Breast cancer remains one of the leading causes of cancer-related deaths among women globally. For effective diagnosis and treatment, precise breast imaging is pivotal. A recent study published on March 14, 2025, evaluated cutting-edge deep learning models to significantly improve breast region segmentation from dynamic contrasts-enhanced magnetic resonance imaging (DCE-MRI).
Utilizing data from 59 patients at Stavanger University Hospital, the research aimed to define optimal techniques for segmenting breast regions, which is fundamental for accurately assessing breast density and identifying abnormalities. Accurate segmentation not only reduces computational demands but also contributes to the environmentally sustainable use of technology for future research.
The study systematically compared seven deep learning architectures: UNet, UNet++, DenseNet, FCNResNet50, FCNResNet101, DeepLabv3ResNet50, and DeepLabv3ResNet101. This comparative analysis explored how these models perform under different evaluation metrics including accuracy, training time, and carbon footprint.
According to the findings, UNet++ achieved the highest performance with the lowest training loss of 0.0112 ± 0.0022. Meanwhile, UNet was noted for its strong generalization capability, achieving validation performance metrics with excellent reliability. The efficiency of these models was particularly highlighted by FCNResNet50, which emerged as the most environmentally friendly, exhibiting the lowest carbon impact during training.
The combination of high computational efficiency and minimal environmental effect makes these models valuable assets for clinical applications. The methods employed included comprehensive preprocessing of the imaging data and advanced quantitative evaluations of each model’s effectiveness through 10-fold cross-validation.
Understanding boundary detection—a significant challenge highlighted by the study—also emphasized the nuanced approach of distinguishing different anatomical features effectively. The research demonstrated how deep learning can assist radiologists by automizing the detection of tumor boundaries and enabling rapid, reliable assessments of breast tissue health.
Notably, the models showcased variances not just in performance but also training times, ranging from 87 minutes for FCNResNet50 to 185 minutes for DenseNet, which indicates varying efficiencies across architectures. UNet also boasted impressive inference times, allowing it to be more suitable for real-time applications.
Further validating the success of this approach, the authors noted, "UNet++ exhibited the best overall model performance, demonstrating superior accuracy in terms of Dice score."
Overall, addressing the gap between technological advancement and clinical practice, this study offers insights and innovations necessary for dissecting complex MRI imaging data.
Concluding, the balance between model performance and computational efficiency allows for the selection of the most suitable architectures per specific clinical requirements. For institutions aiming to implement sustainable solutions for imaging processes, the insights derived from FCNResNet50 will be of substantial importance. This research exemplifies the continual evolution of imaging technologies and highlights potential pathways for future studies aimed at enhancing diagnostic capabilities and environmental sustainability within the medical imaging domain.
Future explorations are expected to focus on refining pre-and post-processing methodologies, integrating various techniques to increase segmentation accuracy, and exploring datasets to reinforce model efficiency and transferability across various health demographics.