Vit-Resnet Assessment of Breast Density From Mri Publisher



Khalaj M ; Arian A ; Torabi A ; Ahmadinejad N ; Gity M ; Yazdi SNM ; Afshari MP ; Tabrizi MS ; Soltanianzadeh H
Source: 2025 7th International Conference on Pattern Recognition and Image Analysis, IPRIA 2025 Published:2025

Abstract

Breast density is a significant risk factor for breast cancer, with higher fibroglandular tissue (FGT) ratios correlating to increased cancer risk. Traditional mammography has limited sensitivity in dense breasts, prompting the use of magnetic resonance imaging (MRI) as a supplemental tool due to its higher sensitivity and three-dimensional (3D) imaging capabilities. While 3D deep learning models are often employed for breast density assessment, they are computationally expensive, requiring specialized and costly hardware. To address these challenges, this study proposes a lightweight two-dimensional (2D) hybrid deep learning model, ViT-ResNet, for efficient breast density classification using MRI, with improved performance. The proposed model combines the strengths of convolutional neural networks (CNNs), which excel at extracting local features, with vision transformers (ViTs), which capture global contextual information, achieving a balance of computational efficiency and robust performance. The model was trained and validated on a dataset of 654 MRI scans, using the 40th, 50th, and 60th percentile slices to represent diagnostically relevant regions. It achieved an average accuracy of 81.7% across ten iterations, with precision, recall, F1 score, and Cohen's kappa of 86.1 %, 78.3 %, 80.8%, and 0.738, respectively. In the iteration with the highest accuracy, the model achieved an accuracy of 84.8%, a precision of 89.2%, a recall of 81.5%, an F 1 score of 84.1%, and a Cohen's kappa of 0.784. Compared to prior studies utilizing 3D approaches that required multiple GPUs, this work demonstrates improved performance and computational efficiency, achieving effective results with a single GPU and showcasing the potential of a 2D strategy for breast density classification using MRI. © 2025 IEEE.