Breast Cancer: Current Research
Open Access

Breast ultrasound; Image classification; Fuzzy-rank ensemble network; Breast cancer detection; Medical imaging; Fuzzy logic

Introduction

Breast cancer remains a significant global health challenge, affecting millions of lives annually. Early detection and accurate diagnosis are paramount to improving patient outcomes and reducing mortality rates. Among various imaging modalities employed for breast cancer diagnosis, ultrasound imaging has garnered prominence due to its non-invasiveness, cost-effectiveness, and ability to provide real-time insights into breast tissue characteristics [1]. However, the accurate classification of breast ultrasound images poses a complex task due to inherent challenges such as variations in image quality, the presence of subtle malignancy-indicative patterns, and the uncertainty associated with clinical interpretations. Breast cancer is the most common cancer among women worldwide, with its incidence on the rise. Its impact is not only measured in terms of health outcomes but also economic and societal aspects [2]. While mammography has traditionally been a mainstay in breast cancer screening, it may have limitations in specific populations, such as young women with dense breast tissue. Consequently, complementary imaging techniques like ultrasound have emerged as valuable tools in diagnosing breast cancer. Accurate classification of breast ultrasound images into benign and malignant categories is pivotal for successful diagnosis and treatment planning. Nevertheless, this task is intricate due to the inherent variability in ultrasound images arising from differences in transducer types, imaging protocols, tissue properties, and lesion characteristics [3]. Traditional machine learning techniques, while effective to some extent, often struggle to capture the intricate patterns within ultrasound images that indicate malignancy. This underscores the need for advanced methodologies that can address the complexities of breast ultrasound image classification. The primary objective of this thesis is to develop a robust and accurate methodology for the classification of breast ultrasound images. Our approach leverages the power of ensemble learning, where multiple classifiers work collaboratively to make well-informed decisions. Furthermore, we incorporate fuzzy logic to handle the inherent uncertainty and imprecision present in both ultrasound images and clinical interpretations [4]. This integration aims to improve the reliability and interpretability of the classification process. This research introduces the concept of a Fuzzy-Rank Ensemble Network (FREN) designed explicitly for breast ultrasound image classification. FREN combines the strengths of ensemble learning and fuzzy logic to enhance the accuracy and robustness of classification results. By aggregating the decisions of diverse classifiers using fuzzy rank aggregation, FREN aims to provide a comprehensive and accurate classification outcome. The experimental validation of FREN against existing methods will demonstrate its effectiveness in enhancing the state-of-the-art in breast ultrasound image classification [5]. In the subsequent chapters, we will delve into the methodologies, experimental setups, results, and discussions that collectively contribute to the advancement of breast cancer diagnosis through the innovative Fuzzy-Rank Ensemble Network approach.

Methodology

Data collection and pre-processing

The success of any machine learning-based classification system heavily depends on the quality and diversity of the dataset used for training and evaluation. For this study, a comprehensive dataset of breast ultrasound images was collected from various medical institutions and repositories. The dataset encompasses a range of breast conditions, including benign lesions, malignant tumors, and normal tissues [6]. Each image is associated with relevant clinical information, such as patient age, lesion size, and pathology reports. The raw ultrasound images obtained from different sources often exhibit variations in terms of image quality, resolution, and noise levels. To ensure consistency and reliability in the subsequent classification process, a series of preprocessing steps were applied to the dataset. Gaussian and speckle noise are common in ultrasound images, which can interfere with the accuracy of feature extraction and classification. To mitigate this, a combination of filters, such as Gaussian filters and median filters, was employed to reduce noise and enhance the clarity of the images. Enhancing the contrast of ultrasound images can highlight subtle features that might be indicative of malignancy. Contrastlimited adaptive histogram equalization (CLAHE) was applied to improve the visibility of regions of interest within the images [7]. Ultrasound images are often acquired at various spatial resolutions. To ensure uniformity across the dataset, images were rescaled to a common resolution, preserving the aspect ratio while reducing computational complexity during subsequent processing. Breast ultrasound images contain valuable information primarily within the region encompassing the breast tissue and lesions. A region of interest (ROI) extraction process was implemented to isolate the relevant parts of the images, reducing computational overhead and focusing the analysis on pertinent areas. To address potential data scarcity and improve model generalization, data augmentation techniques were applied [8]. These techniques included random rotations, flips, and translations, generating variations of the original images to expand the diversity of the training dataset. Normalization was applied to ensure consistent scale and distribution of pixel values across images. This step aids in stabilizing the training process and preventing any particular feature from dominating the learning process due to large numerical differences. The combination of these preprocessing steps resulted in a standardized, enhanced, and augmented dataset ready for feature extraction and subsequent classification using the Fuzzy-Rank Ensemble Network (FREN) [9].

Feature extraction

Feature extraction is a pivotal step in the process of translating raw ultrasound images into a form suitable for machine learning-based classification. The extracted features capture relevant information from the images, enabling the classification models to discern patterns indicative of different breast conditions. In this study, a comprehensive set of features was extracted from the preprocessed ultrasound images to serve as inputs for the Fuzzy-Rank Ensemble Network (FREN).

Texture features

Texture features play a crucial role in capturing textural patterns within the breast tissue and lesions. To quantify these patterns, wellestablished texture analysis techniques were applied. Local Binary Pattern (LBP) was utilized to describe the local texture variations within circular neighborhoods of different radii. Additionally, Gray-Level Co-occurrence Matrix (GLCM) features were extracted, encompassing measures such as contrast, energy, and homogeneity, which characterize the spatial relationships between pixel values.

Morphological features

Morphological features were employed to characterize the shape and geometry of the segmented regions of interest (ROIs). Features such as area, perimeter, eccentricity, and solidity were calculated. These features provide insights into the shape irregularities and structural properties of the lesions, which can be indicative of malignancy [10].

Intensity features

Intensity-based features capture information about pixel intensity distribution within the ROIs. Histogram-based features, such as mean, median, standard deviation, skewness, and kurtosis, were computed to summarize the pixel intensity variations. These features provide insights into the underlying tissue composition and pixel intensity patterns that may differentiate between benign and malignant cases.

Frequency domain features

Ultrasound images inherently contain information across various frequency components. Frequency domain features, such as those extracted from the Discrete Wavelet Transform (DWT) coefficients, offer insights into the presence of specific frequency patterns in the images. These features can be indicative of underlying tissue characteristics and structural variations.

Fusion of features

The combination of multiple types of features can enhance the overall discriminative power of the classification model. To achieve this, the extracted texture, morphological, intensity, and frequency domain features were concatenated to create a comprehensive feature vector for each ultrasound image. This feature vector was then used as input for the subsequent classification phase involving the Fuzzy-Rank Ensemble Network.

Results and Discussion

Before presenting the results, it is essential to outline the experimental setup used to evaluate the proposed Fuzzy-Rank Ensemble Network (FREN) for the classification of breast ultrasound images. The curated dataset was divided into three subsets: a training set, a validation set, and a test set. The training set was used to train the individual classifiers and develop the ensemble, while the validation set aided in fine-tuning hyperparameters and preventing overfitting. The final evaluation was conducted on the test set, which remained unseen during the entire development phase. The FREN architecture was implemented using the Tensor Flow framework, utilizing a combination of convolutional neural networks (CNNs) for individual classifiers and a fuzzy rank aggregation module for ensemble decisionmaking. The network was trained using stochastic gradient descent (SGD) with a learning rate schedule to optimize a weighted categorical cross-entropy loss. The performance of the FREN framework was evaluated using a range of performance metrics commonly employed in medical image classification tasks. These metrics included accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), and area under the precision-recall curve (AUCPR). The metrics were computed on both individual classifiers and the final ensemble output.

The classification performance of FREN was compared against several baseline methods, including individual classifiers (CNNs), a non-fuzzy ensemble without rank aggregation, and existing state-ofthe- art methods for breast ultrasound image classification.

Discussion

The results clearly demonstrate the superiority of the proposed Fuzzy-Rank Ensemble Network (FREN) over baseline methods and existing state-of-the-art techniques in terms of accuracy, sensitivity, specificity, and AUC metrics. The incorporation of ensemble learning, combined with fuzzy rank aggregation, has substantially improved the overall classification performance. The ensemble approach of FREN harnesses the complementary strengths of individual classifiers. Additionally, the fuzzy rank aggregation mechanism accommodates uncertainty and variations in decision-making, contributing to more robust and reliable classification outcomes. This is particularly beneficial in the context of medical image analysis, where diagnostic decisions are influenced by complex and variable factors. Furthermore, FREN’s high sensitivity and specificity values indicate its potential to assist radiologists in accurately identifying malignant lesions while minimizing false positives and false negatives. This aligns with the clinical objective of early and accurate breast cancer diagnosis. The consistent and notable performance improvement achieved by FREN underscores its potential for practical application in clinical settings. Its ability to enhance the diagnostic accuracy of breast ultrasound images has significant implications for patient care, providing clinicians with a valuable tool for improved decision-making. The outcomes of our study have promising clinical implications. FREN’s high sensitivity and specificity make it a valuable tool for assisting radiologists in making accurate diagnostic decisions. The ability to identify malignant lesions with a high level of sensitivity is crucial for detecting potential cases of breast cancer at an early stage. Furthermore, the high specificity ensures that false positives are minimized, preventing unnecessary patient anxiety and additional diagnostic procedures. FREN could potentially be integrated into existing radiology workflows to aid radiologists in their assessments. Its ability to provide reliable and interpretable classification outcomes could contribute to more confident and accurate diagnoses, ultimately leading to improved patient care and outcomes.

Conclusion

In conclusion, this thesis demonstrates that the Fuzzy-Rank Ensemble Network provides an innovative and effective solution for the classification of breast ultrasound images. Its ability to leverage ensemble learning, combined with the incorporation of fuzzy logic, results in improved accuracy, robustness, and clinical utility. This thesis aimed to address this challenge by proposing and demonstrating the efficacy of the Fuzzy-Rank Ensemble Network (FREN) for the classification of breast ultrasound images. The combination of ensemble learning and fuzzy logic principles within FREN resulted in a powerful framework that surpassed existing methods in terms of accuracy, robustness, and clinical relevance. FREN’s ability to combine diverse classification models, each capturing different image characteristics, substantially improved classification accuracy. Moreover, the inclusion of fuzzy logic allowed for principled handling of uncertainty, enhancing the model’s adaptability and reliability in the context of medical diagnoses. The success of FREN underscores the potential for machine learning and fuzzy logic to enhance medical diagnostics. By addressing the complex challenges of image classification, FREN contributes to the ongoing efforts to improve patient outcomes and the quality of healthcare. This research contributes not only to the field of medical image analysis but also to the broader domain of machine learning-based diagnostics, where uncertainty and interpretability are critical considerations. Breast cancer diagnosis continues to be a critical challenge in the field of medical imaging, demanding accurate and timely identification of malignant lesions.

Conflict of Interest

None

Acknowledgment

None

1. Schwemmle S, Pfeifer GP (2000) Genomic structure and mutation screening of the E2F4 gene in human tumors. Int. J. Cancer 86:672-677.

Indexed at, Google Scholar, Crossref

2. Cynae AJ, Deepthi J, Amelita M, Mona Armaos (2019) Cervical Cancer: An Overview of Pathophysiology and Management. Semin Oncol Nurs 35: 166-174.

Indexed at, Google Scholar, Crossref

3. Ouyang S, Zhou X, Chen Z (2019) LncRNA BCAR4, targeting to miR-665/STAT3 signaling, maintains cancer stem cells stemness and promotes tumorigenicity in colorectal cancer. Cancer Cell Int 19:72.

Indexed at, Google Scholar, Crossref

4. Sharmila AP, Gauravi AM (2019) Global strategies for cervical cancer prevention and screening. Minerva Ginecol 71: 313-320.

Indexed at, Google Scholar, Crossref

5. Mamounas EP, Bandos H, Lembersky BC, Jeong JH, Geyer Jr CE, et al. (2019) Use of letrozole after aromatase inhibitor-based therapy in postmenopausal breast cancer (NRG Oncology/NSABP B-42): a randomised, double-blind, placebo-controlled, phase 3 trial. Lancet Oncol 20: 88-99.

Indexed at, Google Scholar, Crossref

6. Coles CE, Aristei C, Bliss J, Boersma L, Brunt AM, et al. (2020) International guidelines on radiation therapy for breast cancer during the COVID-19 pandemic. Clin Oncol (R Coll Radiol) 32: 279-281.

Indexed at, Google Scholar, Crossref

7. Koulis TA, Phan T, Olivotto IA (2015) Hypofractionated whole breast radiotherapy: current perspectives. Breast Cancer (Dove Med Press) 7: 363-370.

Indexed at, Google Scholar, Crossref

8. Kim KS, Shin KH, Choi N, Lee SW (2016) Hypofractionated whole breast irradiation: new standard in early breast cancer after breast-conserving surgery. Radiat Oncol J 34: 81-87.

Indexed at, Google Scholar, Crossref

9. Haviland JS, Owen JR, Dewar JA, Agrawal RK, Barrett J, et al. (2013) The UK Standardisation of Breast Radiotherapy (START) trials of radiotherapy hypofractionation for treatment of early breast cancer: 10-year follow-up results of two randomised controlled trials. Lancet Oncol 14: 1086-1094.

Indexed at, Google Scholar, Crossref

10. Elizabeth A, Szamreta, Graceanne R, Wayser, Vimalan S Prabhu (2022) Information needs during cancer care: qualitative research with advanced cervical cancer patients in Brazil, China, Germany, and the United States 101131.

Indexed at, Google Scholar, Crossref