tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Fibroglandular Tissue Classification in Breast MRI: A Comparative Study of Automated Decision Strategies

Authors :

Meysam Khalaj¹ Arvin Arian² Ala Torabi³ Nasrin Ahmadinejad⁴ Masoumeh Gity⁵ Seyedeh Nooshin Miratashi Yazdi⁶ Mohammad Pooya Afshari⁷ Melika Sadeghi Tabrizi⁸ Hamid Soltanian-Zadeh⁹

1- University of Tehran 2- Tehran University of Medical Sciences 3- Tehran University of Medical Sciences 4- Tehran University of Medical Sciences 5- Tehran University of Medical Sciences 6- Tehran University of Medical Sciences 7- University of Tehran 8- University of Tehran 9- University of Tehran

Keywords :

Fibroglandular Tissue Classification،Breast MRI،BI-RADS Assessment،Deep Learning،Shannon Entropy

Abstract :

Fibroglandular tissue (FGT) assessment in breast magnetic resonance imaging (MRI) is clinically important for breast cancer risk evaluation and is standardized in the Breast Imaging Reporting and Data System (BI-RADS) lexicon. While automated approaches have largely focused on segmentation, classification-based methods remain underexplored. Previous automated FGT classification studies have generally analyzed both breasts together, overlooking BI-RADS recommendations for side-specific evaluation and alternative strategies such as probability averaging or uncertainty-based rules. This study evaluates three assessment strategies: the conventional BI-RADS Maximum Rule, a novel Probability Averaging Rule to integrate bilateral information, and a novel Lower-Uncertainty Rule based on Shannon entropy to prioritize more confident predictions. These strategies were assessed using three diverse deep learning architectures, MobileNetV2, ResNeXt-26, and a hybrid ViT-ResNet, selected to analyze performance across models with different architectures and feature extraction mechanisms. The dataset comprised 654 pre-contrast 3D axial T1-weighted fat-saturated breast MRI scans, with each breast evaluated independently. Across ten independent runs, ViT-ResNet with Probability Averaging Rule achieved the highest test accuracy (0.85), F1 score (0.84), and Cohen’s kappa (0.78), while violin plot analysis showed that the Lower-Uncertainty Rule produced the lowest predictive entropy. Both proposed strategies consistently outperformed the conventional rule. The curated, expert-annotated dataset is publicly released to support reproducible research in this domain.