Region of Interest-Based Breast Cancer Detection with Oversampling Technique

Breast cancer is a major global health concern and one of the most common malignancies affecting women globally [1-3]. After lung cancer, breast cancer ranks as the second most prevalent cause of death among women [4]. Breast cancer incidence continues to rise according to statistics from respectable health organizations, emphasizing the essential need for early detection and intervention. Early detection not only improves treatment outcomes but also increases the probability of a full recovery. Studies show that women diagnosed with breast cancer in its early stages have substantially better survival rates, enabling more effective and less invasive treatments [5].

Despite advancements in medical imaging technologies, accurate breast cancer detection remains a challenging task. The complexity lies in the vast volume of mammographic images and the relatively small size of abnormalities compared to surrounding breast tissue [6]. Moreover, the imbalanced nature of breast cancer datasets, where normal cases significantly outnumber cancerous ones, exacerbates the problem by skewing model predictions towards the majority class [7]. One of the key issues in breast cancer detection is how to efficiently process large volumes of medical images while maintaining a high level of accuracy in detecting malignancies. Manual examination of such datasets is time-consuming, labor-intensive, and prone to human error [8]. This highlights the need for automated methods that can accurately identify suspicious regions within images without overwhelming medical practitioners with data overload [9].

Breast cancer detection methods can be broadly classified into two types: region-of-interest (ROI)-based and patch-based approaches [10]. ROI-based methods focus on isolating specific areas of the image that are most likely to contain abnormalities, excluding irrelevant regions to streamline analysis. In contrast, patch-based methods divide the image into smaller segments, conducting feature extraction or detection on each segment individually [11]. Determining whether patch-based or ROI-based methods are better depends on various factors including the specific application, dataset characteristics, computational resources, and the desired balance between accuracy and efficiency [12]. Patch-based methods often offer more granularity and flexibility in analyzing images as they operate on smaller segments or patches, allowing for detailed feature extraction and potentially capturing local patterns effectively. However, they may require processing multiple patches across the entire image, which can be computationally intensive and time-consuming, especially for high-resolution images [13]. On the other hand, ROI-based methods focus only on the regions of interest containing relevant information, potentially reducing computational overhead and processing time. Choosing ROI is a common step in medical image analysis across all imaging modalities [14]. By isolating the target areas, ROI-based approaches may also enhance the signal-to-noise ratio and improve the efficiency of subsequent analysis steps. However, identifying accurate ROIs upfront can be challenging, and there is a risk of overlooking important information present in other parts of the image.

This study proposes a ROI-based approach combined with oversampling techniques to address these challenges. By focusing on specific areas within the mammograms that are likely to contain abnormalities, the ROI-based method reduces the computational load and allows the model to concentrate on the most critical parts of the image [15]. Additionally, SMOTE Tomek Link, an oversampling technique, is applied to handle the dataset imbalance by generating synthetic samples for the minority (cancerous) class while removing noisy or overlapping samples between the classes [16]. This combined approach not only improves the model's ability to detect cancerous regions but also mitigates the effects of class imbalance, leading to more reliable predictions.

The integration of ROI-based methods with SMOTE Tomek Link for breast cancer detection is a novel contribution in this field. While ROI-based methods have been explored in medical image analysis, most studies have focused on either ROI extraction or oversampling techniques in isolation. Our approach stands out by addressing two critical issues simultaneously: focusing on the region’s most likely to contain abnormalities and dealing with the class imbalance inherent in breast cancer datasets. In contrast to traditional patch-based methods that divide the entire image into smaller sections for analysis, our ROI-based method specifically isolates the most relevant areas, which reduces both computational costs and the risk of false positives. Moreover, the use of SMOTE Tomek Link further enhances the model's performance by balancing the dataset, allowing the CNN architectures to learn effectively from both normal and cancerous cases. By addressing both data imbalance and precise feature extraction, this study sets a new standard for breast cancer detection, making it a promising approach for clinical applications where accuracy and efficiency are paramount.

3.1 Dataset

The RSNA screening mammography breast cancer detection dataset is a large dataset specifically designed for developing machine learning models for breast cancer detection. It consists of 54,713 mammographic images from around 11,000 patients, with each patient having at least four images taken from different projection angles and laterality to ensure comprehensive coverage of the breast tissue. The dataset includes crucial metadata such as age, implantation status, BIRADS scores, and breast density, which are important features for classification and risk assessment. The mammograms were originally stored in DICOM format and later resized to 512×512 pixels and converted to 8-bit grayscale to standardize the images for deep learning model training. The example of the images is shown in Figure 1. Figure 1 illustrates a visual comparison between mammogram images representing two distinct classes: the normal class (first row) and the cancer class (second row). In the first row, images depict mammograms from patients classified as normal, meaning no cancerous regions are detected. These images exhibit smooth, uniform breast tissue without any apparent masses or irregularities. The mammograms show clear areas with no signs of abnormal growth or suspicious features. In contrast, the second row contains mammograms from the cancer class, where cancerous regions are present. These images demonstrate noticeable differences compared to the normal class, with visible irregular masses or calcifications that may indicate the presence of malignant tumors. These abnormalities are typically denser and have an uneven shape, which is characteristic of cancerous tissue in mammograms. The visual differences between the normal and cancerous mammograms highlight the importance of feature extraction and detection, as well as the role of the deep learning model in distinguishing between these two classes to support early and accurate diagnosis.

1.png

Figure 1. Sample images from the RSNA dataset [21]

3.2 Proposed method

The proposed method of this study shown in Figure 2. It begins by utilizing the RSNA dataset, a widely recognized collection of mammographic images for breast cancer detection. The original RSNA dataset appears highly imbalanced as only about 1.8% of the images fall into the cancer class. To prepare the data for analysis, the dataset undergoes data splitting using two distinct ratios. The first split follows a 5:1 ratio, resulted 5,790 of normal class and 1,158 of cancer class. A second split with a 3:1 ratio is also employed, resulted 3,474 normal class and 1,158 cancer class. Additionally, a Random Under Sampling (RUS) technique is applied to address the issue of class imbalance in the dataset. This involves reducing the number of majority class samples, ensuring that the model does not become biased towards one class during training. The class distribution of dataset shown in Figure 3. Once the dataset is balanced, the images proceed through a preprocessing stage. During preprocessing, the images are converted to grayscale, which simplifies the image by removing color information and highlighting the most critical structures. This is a crucial step for reducing computational complexity while preserving key details necessary for detecting abnormalities. Following this, a ROI extraction method is applied to isolate specific areas of the mammograms that are most likely to contain cancerous lesions. By focusing only on these regions, the model becomes more efficient at detecting abnormalities while reducing noise from irrelevant parts of the image.

2.png

Figure 2. Proposed method

3.png

Figure 3. Class distribution

To further address the issue of class imbalance, a hybrid resampling method combining the SMOTE and Tomek Link is applied. SMOTE works by generating synthetic samples for the minority class (cancerous images) through interpolation between existing minority samples. This approach effectively increases the representation of the minority class, reducing the bias towards the majority class. However, while SMOTE addresses the imbalance, it can introduce noisy or overlapping data points. To counteract this, Tomek Link is employed as a cleaning technique. It identifies pairs of samples (one from the minority class and one from the majority class) that are closest to each other and misclassified by the classifier, known as Tomek Links. These pairs are removed to increase the separation between classes. By combining SMOTE and Tomek Link, the method not only oversamples the minority class but also removes borderline examples that may lead to model confusion, thus creating a cleaner and more distinct decision boundary. This hybrid approach offers an advantage over other oversampling techniques by simultaneously increasing minority class representation and reducing noise and class overlap, improving the model’s ability to accurately distinguish between cancerous and non-cancerous images.

The core of the analysis involves the use of several pre-trained CNN models, including InceptionV3, InceptionResNetV2, ResNet152V2, Xception, DenseNet201, and EfficientNetB7. These models have been previously trained on large image datasets, which allows them to perform highly effective feature extraction from medical images. In this study, these pre-trained models are fine-tuned to detect breast cancer within the extracted ROIs from the mammograms. By leveraging transfer learning, the research takes advantage of the powerful feature representations learned by these models, ensuring that they can accurately identify subtle patterns associated with breast cancer.

Finally, the performance of these models is evaluated using a range of standard metrics to assess their effectiveness in detecting breast cancer. Metrics such as accuracy, precision, recall, and F1-score are likely used to provide a comprehensive evaluation of each model's ability to correctly classify the images. This performance evaluation allows the study to determine the most effective approach for detecting breast cancer using the combination of ROI extraction, data balancing techniques, and pre-trained CNN models. This detailed methodology demonstrates the rigorous approach taken to develop an accurate and robust system for breast cancer detection.

3.3 Experimental setting

This experimental setup ensure that the models are well-optimized to achieve high accuracy and robustness in detecting breast cancer, especially in a dataset with significant class imbalance. The dataset is divided into two parts, with 80% allocated for training and 20% for testing, while 10% of the training data is set aside as validation data to monitor performance and prevent overfitting. The dataset is processed using various CNN architecture models, with optimized hyperparameters as detailed in Table 1. These include a learning rate of 0.0001, a batch size of 32, 100 epochs, the Adam optimizer, and the ReLU activation function to ensure optimal performance. The chosen learning rate allows for fine-tuned weight adjustments, while the batch size balances memory usage and computational efficiency. The Adam optimizer is selected for its adaptability and ability to handle noisy gradients, aiding faster convergence, and the ReLU activation function enables the model to effectively capture non-linear relationships within the data. The entire experiment is conducted using Python on a DGX A100 server equipped with 40GB of RAM, leveraging GPU acceleration to handle large-scale computations. This setup ensures the model can efficiently process high-resolution mammographic images, reducing training time and overcoming memory constraints, while enabling high performance in breast cancer detection.

Table 1. Hyperparameter setting

3.4 Performance metrics

The performance of the models was evaluated using accuracy, precision, recall, and F1 score. These metrics were derived from the model’s confusion matrix, which is based on the values of True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN), as shown in Eqs. (1)-(4). Each metric provides a different perspective on the model’s ability to correctly classify cancerous and non-cancerous regions, offering a comprehensive assessment of its performance.

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$          (1)

Precision $=\frac{T P}{T P+F P}$          (2)

Recall $=\frac{T P}{T P+F N}$          (3)

$F 1-Score =\frac{2 *( { Precision }+  { Recall })}{{ Precision }+  { Recall }}$          (4)

This study proposes the use of ROI in the detection of breast cancer, assessing various image proportion schemes to address the significant imbalance observed in the original dataset. The original dataset, along with three modified versions—Split 1, Split 2, and RUS datasets—were evaluated for effectiveness. The Split 1 and Split 2 datasets were derived by extracting data at ratios of 5:1 and 3:1, respectively, from the original dataset, resulting in progressively less severe imbalances. The RUS dataset represents a balanced version achieved through RUS method. Experimental results reveal that the Split 2 ratio scheme yields superior accuracy compared to the other datasets, with the EfficientNetB7 model achieving the highest accuracy of 97.41% after applying SMOTE+Tomek Link oversampling. Although the study yielded promising outcomes, opportunities for further refinement exist. Future work could include incorporating more advanced data augmentation techniques to enhance dataset diversity. Additionally, deploying the system in a real-world clinical setting and testing its efficacy on a larger and more varied dataset could significantly contribute to its validation, enhancing its robustness and generalizability for practical applications.