Automated lesion segmentation is essential for DR screening, but current deep learning models often lack robustness, generating false positives in low-contrast or artifact-heavy regions. This instability largely stems from a lack of anatomical understanding. While incorporating vessel structures can guide the model, obtaining pixel-level vessel annotations for training is notoriously expensive and scarce.

To address this dilemma, the research team proposed MedFuse on 15 March 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

Sub-headline: HIT (Shenzhen) researchers develop FedPD to enhance personalized cross-architecture collaboration

Researchers from Harbin Institute of Technology (Shenzhen) proposed FedPD, a personalized federated learning method based on partial distillation. By assessing knowledge relevance for selective transfer, FedPD enables efficient collaboration among clients with diverse model architectures while significantly improving performance on heterogeneous data.

Applying biochar to soil is a recognized strategy for combating climate change, primarily by locking away carbon for long periods. Yet, its broader impact is complex; under different conditions, biochar can either suppress or unexpectedly release other potent greenhouse gases like nitrous oxide and methane from the soil. This inconsistency has been a significant barrier to its widespread adoption. A new set of predictive models developed by researchers Beatriz A. Belmonte, Raymond R. Tan, and their colleagues at the University of Santo Tomas and De La Salle University brings clarity to this issue. The team created a system to predict how soils will respond to biochar, offering a way to tailor its application for maximum climate benefit.

By introducing fairness from the beginning with ‘fuzzy’ systems that understand ambiguity and shades of correctness, the evolved AIs balanced fairness and accuracy even when tasked with coming up with solutions for complicated financial and ethical issues.

At Oak Ridge National Laboratory’s Center for Nanophase Materials Sciences, Yongtao Liu is building AI-driven “closed-loop” nanomaterials experiments that can plan measurements, interpret results in real time and choose the next step — accelerating discovery without removing human judgment. His focus is not just speed but trustworthy autonomy: systems must be interpretable, resilient to instrument artifacts and designed to avoid “false novelty,” where noise masquerades as new physics. Drawing on work such as novelty detection in conductive AFM studies of halide perovskites — linking local microstructure to unusual hysteresis behavior — Liu emphasizes that autonomous labs also demand better methods to validate and understand the massive data they produce. He is developing practical tools like AEcroscopy to standardize automated microscopy workflows and a Gated Active Learning Framework to prevent models from confidently learning from out-of-assumption data, while also pushing cross-facility autonomy that links fast measurements with slower synthesis. Ultimately, Liu envisions AI that helps scientists reason and explore vast experimental spaces — freeing researchers from repetitive tasks so they can focus on asking sharper questions.