Classification of Breast Malignant Tumor Using Ensemble Deep Learning Approach and Magnetic Resonance Imaging

Abstract

Breast cancer stands out as a predominant subject in both biology and health practices, mainly because it remains one of the leading causes of death among women. DCE-MRI and ultrasound together with mammograms, are the most commonly used imaging techniques for breast tumor diagnosis, and each delivers uniquely unique information regarding tumor areas. Even though many algorithms for breast tumor classification exist in the area of machine learning, evaluating multiple modalities separately, it is questionable how their classification performance can be advanced even further. For addressing these challenges, the hybrid deep learning architecture is suggested for the classification of the breast tumor from MRI and US data. First, the input samples are passed through the 2D and 3D convolutional layers for extracting features from each modality separately. Secondly, a discrimination-adaptation module is added to get data features that are independent of the modality by adopting a trio of discriminators through adversarial training. Then, a feature fusion is introduced, using the Multi-Modality Adaptive Feature Fusion (MMAFF) which combines the generic features from each modality and gives more compact features by constructing the affinity matrix and selecting the nearest neighbors only. An integrated MRI-US breast tumor classification dataset comprising 502 cases, including three clinical indicators: Patient data on lymph node metastasis, histological grade, and Ki-67 level is collected to assess the validity of the a priori proposed approach. OUTCOMES: The accuracy is 82%. 5%, 85. 7%, and 89. The detected DPYD variants affected treatment stratification for lymph node metastasis with 6% variance, histological grade, and Ki-67 level clinical indicators.