New SurfNet framework for cortical surface reconstruction from MRI

I just came across an incredible paper in Nature Communications (Oct 2025) introducing FAST (FrAme-multiplexed SpatioTemporal learning) a new deep learning framework that brings real-time denoising to high-speed fluorescence neural imaging. In simple terms, this technology helps neuroscientists see clearer images of brain activity as it happens processing data at over 1000 frames per second! What really stood out to me: 🔹 It’s self-supervised no need for ground truth data. 🔹 It uses a super lightweight CNN (only 0.013M parameters!). 🔹 It adapts intelligently between space and time to avoid losing details in fast-changing signals. 🔹 It even comes with a GUI, making real-time neural imaging accessible to more labs. This could be a game-changer for brain research especially in closed-loop neuroscience where every millisecond matters.

Deep Learning Enhanced Shear Wave Elastography with Adaptive Kernel Optimization for Liver Fibrosis Assessment This paper introduces a novel approach to Liver Fibrosis assessment using Shear Wave Elastography (SWE), enhanced by deep learning and adaptive kernel optimization. Our method surpasses existing SWE techniques by incorporating a dynamically adjusted kernel in the ultrasonic signal processing phase, leading to a 25% improvement in accuracy and a drastic reduction in measurement error rates. This innovation promises widespread clinical utility, significantly impacting the diagnosis and management of liver disease, and market potential is projected to reach $5 billion within 5 years. Our approach utilizes a two-stage methodology. First, a Convolutional Neural Network (CNN) extracts features from raw SWE B-mode images, classifying tissue types and identifying regions of interest. This information feeds into a second stage – an adaptive kernel optimization framework which dynamically adjusts the kernel used in the SWE signal processing via a Reinforcement Learning (RL) agent. The RL agent l https://lnkd.in/geCJ-cAX

Optimization in Machine Learning and Neuroscience Unifies Gradient Descent and Neural Adaptation via Derivative-Free Methods Recent research unifies diverse optimisation techniques, revealing that methods relying solely on trial-and-error, rather than complex calculations, can achieve comparable performance to standard approaches in artificial intelligence and offer a compelling mathematical framework for understanding how brains learn. #quantum #quantumcomputing #technology https://lnkd.in/exkCU6gs

I am pleased to share that I recently presented a segment of our current research on utilizing domain-specific BERT models, integrated with a tuned neural network, to enhance medical concept relationship classification in SNOMED CT at the MidAtlantic Bioinformatics Conference in Philadelphia, 2025. Our study examines how specialized biomedical language models (BioLinkBERT, PubMedBERT, ClinicalBERT, etc.) combined with a hyperparameter-tuned neural network can enhance the identification of semantic relationships between medical concepts. This approach aims to automate ontology updates by automatically classifying the correct relationship types for new medical concepts, eliminating the need for manual curation by ontology editors. It was a great experience discussing the work, receiving valuable feedback, and learning from other amazing researchers. #Research #AI #Healthcare #BiomedicalInformatics #SNOMEDCT #NJIT

AI tools are transforming how movement is studied in neuroscience — turning complex behaviors into data that help reveal how the brain works. One leading example is DeepLabCut, created by Mackenzie Mathis and Alexander Mathis, a deep learning software tool that tracks animal movements with remarkable precision. Now, extensions like SuperAnimal and AmadeusGPT are lowering barriers for researchers, expanding access to powerful tools that illuminate how neural activity gives rise to behavior. Read more: https://bit.ly/4qnX7JF

Excited to share our new paper, “Bridging Traditional and Deep Learning Methods in H&E Histological Image Normalization: A Comprehensive Review and Introducing a Novel Framework for Comparative Analyses,” is now published in the Journal of Advanced Research (Elsevier). This work provides an in-depth analysis of histopathology image normalization methods, from classical approaches to modern deep learning and diffusion models, and introduces a novel framework for evaluating and comparing normalization strategies. We hope this comprehensive review helps researchers and practitioners in digital pathology, AI-based diagnosis, and medical image analysis to select the most suitable normalization techniques for their studies. 💡 Highlights: - Unified categorization of traditional, deep learning, hybrid, and signal-processing-based normalization methods - Comparative analysis of eight state-of-the-art techniques - Introduction of a new deep learning–based framework for evaluating normalization performance 📖 Read the full open-access article here: 👉 https://lnkd.in/gtYmbGF3 I’m truly grateful to have worked with such an incredible team. A special thanks to Ali Masoudi-Nejad, Professor, whose guidance and insight have been invaluable throughout this research. Journal of Advanced Research #DigitalPathology #Histopathology #DeepLearning #AIinMedicine #MedicalImaging #ImageNormalization #ComputationalPathology #ReviewPaper #OpenAccess #Elsevier #Research

🚀 Excited to share my recent research work titled “A Comparative Study of Deep Learning Architectures for Brain Tumor Detection Using MRI Images.” In this study, I explored and compared CNN-based classification and 3D U-Net-based segmentation models for automated brain tumor detection. Using the BraTS 2020 dataset, the CNN achieved 96.8% classification accuracy, while the 3D U-Net delivered a Dice score of 0.87 for precise tumor localization. This project deepened my understanding of computer vision, neural networks, and medical image analysis, and strengthened my passion for applying AI in healthcare. 📄 You can read the full preprint here: 🔗https://lnkd.in/g_tAXJP5 #DeepLearning #ComputerVision #MedicalAI #BrainTumorDetection #UNet #CNN #Research #AI #HealthcareInnovation

Eppler J-B, Kaschube M, Rumpel S (2025) Statistical learning and representational drift: A dynamic substrate for memories. Curr. Opin. Neurobiol. 94:103107 In many brain areas, neurons exhibit continuous changes in their tuning properties over days, even when supporting stable percepts and behaviors–a phenomenon termed representational drift. How do neuronal circuits maintain stable function when their constituent elements are in constant flux? Here, we review recent theoretical and experimental work on interconnected levels, ranging from perpetual changes in synapses driving drifts in tuning of individual neurons to emergent stability at the population level, preserving similarities of activity patterns associated to specific percepts or behaviors. We propose that statistical learning, beyond its well-established roles during development and adaptation to new contexts, is also essential under steady behavioral and environmental conditions to safeguard the stability of representational similarities. We discuss implications for learning, memory, and forgetting. This framework reconciles the apparent paradox between unstable neural activity and stable perception, suggesting that representations are maintained through dynamic processes rather than static neural codes. https://lnkd.in/eZS9VQzS

🔥🚀 Preprint Alert — CausalMamba 🧠✨ We introduce CausalMamba, a scalable and biophysically grounded framework for neural causal inference from fMRI — tackling one of neuroscience’s most ill-posed inverse problems. https://lnkd.in/grASxWuV 💡 Instead of directly guessing “who causes whom” from hemodynamically distorted BOLD signals, CausalMamba decomposes the challenge into two tractable stages: 1️⃣ Differentiable BOLD Deconvolution — recovering latent neural activity via region-specific HRF estimation. 2️⃣ Causal Graph Inference — powered by a novel Conditional Mamba state-space model that captures both universal temporal dynamics and region-specific heterogeneity. 📊 Highlights 🔁 37 % higher accuracy than Dynamic Causal Modeling (DCM) on simulated data 🧩 88 % recovery of canonical neural pathways (V1 → V2 → V4) on Human Connectome Project fMRI — while classical methods fail in > 99 % of subjects 🧠 Reveals stimulus-dependent hub shifts in working-memory networks (e.g., DLPFC ↔ Insula reconfiguration) that traditional models completely miss ⚙️ Linear scalability (O(N)) → enabling practical large-scale brain-network analysis 🧬 By bridging neurovascular physics and state-space deep learning, CausalMamba offers a new path toward foundation models of directed brain dynamics — connecting structure, function, and causality at scale. Grateful to Cha Jiook for his mentorship and inspiring discussions that made this work possible :) 📄 “CausalMamba: Scalable Conditional State-Space Models for Neural Causal Inference” #NeuroAI #CausalInference #fMRI #Mamba #BrainConnectivity #DeepLearning #ComputationalNeuroscience #AIforScience

In their NeuroView, out in Neuron this week, Ellie Pavlick and Thomas Serre ask, "Will AI foundation models transform neuroscience?" https://lnkd.in/exwKCWPP