Prediction and Classifications of Breast Cancer Using Enhanced Convolutional Neural Network Approaches

Artificial Intelligence plays an important role in all aspects of regular life, and it is present in the medical field for predicting the severity of an issue, and accordingly, medical treatment can be taken for diagnosis. Machine learning (ML) contributes to the process of predicting breast cancer at an earlier stage and helps women to get healthy life. Various classification algorithms are used in the study conducted by Naji et al. [1], like K-Nearest neighbour (KNN), Random Forest (RF), Logistic Regression (LR), Decision Tree (DT), and Support Vector Machine (SVM) for predicting, diagnosing breast cancer, and identified that SVM outstripped the rest of the classifiers and produced 97.2% highest accuracy. Biosensor and several ML approaches are investigated in the research [2] for earlier prediction of breast cancer, where ML doesn’t require any pre-programmed models, it identifies patterns based on received data and forecasts outcomes by models, and biosensor quantifies organic characteristics of body fluids and tissues. Analysing breast cancer using genetic level is more expensive, and instead of that, histopathological image analysis is a common approach that is frequently used for cancer detection and analysis. Nassif et al. [3] processed a histopathological image using deep learning (DL), and previous works are reviewed to detect the seriousness of the disease. That provides opportunities for researchers to forecast the status of patients using Recurrent Neural Network (RNN) and LSTM methods.

Most research articles focus on the accuracy matrix, not on the confusion matrix, where it fails to distinguish false negative and false positive classifications, so it recommends that future research should concentrate on the confusion matrix for assessing the performance. Before introducing DL and ML approaches for detecting breast tumour by building models, a public data set was used for predicting and detecting the severity of the issue, and it gives a maximum of 99% accuracy using a Convolutional Neural Network (CNN). For creating a model, data should be analysed, and then it should be pre-processed with ML algorithms for measuring precision, recall, and accuracy. Khan et al. [4] divided the dataset into three portions—training, validation, and testing—with training conducted for 50 epochs. The model’s performance improved progressively with each epoch, ultimately achieving 99% accuracy using a CNN. LR was employed for classification and sequential optimization. For performing automated diagnosis, recursive feature elimination is used for feature selection and CNNs as a classifier exemplary also various algorithms were compared in Bhise et al. [5] for accuracy and precision. Feature selection is the process of reducing ambiguity and reducing the complication of the dataset, also reducing data size and training time. Embedded methods, filter, and wrapper methods are commonly used feature selection methods, where the wrapper method is used by recursive feature selection, and in contrast, the filter–based method is used for identifying the largest score.

The proposed scheme distinguishes malignant and benign tumour with earlier rate. Pre-processing is not mandate extracting features inevitably, when a complicated classifier like CNN is used, and is more proficient because it screens significant parameters and works extremely well on image data. Naji et al. [6] incorporated a 10-cross-field procedure for identifying best classifiers in ML approaches using the voting ensemble technique. This scheme gives a maximum accuracy of 98.1 and a minimum error rate of 0.01 is fashioned. Parameters like sensitivity, accuracy, precision, and F-measures are compared for finding the best classification efficiency using the voting ensemble technique, and at the end, it produces a low error rate. Medical advancements like mammograms, histopathological imaging, Magnetic Resonance Imaging (MRI), and computed tomography are used for the detection of cancer in earlier and malignant states, but it is error prone and cost-effective.

A characteristic of DL is that it can analyse and extract features using high-dimensional correlated data. It has been utilized in the study [7] for the histopathological and radiographic images, classifying breast cancer image modalities like ultrasound, mammography, thermography, and MRI using DL, which is the main aim of the research with openly offered datasets. Methods like filters and DL simulations were incorporated in the study [8] to focus on the prominence of microcalcification in mammograms. Here, filters are used for filtering standard and unusual mammogram descriptions, and the riddled imageries are taken as input and given to the Yolov4 DL model for classifying standard and unusual mammogram images. To prove the significance of a filter in classifying images, it was implemented and classified without using a filter that gives a significant difference in the results. Yolov4 is a DL model that detects an object in a frame in real-time, by capturing 10% accuracy and 12% frames. As the number of images increases, the success ratio increases using a data augmentation technique that changes the brightness of images, rotates them, and scales the size of an image. A DL scheme was applied to images by converting them into one-dimensional data, which increased the performance of accuracy in classification [9]. In this work, DL is applied for both histopathological images and a genetic factor dataset, which is a one-dimensional dataset, and it uses the convex hull algorithm and disseminated stochastic neighbour technique. For an enlightening performance, the proposed method uses the dataset and its decomposition, whereas the previous method accesses the dataset directly for classification.

CNN-based classifiers, dissimilarity mode decomposition, are used for data decomposition along with wavelet transformation. Hybrid CNN-based cancer detection is incorporated in Sahu et al. [10], which reveals improved performance than base classifiers and threshold value, weight factor plays a vital role for hybridization. Unlike conventional DL models, a high-accuracy CNN performs well, not only for large datasets, even for smaller dataset too.

Though the use of DL models like ResNet, DenseNet, and EfficientNet in medical image analysis has shown remarkable success, these models demand extensive datasets and heavy computational power, making them less feasible for use in clinical settings. In addition to this, most conventional approaches emphasize accuracy but fail to take into account the rate of false negatives and the explainability of the model.

In an attempt to tackle these challenges, this study presents an Enhanced Convolutional Neural Network (ECNN) that offers computational efficiency, small sample size requirements, enhanced classification results, and comprehensive evaluation criteria. The proposed ECNN was selected as the core architecture because CNNs excel at automatically learning hierarchical spatial features from raw pixel data in medical images. Unlike RNN or Long Short-Term Memory (LSTMs), which are designed for sequential or time-series data (e.g., gene sequences), CNNs effectively capture local patterns, edges, textures, and spatial hierarchies present in mammograms, ultrasound, or histopathological images through convolutional filters and pooling layers. Hybrid models, while powerful, often demand substantially larger datasets and higher computational resources, whereas the lightweight ECNN achieves strong performance with limited data, making it more practical for real-world clinical deployment.

The foremost purpose of this proposed work is to recognize an effective DL approach for detecting and diagnosing breast cancer in earlier and later stages, which is a subset of ML algorithms. Supervised Learning (SL), Unsupervised Learning (UL), and Reinforcement Learning (RL) are the major types of algorithms, whereas supervised learning is in need of a dataset for predicting the test data with a trained dataset with external supervision. It can be categorized into regression and classification. UL doesn’t require any external supervision; it is trained using an unlabelled dataset, and it supports solving association and clustering-related problems. RL is completely different from the above-mentioned algorithms; it contains one agent, and it acts with an environment, according to the feedback received it takes decision. Here, Q-Learning is considered the main algorithm, and it is supported based on actions and rewards.

3.1 Deep learning for breast cancer detection

DL is the scope of ML that emphases on development and application of Artificial Neural Networks (ANN) with several layers. It is stimulated by the construction and purpose of the humanoid brain and aims to enable computers to learn and make decisions autonomously without explicit programming. DL algorithms have demonstrated promising results in predicting breast cancer risk by analysing various factors such as mammographic images, genetic data, and patient demographics. CNNs are commonly used in image analysis tasks. They learn to identify complex patterns within mammograms, such as micro calcifications, masses, and architectural distortions.

By training on large datasets, CNNs can learn discriminative features indicative of breast cancer presence or absence. Such algorithms enable risk stratification, assisting clinicians in identifying high-risk individuals who may require further screening or preventive interventions. CNNs are particularly effective in analysing images and have been widely used in medical imaging tasks. They can learn complex patterns and features from mammograms or other imaging modalities to aid in breast cancer detection and diagnosis. CNNs consist of multiple layers of interconnected units, including convolutional layers, pooling layers, and fully connected (FC) layers. The convolutional layers apply filters or kernels to the input images, which helps in capturing local patterns and features. The pooling layers reduce the spatial dimensions of the features, making the network more robust to translation and scale variations. Finally, the FC layers enable the network to make predictions based on the learned representations.

For CNNs for breast cancer prediction, a dataset of labelled images is required. This dataset typically consists of mammograms or other relevant imaging data, along with corresponding labels indicating the presence or absence of breast cancer. The CNN is trained on this dataset, where it learns to automatically extract discriminative features from the images and associate them with the cancer diagnosis.

In the above pictorial representation, Figures 1 and 2 show deformities of breast cancer using a CNN using DL. The proposed ECNN architecture consists of multiple layers designed to improve feature extraction and classification performance.

Feature extraction, parameter sharing, spatial hierarchies, non-linearity and down-sampling are key responsibilities of the convolution layer in DL. Convolutional layers often include pooling operations, such as max pooling or average pooling, to down-sample the spatial dimensions of the feature maps. Pooling helps reduce the computational complexity of subsequent layers, provides local translational invariance, and helps the model to focus on the most salient features. Dimensionality reduction, feature selection, translation invariance robust to noise, and computational efficiency are the key responsibilities of the pooling layer in DL.

The pooling layer plays a crucial role in dimensionality reduction, translation invariance, feature selection, noise robustness, and computational efficiency within DL models. These responsibilities contribute to the effective extraction and representation of important features, enabling the network to learn and make accurate predictions. Parameter learning, classification, representation learning, and learning complex relationships are responsibilities of FC layers in DL. FC layer serves as the bridge between the preceding layers, responsible for learning complex relationships, introducing non-linearity, performing representation learning, producing the final output, and adapting the model's parameters during training. It plays a vital role in enabling the model to make accurate predictions or classifications based on the learned features.

• Input (224 × 224 grayscale) → Conv layers + BatchNorm + ReLU + MaxPooling + Dropout → Flatten → FC layers → Softmax output
• Prepare dataset (70/15/15 split) and apply preprocessing
• Compile model: Adam optimizer (Lr = 0.001), Categorical Crossentropy loss
• Train for 50 epochs (batch size = 32), monitor validation loss
• Evaluate on test set using accuracy, precision, recall, F1-score, confusion matrix, and Receiver Operating Characteristic–Area Under the Curve (ROC-AUC)
• Save the trained model

The above pseudocode illustrates the CNN algorithm that supports predicting the occurrence of breast cancer in earlier and later stages. Initially we should define CNN architecture and model model = Conv2D (input, filters, kernel_size, activation), and it is mandatory to add pooling layers for down-sampling. After flattening the output connection layers are activated with functions output = Dense (input, units, activation).

Then compiled using parameters like loss, optimizer, and metrics. Followed by that, the dataset is prepared, and it is trained using a CNN model. With the help of a trained model, the same procedure is repeated until the convergence of performance is achieved. After achieving the accuracy in the result, it is saved model.save (model_filename).

3.2 Data pre-processing

The dataset used in this research is the Breast Cancer Wisconsin Dataset acquired from the UCI ML Repository. It comprises 569 records, where 357 instances are benign, while the rest 212 are malignant. The dataset is split into training, validation, and testing datasets, accounting for 70%, 15%, and 15% of the dataset size, respectively. Traditional image operations (resizing, normalization, grayscale conversion, filtering) are applied first, followed by deep-learning-specific steps (convolutional feature extraction, Batch Normalization (BatchNorm), pooling, and dropout) within the ECNN model. This clear separation ensures that raw image quality is optimized before hierarchical feature learning.

Predicting breast cancer using a CNN involves medical image analysis, such as mammograms, and prediction is a complex task that requires extensive data pre-processing through various steps. Those steps are resizing, normalization, grayscale conversion, image filtering, and image enhancement.

Resized Image = Resize (Image, width, height)     (1)

Normalized Image $=\frac{\text { Image }- \text { Min }}{\text { Max }- \text { Min }}$     (2)

Grayscale Image = 0.2989 × Red + 0.5870 × Green + 0.1140 × Blue     (3)

$\operatorname{Output}(x, y)=\sum_{i=0}^{M-1} \sum_{j=0}^{N-1} \operatorname{Input}(i, j) \times \operatorname{Filter}(x-i, y-j)$     (4)

$\operatorname{Output}(x, y)=\frac{C D F(\operatorname{Input}(x, y))-C D F_{\min }}{M \times N-C D F_{\min }} \times(L-1)$     (5)

In the above Eq. (1), the image is resized using parameters like image, height, and width. In Eq. (2), the resized image is normalized with its min and max values. In Eq. (3), the normalized image is converted into a grayscale image. In Eq. (4) and Eq. (5), the pre-processed image.

3.3 Feature extraction

Using CNN for image processing feature extraction involves mining relevant and discriminative features of images. CNN learns hierarchical representation of visual data and patterns of images. For detecting special features of images, CNN consists of multiple filters, and it supports computing element-wise estimation. Complex patterns are enabled by introducing activation functions; here, CNN includes the rectified linear unit (ReLU) that replaces negative values with zero. Spatial dimensions are reduced by pooling layers, and it supports for extracting more relevant information. Local Response Normalization (LRN) layers are used to normalize responses within a local region and promote against adjacent sectors. Feature maps are visualized to understand patterns in a specific layer, which helps to interpret network behaviour.

3.4 Image classification

Image classifications are done in breast cancer prediction using CNN DL model by considering various parameters like dataset preparation, data pre-processing, model architecture, training, validation, and hyper parameter tuning, interpretation, and deployment, and testing are considered carefully for various influences like data quality, architecture selection, and interpretation techniques. This technique analyses the model's predictions and investigates the learned features to gain insights into its decision-making process. The interpretation of mammography images can be improved by using interpretability approaches like gradient-based class activation maps (Grad-CAM) or occlusion sensitivity to pinpoint key areas that are crucial to the model's categorization. Once the model's performance has been validated, it can be used to classify breast cancer on fresh, previously unveiled mammography pictures.

The recent developments in the field of DL have enabled new architectures to be built that excel at solving medical image classification problems. For instance, ResNet uses residual connections that allow training of extremely deep neural networks and overcome the problem of the gradient vanishing. Similarly, DenseNet improves feature propagation through the connection of each layer to all others, which increases information exchange and decreases redundancies. EfficientNet scales the depth, width, and input resolution of a neural network, ensuring maximum performance and a minimal number of parameters.

However, there are some downsides to the use of these types of models for detecting breast cancer. The main one is the requirement of an enormous amount of annotated data for effective training. Moreover, due to complex architecture and the need to employ transfer learning, training of such models requires many computing resources, which makes using them in clinical practice unfeasible. In addition, they are likely to be overfitted on small or imbalanced datasets, which is common in medicine.

Conversely, the proposed ECNN employs a simple architecture supplemented by BatchNorm, dropout, and optimal hyperparameters. This allows the model to achieve high classification precision and low computational complexity, making it suitable even for smaller datasets without the risk of overfitting. In addition, ECNN puts special emphasis on minimizing false negatives, which is particularly important in breast cancer detection.