is calculated as:

Sn=𝑇𝑃𝑇𝑃+𝐹𝑁,

where TP is the number of true positives and FN is the number of false negatives.

2.4.3Specificity

The specificity (Sp) is expressed in terms of the true negative rate and is defined as the proportion of negative tuples that the prediction model correctly predicts. Specificity is calculated as:

Sp=𝑇𝑁𝑇𝑁+𝐹𝑃,

where TN is the number of true negatives and FP is the number of false positives.

2.4.4Positive predictive values

Positive predictive values (PPV) represent the percentage of appropriate instances among the recovered instances and are known as precision. PPV is calculated as:

PPV =𝑇𝑃𝑇𝑃+𝐹𝑃,

where TP is the number of true positives and FP is the number of false positives.

3.Results and discussion

The histogram distribution of the seven colour space dimensions of the images was extracted, this included the greyscale, the three layers of the H, S, and V dimensions, and the three layers of the L*a*b space. Further, the performances of the six classifiers for the histogram features were calculated as shown in Fig. 3.

Figure 3 shows that the greyscale and the S-dimension gave the best classification accuracy for the classifiers k-NN, SVM, and RF with an accuracy of 97.6%, 95.8%, and 95.3% for greyscale images, and 97.3%, 95.8%, and 95.6% for S-dimension images, respectively. These results are in agreement with Altunkeser and Körez [36], who recommended the use of greyscale images for evaluation of microcalcification; as improved detection of intramammary lymph nodes and microcalcifications can be obtained from greyscale images compared to standard ones. On the other hand, the V-dimension achieved the best accuracy using pl-DA, DT, and SVM classifiers with 89.0%, 95.0%, and 95.8%, respectively. This agrees with the review of Avcı and Karakaya [19] who pointed out that in the literature the SVM classifier generally gives higher classification accuracy in comparison with other methods. For all classifiers, the L-dimension did not classify correctly and just allocated all results to the cancer group. This could be explained by the fact that the L*a*b space comprises two colour channels, i.e., a* and b*, alongside one channel dedicated to luminance (lightness), denoted as L*. The L* channel represents the overall lighting conditions rather than the graduation in lighting, which constrains the information within the L-dimension [37]. The SVM classifier gave approximately similar classification accuracy (95.8%) for the greyscale and the three dimensions of HSV space. Further, both the k-NN and SVM classifiers achieved approximately similar accuracy among the seven features with an absolute accuracy difference of ⩽ 1.8%. Among the six classifiers, the NB and pl-DA classifiers resulted in the lowest accuracy ranging between 60.4%–89.0%.

Figure 4.

Classification accuracy of the six classifiers using the five features, i.e., histogram of greyscale images, Fast Fourier Transform (FFT), Harris Points, MinEigen Points, and Wavelet analysis.

As the histogram of greyscale images was the best feature for classification compared to the other colour dimensions (Fig. 3), the classification metrics for the greyscale feature were calculated to analyse its performance details, as shown in Table 2. Except for pl-DA classifier, the specificity (94.74% to 97.74) was higher than the sensitivity (92.12% to 97.54%) for all classifiers. This indicates that the classifiers can recognise healthy (cancer-free) images better than cancer images. The PPV values were ⩾ 96.39, which indicates the high precision of the classifiers.

Moreover, given that the histogram of grayscale images consistently outperformed the histograms of other colour dimensions across various classifiers, it is selected for further comparison with the remaining four features of the shape characteristics (i.e., Harris Points and Minimum Eigenvalue (MinEigen) Points) and frequency domain characteristics (i.e., The Two-dimensional Fourier Transform and the Wavelet Transform), see Fig. 4. Based on Fig. 4, the performance of the classifiers for the histogram of the greyscale image resulted in the best classification accuracy compared to the other four features with accuracy up to 97.6%. Except for the NB classifier, the wavelet feature achieved the second-best accuracy in all classifiers with an accuracy of up to 94.6%. While comparing the classifiers, the pl-DA and the NB classifiers showed the lowest classification accuracies (60.8%–87.5%) while the k-NN, SVM, and RF showed better classification accuracies (61.3% to 97.6%).

The combination of different features was investigated for different classification schemes, with no significant improvements in the overall performance. Even the all-feature option did not result in better performance. This indicated that many features had redundant values and did not contribute to any additional information.

Various studies investigated breast tissues and possible features and classifiers for the detection of breast abnormalities. Table 3 summarises the results of sample literature works along with the results of this work. Among these studies, the use of Biglycan biomarker images was helpful in the detection of breast cancer, resulting in 97.6% accuracy of detection using the k-NN classifiers.

4.Conclusion

While the accurate detection of breast cancer has been a challenge for physicians, extracting appropriate features from histopathology images along with using classification techniques are believed to facilitate the diagnosis process. In the current study, feature extraction and classification techniques were investigated and compared in diagnosing breast cancer. Five features were extracted and compared based on shape characteristics (i.e., Harris Points and Minimum Eigenvalue (MinEigen) Points), spatial characteristics (i.e., The Two-dimensional Fourier Transform and the Wavelet Transform), and statistical characteristics (i.e., histogram). Six different commonly used classification algorithms were used; i.e., K-nearest neighbours (k-NN), Naïve Bayes (NB), Pseudo-Linear Discriminate Analysis, Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF). These features and classifiers were compared with accuracy, sensitivity, and specificity being computed as part of the evaluation. The histogram of greyscale showed the best performance among other colour spaces as well as among the other four features. Thus, physicians can use this feature in breast cancer diagnosis. Interestingly, the wavelet feature provided a promising accuracy in most classifiers.