arXiv cs.AI
·
6 hours ago
Researchers developed an adaptive layer-freezing strategy for federated learning that reduces energy consumption in medical imaging tasks by selectively freezing encoder weights during training. The approach achieved up to 23% reductions in training time, energy consumption, and CO2 emissions while maintaining performance on MRI-to-CT conversion across multiple architectures. The method enables healthcare institutions with limited computational resources to participate in collaborative model training without significant performance trade-offs.
arXiv cs.AI
·
6 hours ago
Researchers proposed a method to improve calibration of segmentation models trained with region-based losses like Dice loss by modifying the gradient vector field during training. The technique scales gradient magnitude linearly with prediction error, adding a factor to the loss's partial derivative. Medical imaging segmentation tasks showed improved calibration while maintaining prediction accuracy across 2D and 3D applications.
arXiv cs.AI
·
6 hours ago
Researchers developed a method using demographically-conditioned synthetic medical images generated with fine-tuned Stable Diffusion 2.1 to address fairness issues in COVID-19 chest CT classifiers. When synthetic images were used for pretraining followed by fine-tuning, the classifier achieved comparable performance to using 100 times more real data, and synthetic minority cohorts reliably ranked subgroup performance (Spearman ρ = 1.00) compared to real test sets. This approach enables both training bias mitigation and evaluation bias detection by generating synthetic data in demographic groups where real test samples are too scarce for reliable fairness audits.
arXiv cs.AI
·
6 hours ago
Researchers created a new dataset of 3D MRI brain scans paired with text reports and built a vision-language model that uses multiple LLMs working together to generate medical reports for brain tumors. The model outperformed existing 2D and 3D methods on report generation and visual question answering tasks. This approach enables AI systems to better analyze 3D medical imaging and could improve diagnostic accuracy in brain oncology.
arXiv cs.CL
·
6 hours ago
ReportMedSAM is a framework that uses radiology reports to guide medical image segmentation by creating a learnable bank of organ concepts aligned through contrastive learning with a frozen medical vision-language encoder. The system achieved competitive segmentation accuracy on the AbdomenAtlas 3.0 dataset while remaining robust to clinical language variations like different names for the same organ. The approach allows new anatomical structures and segmentation tasks to be added without retraining existing components, addressing scalability limitations of rule-based extraction methods.
