arXiv cs.AI
·
6 hours ago
ViPSAM is a visual prompting framework built on the Segment Anything Model that segments lesions in low-contrast CT scans by incorporating guidance from contrast-enhanced MRI images. The method was evaluated on liver lesion segmentation in non-contrast CT scans acquired for proton therapy planning and outperformed U-Net and baseline SAM approaches. This approach enables more accurate lesion delineation in medical images where contrast between lesions and background tissue is poor, improving segmentation reliability for clinical applications.
arXiv cs.CL
·
6 hours ago
Researchers developed LLM-T1D, a system that combines reinforcement learning with large language models to control insulin delivery for Type 1 Diabetes patients, making the decision-making process transparent and explainable. The system achieved 73.5% Time in Range on the FDA-approved UVA/Padova simulator while maintaining formal safety verification against hallucinations. The approach enables patients and doctors to understand why the insulin pump makes specific decisions, addressing the trust issues associated with black-box artificial pancreas systems.
arXiv cs.CL
·
6 hours ago
Researchers analyzed nine systems for medical question-answering on endoscopy images to understand how multimodal AI combines visual and textual information while maintaining reliability. Parameter-efficient adaptation of pretrained models achieved strong performance on the MediaEval Medico 2025 challenge, but improvements in answer accuracy did not guarantee faithful clinical reasoning. Systems using structured reasoning and explicit evidence grounding performed more reliably, suggesting healthcare AI design should prioritize explainability and robustness evaluation over raw performance metrics.
