Breast cancer diagnosis using radiomics-guided DL/ML model-systematic review and meta-analysis

Cancer is one of the leading causes of death on a global scale, whereas breast cancer is the type of cancer that affects the most women. Early detection and accurate staging are essential for effective cancer treatment and improved patient outcomes. Recent developments in medical imaging and artificial intelligence (AI) have created new opportunities for breast cancer detection and staging. Medical image analysis techniques, including radiomics, machine learning and deep learning, have shown promise for breast cancer detection and stage estimation. The goal of the systematic review and meta-analysis is to evaluate and examine the state-of-the-art implications of radiomics-guided deep learning (DL) approaches for breast cancer early detection utilizing different medical image modalities. The selection criteria were established on the basis of the PRISMA statement. Our research employs a PICO structure and text mining technique (Topic Modeling) using Latent Dirichlet allocation (LDA) approach. The primary objective of the search was to conduct a thorough evaluation of the literature related to radiomics analysis and breast cancer in the fields of medical informatics, computer vision, and cancer research. Subsequently, the investigation concentrated on the fields of medical science, artificial intelligence, and computer science. The inquiry encompassed the years 2021 to 2024. The QUADAS-2 instrument is employed to evaluate the articles to ensure their quality and eligibility. Feature extraction methods that employ radiomics and deep learning are extracted from each study. The sensitivity value was pooled and transformed using a random-effects model to estimate the performance of DL techniques in the classification of breast cancer. The systematic review comprised 40 studies, while the meta-analysis consisted of 23 studies. The research studies employed a variety of image modalities, radiomics, and deep learning models to diagnose breast cancers. Ultrasound and DCI-MRI are the most frequently employed image modalities. The pyradiomcs pyhon package is employed to extract the radiomic features, and CNNs, ResNet, and DenseNet models are employed to extract the deep features. The LASSO (13) and T -test (9) statistical models are the most commonly used for feature selection. The most widely used deep learning models for breast cancer classification are ResNet and VGG. This systematic review and meta-analysis examined the feasibility of employing radiomics-guided deep learning/machine learning models for identifying breast cancer. The studies yielded positive results, as specific models demonstrated remarkable precision in distinguishing between malignant and benign breast tumors. However, there is a wide variety of variations in the designs of studies, the architectures of models, and the methodologies used for validation. Further research is required to verify the results of this study and to investigate the potential of deep learning models guided by radiomics in the early detection of breast cancer.