Fine-tune Gemma 4 models for state-of-the-art performance

Modern AI was accidentally brute-forced by crypto greed demanding faster GPUs. The second Amodei realizes a cancer cure will leave thousands of oncologists unemployed, he’ll have Claude ship it in the next release.

Swedish scientists have invented next-gen, miniature biosensors – through integrating lasers and optics, compacted into a centimetre-sized single semiconductor chip. This invention can make at-home medical investigations feasible. https://lnkd.in/gG3fH7vN

I recently developed a CT-based radiation dose simulation framework using Monte Carlo photon transport and deterministic beam modeling. The project focuses on simulating how radiation interacts with heterogeneous human tissue using real CT (DICOM) imaging data. Pixel values are converted into Hounsfield Units (HU), enabling tissue classification into air, soft tissue, and bone, which are then mapped to material-specific attenuation coefficients. Two models were implemented: • A deterministic model based on exponential attenuation (Beer–Lambert law) • A Monte Carlo simulation capturing stochastic photon interactions, scattering, and energy deposition A key extension of the system is targeted beam delivery, where photons are initialized near a tumor region embedded in soft tissue. This allows comparison between random and targeted irradiation strategies. Key findings: • ~40× increase in mean tumor dose using targeted beam delivery   • Monte Carlo model produces realistic heterogeneous dose distributions   • Underdose probability analysis (~40%), highlighting limitations of single-beam approaches   • Strong dependence of dose distribution on tissue composition and density  The project also includes an interactive Streamlit application for real-time visualization of CT images, tissue segmentation, and dose maps. GitHub: https://lnkd.in/dwDsDK6T This project combines medical imaging, physics-based modeling, and computational simulation to demonstrate principles used in radiotherapy planning systems. #MedicalPhysics #Python #MonteCarloSimulation #ComputationalModeling #MedicalImaging #Radiotherapy

X01 addresses the problem of inconsistent results on the mechanical properties of ultrasoft materials such as brain tissue by reconciling data from different ex vivo and in vivo testing techniques, including MRI, MRE, vibration tests, rheometry, indentation, and AFM. The goal is to develop a continuum-based model that unifies different experimental observations using advanced simulations and validations of experimental data from phantoms, animal brain tissue samples, and in vivo mouse and human brains to predict disease-induced changes in cerebral (poro-)visco-elastic properties.

Today, one project will be named the winner of the Photonics Frontiers Award 2026 at Optatec Messe. And with a shortlist showcasing some of the most exciting applied photonics innovation in the world today, anticipation is building ahead of this afternoon’s announcement. From quantum photonic co-processors and AI-powered optical interconnects to breakthrough cancer diagnostics, advanced industrial laser systems, and next-generation sensing technologies — this year’s finalists represent the cutting edge of applied photonics innovation. These are not concepts. They are real-world technologies tackling real-world challenges right now. The Photonics Frontiers Award was created to recognise projects driving measurable impact across industry, science, healthcare, communications, manufacturing, aerospace, and beyond — and the 2026 shortlist showcases just how rapidly photonics is reshaping the future. Who will take the title? Will it be: 🔹 The photonic co-processor described as an industry first? 🔹 A breakthrough AI data centre interconnect technology? 🔹 A new approach to laser manufacturing powered by machine learning? 🔹 A photonics innovation transforming medical diagnostics and surgery? The winner will be revealed today at: 📍 Optatec 2026 — Exhibitor Forum, Booth 625 🕞 3:30 PM Join us for the official award ceremony as we celebrate the projects, collaborations, and innovators pushing photonics into entirely new frontiers. With finalists spanning biophotonics, quantum technologies, industrial imaging, optical communications, sensing, aerospace, and laser systems, this year’s award reflects the extraordinary breadth — and growing influence — of photonics across the global technology landscape. If you’re attending Optatec, don’t miss it.

Today, one project will be named the winner of the Photonics Frontiers Award 2026 at Optatec Messe. And with a shortlist showcasing some of the most exciting applied photonics innovation in the world today, anticipation is building ahead of this afternoon’s announcement. From quantum photonic co-processors and AI-powered optical interconnects to breakthrough cancer diagnostics, advanced industrial laser systems, and next-generation sensing technologies — this year’s finalists represent the cutting edge of applied photonics innovation. These are not concepts. They are real-world technologies tackling real-world challenges right now. The Photonics Frontiers Award was created to recognise projects driving measurable impact across industry, science, healthcare, communications, manufacturing, aerospace, and beyond — and the 2026 shortlist showcases just how rapidly photonics is reshaping the future. Who will take the title? Will it be: 🔹 The photonic co-processor described as an industry first? 🔹 A breakthrough AI data centre interconnect technology? 🔹 A new approach to laser manufacturing powered by machine learning? 🔹 A photonics innovation transforming medical diagnostics and surgery? The winner will be revealed today at: 📍 Optatec 2026 — Exhibitor Forum, Booth 625 🕞 3:30 PM Join us for the official award ceremony as we celebrate the projects, collaborations, and innovators pushing photonics into entirely new frontiers. With finalists spanning biophotonics, quantum technologies, industrial imaging, optical communications, sensing, aerospace, and laser systems, this year’s award reflects the extraordinary breadth — and growing influence — of photonics across the global technology landscape. If you’re attending Optatec, don’t miss it.

Photon Counting CT -PCCT - provides up to 27 times better resolution than conventional CT, allowing for imaging fine structures in high detail. A game changer technology , a real revolution in cardiac imaging, a faster and reliable way to better definition of cardiac tissue and structure.

Quantum Computing & Cancer Detection For decades, cancer detection has been viewed as a race against time. The sooner you detect it, the better your chances. Yet, beneath that urgency is a significant limitation-our struggle to truly understand what happens at the most basic level of the human body. We don’t catch cancer early because we fully grasp it; we catch it early because we’ve learned to spot patterns just before they become disastrous. This is where quantum computing starts to change the conversation-not by speeding up current methods but by changing how we approach the problem. Classical computers, no matter how powerful, face limits.....(to continue reading visit-https://lnkd.in/gGdpwFiB

New Post: Design and Evaluation of a Compact Adaptive Beamforming Ultrasound Probe for Point‑of‑Care Cardiac Screening - ## Abstract Point‑of‑care cardiac ultrasound requires probes that are portable, low‑cost, and capable of delivering image quality sufficient for clinical decision‑making. This exploratory study proposes a **synthetic** design for a 32‑element flexible linear array that integrates a low‑noise analog front‑end, a 12‑bit 50 MS/s analog‑to‑digital converter, and an embedded conjugate‑gradient adaptive beamformer running on an ARM Cortex‑M7 \[…\] \[Source & Legal Disclaimer\] This is an AI-generated simulation research dataset provided by Freederia.com, released under the Apache 2.0 License. Users may freely modify and commercially use this data \(including patenting novel improvements\); however, obtaining exclusive patent rights on the original raw data itself is prohibited. As this is AI-simulated data, users are strictly responsible for independently verifying existing copyrights and patents before use. The provider assumes no legal liability. For future Enterprise API access and bulk dataset purchase inquiries, please contact Freederia.com.

V2 is done. Here are the results. 🧪 When we entered the AMD Developer Hackathon lablab.ai, we made a promise: build the next version of research we believe in. Here's what we delivered. 📊 Final numbers — 199 I-SPY 2 patients, 5-fold stratified CV, 3 seeds: • AUC: 0.771 ± 0.011 → +6.7 pts vs V1 • Sensitivity: 0.813 — the model finds 81% of pCR patients • Best single fold: AUC 0.926 • TNBC folds: AUC up to 0.933 But the numbers don't tell the full story. Here's what actually moved the needle: 🔹 GATv2 Clinical Graph: +7 AUC points HR status, HER2, TNBC, age, and treatment arm — modeled as a graph with 16 biologically-grounded edges, not a flat vector. Zeroing the GNN and keeping only imaging drops AUC from 0.728 to 0.673. Clinical molecular features aren't redundant with DCE-MRI. They're complementary. 🔹 RadImageNet backbone: +10.6% sensitivity Swapping ImageNet for a backbone pretrained on 1.35M radiological images improved true pCR detection from 67% → 81%. That gap is the difference between a patient receiving breast-conserving surgery or a mastectomy. 🔹 GroupWise normalization: the stability breakthrough 6 PyRadiomics features had values at scales up to 10⁶ — silently suppressing the 1,024 CNN features that RadImageNet was optimized to extract. A signed log transform + winsorization per feature group fixed it. Seed-to-seed AUC variance dropped from ±0.025 → ±0.011. The biggest stability gain in the entire project came from a normalization fix, not from adding more parameters. 🔹 Only 23.8% of parameters are trained 518K trainable out of 2.17M total. The frozen Phase LSTM encoder does the heavy lifting. Efficient by design — and necessary when your dataset is 199 patients. 🔬 Molecular subgroup analysis: TNBC reaches AUC 0.933 in the best fold — consistent with the higher tumor vascularization changes visible in DCE-MRI. HR+/HER2- remains the hardest subtype: only 14 pCR+ cases in the entire dataset, and a single misclassification in a validation fold collapses AUC below chance. That's not a model failure — it's a statistical reality of small clinical datasets that we report transparently. The entire pipeline runs on AMD ROCm + PyTorch on an RX 9070 + Ryzen 5 7600X — with specific kernel-level optimizations for gfx1201 that weren't documented anywhere. No CUDA. No cloud. Built in Lima, Perú 🇵🇪 Proof that clinical-grade medical AI is achievable on AMD hardware. #AMDHackathon #lablabai #AMD #MercuryIngenium #BreastCancer #MedicalAI #GATv2 #DeepLearning #Peru #UNMSM #IEEE #GraphNeuralNetworks #pCRPrediction #ROCm