(1) In this study, some conventional machine learning techniques are proposed for more accurate breast cancer classification.

(2) The study additionally proposes a classical deep learning model for more accurate breast cancer classification. In this case, a softmax activation function is used to improve the training models and accuracy estimations.

(3) Machine learning and deep learning models are employed to detect benign, malignant, and normal masses.

(4) Instead of using single observation performance, this study applies 10-fold cross-validation for fair performance analysis. In addition, four standard evaluation metrics were used to assess the models’ performance.

(5) The study employs a combination of quantitative and qualitative techniques to provide many viewpoints on the subject issue of breast cancer classification.

(6) It is a fresh contribution to the literature in that it proposes new AI models for breast cancer classification.

It should be noted that this study adds to the body of knowledge on breast cancer classification. Despite numerous research have been conducted in the domain of breast cancer classification, this study gives an improvement with similarities and differences of the utilized state-of-the-art models. This study suggests utilizing conventional machine learning and classical deep learning approaches to classify breast cancer. Table 1 (Ref. [12, 13, 16, 17, 18, 19]) summarizes the proposed model’s differences and similarities with previous research.

The results in Table 1 show that the dataset used comprises non-binary breast ultrasound images; consequently, it contains benign, malignant, and normal masses in comparison to the state-of-the-art. Furthermore, the proposed models attempt to learn the dataset using conventional ML of RF and KNN, as well as classical DL of convolutional neural networks (CNN).

The remainder of the study is structured as follows: The methodological approach employed in this study is presented in Section 2. Section 3 follows directly after Section 2 and shows the results achieved from using the methodology used in this study, including both quantitative and qualitative outcomes. Section 4 contains the discussion, which explains the study in depth. The limitations of the sudy is presented in Section 5. Finally, Section 6 summarizes the study’s findings and future plans.

The present study made use of the publicly accessible Breast Ultrasound Images (BUSI) dataset, that can be found at https://scholar.cu.edu.eg/?q=afahmy/pages/dataset[20] and United States (US) breast ultrasound images found at https://qamebi.com/breast-ultrasound-images-database/[21], both of which were accessed on May 1, 2023 and August 10, 2023 respectively. The ultrasound scans (BUSI) were collected from women between the ages of 25 and 75. The data was acquired in 2018 at Cairo’s Baheya Hospital for Early Detection and Treatment of Women’s Cancer using a LOGIQ E9 ultrasound system and a LOGIQ E9 Agile ultrasound [22]. There are 600 female patients in all. The data is divided into three categories: normal, benign, and malignant. The dataset contains 780 images with a resolution of 500 $\times$ 500 pixels. There were 133 normal images with no cancerous masses, 437 images with cancerous masses, and 210 images with benign masses among the 780 cancerous images. Furthermore, the US breast ultrasound images include 123 and 109 ultrasound images of malignant and benign breast tumors, respectively.

This study employs a framework for classifying the breast masses shown in Fig. 1. The BUSI and US ultrasound images data underwent several preprocessing stages, such as, image augmentation to improve the deep learning models’ performance and generalization capabilities for computer vision tasks. To facilitate quick training, the RGB (Red, Green, Blue) images were transformed to grayscale. As a result of the lower dimensionality, training models on grayscale images is often faster than training models on RGB images. It should be noted that converting RGB to grayscale is a typical preprocessing step in deep learning models, especially when color information is not critical or when lowering input dimensionality is crucial. Deep learning models might suffer from noise in the dataset. It can introduce inconsistencies, outliers, or irrelevant data from which the model can learn and generalize. The suspected noise in the dataset was removed so that the model could focus on the key patterns and characteristics in the data, resulting in increased performance and more accurate predictions.

Furthermore, feature extraction is an important stage in deep learning. This was done to enable automated learning of complex representations, reduce data dimensionality, improve computer efficiency, enhance transfer learning, promote generalization, and facilitate interpretability. The dataset was separated into training (70%) and testing (30%) sets after the data preparation phases. A percentage of the testing set (30%) was utilized to validate the models to be trained. The random forest (RF), K-Nearest Neighbor (KNN), and convolutional neural network (CNN) are the machine learning (ML) and deep learning (DL) classifiers utilized to train the dataset in this study. These artificial intelligence (AI) techniques were utilized to categorize breast masses as benign, malignant, or normal. Finally, the proposed models’ outputs were evaluated utilizing several evaluation metrics.

The preprocessing techniques used in this study are explained in the following subsections. Data preprocessing is the process of modifying and preparing data in order to make it more suited for the job at hand and to improve the performance of machine learning and deep learning algorithms [15]. Data preprocessing is critical for ensuring the quality, dependability, and efficacy of future analysis or modeling techniques [23]. Figs. 2,3 show the distribution of the BUSI and US ultrasound image datasets before augmentation. It is essential to note that the BUSI dataset is multiclass (normal, benign, and malignant), whereas the US ultrasound images dataset is binary (benign and malignant).

In this study, three artificial intelligence techniques are employed to train the dataset to classify breast masses as benign, malignant, or normal: random forest, K-Nearest Neighbor, and convolutional neural network. It should be stressed that optimization is an important part of model training. The goal of optimization is to reduce the risk of mistakes or loss from these forecasts while also improving the model’s accuracy [33]. Depending on the dataset and the complexity of the task at hand, optimization strategies may differ. The hyperparameter and validation strategies were employed to optimize the models in this study. The 60% training, 20% validation, and 20% testing hyperparameter approach is used. To evaluate the model’s performance and prevent overfitting, 5-fold cross validation was used. The 5-fold aids in estimating the model’s generalization capabilities.

Model evaluation is critical since it measures a classifier’s performance as a generic model [41]. The purpose of a performance evaluation is to measure a model’s generalization accuracy on unseen/out-of-sample data [42]. Accuracy, precision, recall, and Receiver Operating Characteristics (ROC) are among the performance evaluation measures utilized. Tenfold cross validation was used to validate the dataset. The accuracy, precision, and recall are represented by Eqns. 4,5,6, respectively.

(4)$$\text{Accuracy}=\frac{TP+TN}{TP+FP+TN+FN}$$

(5)$$\text{Precision}=\frac{TP}{TP+FP}$$

(6)$$\text{Recall}=\frac{TP}{TP+FN}$$

Where true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values. The area under the curve-receiver operating characteristics (AUC-ROC) curve is a classification performance statistic with variable threshold values. The AUC is a measure of separability, whereas the ROC is a probability curve. The ROC curve is plotted as a function of true positive rate (TPR) vs. false positive rate (FPR), with TPR on the y-axis and FPR on the x-axis. The flowchart of the proposed method from start to end, taking into account the phases involved in achieving model classification results and performance assessments, is shown in Fig. 7.

The study used a non-binary (benign, malignant, and normal) data description in the training of the proposed models, which included random forest, K-Nearest Neighbor, and convolutional neural network. The performance scores of the models used in this study are shown in Table 3 below. The presented findings are the outcome of training models based on the techniques employed. These classifiers were trained using the models and evaluation scores described in the deep learning and machine learning classifiers and their accompanying evaluation metrics. The Google Colaboratory (Colab) platform is used to train the models. The Google Colab assists in the creation and execution of arbitral decisions and is well-suited to ML and DL, data analysis, and education.

Table 3 shows that the CNN achieved the highest accuracy with 0.9610, followed by the KNN with 0.6439. With 0.6146, the RF had the worst performance. Furthermore, the CNN obtained significant precision (0.9724) and recall (0.912) values. The results reveal that deep learning outperforms conventional machine learning techniques in training image datasets. The results obtained from the three models used in this study are summarized in Fig. 9.

Additionally, the receiver operating characteristics (ROC) of the models were calculated for each of the artificial intelligence models. The Receiver Operating Characteristics (ROC) scores are used to assess the efficacy of the models. The ROC score is a well-known machine learning assessment metric. The capacity of a classification model to differentiate between the breast masses data across numerous classification criteria is measured by the ROC. Figs. 10,11,12 show the ROC values for CNN, RF, and KNN, respectively. At various classification thresholds, the ROC shows the true positive rate (TPR) vs. the false positive rate (FPR). Based on the training sets, the CNN had a ROC value of 0.99, the RF had a ROC value of 0.85, and the KNN had a ROC value of 0.56.

From the Fig. 10, it can be observed that, the CNN classifier achieved a value of ROC = 0.99 in terms of the training and validation sets. The ROC score shows that the model achieved their best discrimination between classes, which means that it properly categorized all positive examples with no false positives. It should be noted that the data used in this study is well labeled and free of imbalances. Other evaluation metrics were utilized to examine the model’s performance in order to determine these values. This is evident when compared to the precision and recall values obtained by the CNN classifier in Table 3.

Furthermore, the ROC values obtained in Fig. 11 illustrate the distribution of the RF model. The validation ROC score for the RF is 0.54 while the training ROC value is 0.85. During the model’s training operation, the RF attained critical thresholds. It is also important to remember that a ROC of 0.5 or more is considered noteworthy in all circumstances [10]. This means that the RF did well in classifying the breast masses in all cases.

In the case of the KNN model (Fig. 12), the model attained crucial values of more than 0.5 in the ROC, indicating that the model can differentiate between the classes and is superior to random guessing.

The ROC score is less impacted by unbalanced class distributions than other measures such as accuracy. When one class dominates the dataset, accuracy can be deceiving, but ROC provides a more reliable judgment. It is important to recognize, however, that the ROC score has a few drawbacks. Table 4 presents the ROC curve values for the models’ training and validation outcomes. The CNN earned the highest ROC value in both the training (0.99) and validation (0.99) sets. The RF had the second-best ROC score in the training set but the worst fared in the validation sets is to the KNN (0.48).

A range of assessment metrics were utilized in this study to provide a full understanding of models’ performance. This approach provides vivid understanding of model performance in the study of machine learning and deep learning. Table 5 shows the performance of the ML and DL on US ultrasound image datasets. It is clear that the CNN had the greatest accuracy of 0.8251, followed by the KNN with 0.7557. The RF had the lowest accuracy (0.6602). Furthermore, the models attained significant recall and precision values. It should be noted that the models utilized in this study can train both multiclass and binary class datasets. The capacity of the models to survive data imbalance and overfitting in all parts of the training models is what makes them unique.

The ROC of the training and validation sets for the CNN, RF, and KNN models in the US ultrasound images dataset is shown in the Figs. 13,14,15.

Table 6 summarizes the ROC for the CNN, RF, and KNN models in the US ultrasound images dataset.

The CNN, RF, and KNN models were therefore trained using the Google Colab platform using Python software version 3.11.4 (The Python Software Foundation (PSF), 1209 Orange Street, Wilmington, DE, USA). The installed Random Access Memory (RAM) is 12.0 gigabytes (GB), and the operating system is 64-bit. Table 7 provides a tabular representation of the system specs and experimental settings used to reach the results.

Our experimental system consists of five nodes, each with an Intel(R) Core(TM) i5-1155G7 @ 2.50GHz, eight cores, and twelve gigabytes of RAM. The platform stack is made up of the Google Colab platform. In these studies, we have one master and up to four slaves. We developed a series of tests based on the system’s hardware limits. The product of ‘number of cores per executor’ and ‘number of executors’, for example, can range from 1 to the maximum number of cores in the cluster. Table 8 shows the experimental configurations used to acquire training data. We created and evaluated AI prediction models of three sorts, namely CNN, RF, and KNN, with data volumes ranging from 5 GB to 15 GB. A windows bash script runs each application for each of the parameter values listed in Table 8. A few possibilities are incorrect owing to resource mismatches and are thus skipped. We could gather around 400 data points as a training set for each application in order to create the AI models.

Breast cancer is a major health problem for women all over the world. Early identification and preventative actions are critical in lowering the disease’s effect. Women can take important precautions to protect their breast health. This study shows some qualitative empirical performances to lessen the consequences of breast cancer in women. To begin, knowing one’s family history and personal risk factors is critical to taking charge of breast cancer prevention. Women should enquire about their family’s medical history, particularly about breast cancer, and share this information with their healthcare professional. Women who are aware of their risk factors are better able to make educated decisions and take preventative steps. Women are urged to follow a healthy lifestyle after learning about their family history of breast cancer as a proactive start toward minimizing their chance of having breast cancer. Women can actively lower the likelihood of breast cancer incidence by cultivating a healthy lifestyle. Regular physical exercise aids in weight control and adds to a healthy body mass index (BMI) [13]. Limiting alcohol use, not smoking, and limiting exposure to environmental contaminants all help mitigate the risk.

In addition, women are advised to practice breast self-examination (BSE) as a simple yet powerful technique for monitoring changes in their breast health [16]. By conducting BSE on a regular basis, women become acquainted with the usual appearance and feel of their breasts, allowing them to detect any abnormal changes quickly. To guarantee consistency and accuracy, BSE should be tested periodically, ideally a few days after menstruation. Women can take an active role in their breast health by incorporating BSE into their routine, as well as visiting healthcare professionals for clinical breast examination (CBE) to provide an additional layer of assessment and expertise. It should be noted that the researchers in this study are not health practitioners; however, they use technological tools to train health data, as well as experience and encounters with women in community services, recommend the aforementioned approaches to mitigate breast cancer risk. Remember that these are basic guidelines, and women should speak with their healthcare professionals to build a tailored strategy based on their unique risk factors and medical history.