Untrained Convolutional Neural Networks Match Pretrained Models for MRI Feature Extraction
Researchers demonstrated that untrained convolutional neural networks (un-CNN) can extract features from structural MRI scans with performance comparable to or better than state-of-the-art pretrained models across multiple datasets and tasks. Untrained CNNs offer practical advantages including lower computational costs, reduced memory requirements, and elimination of data leakage risks. This finding could simplify MRI analysis workflows and improve reproducibility in medical imaging research.
A new study published on bioRxiv shows that untrained convolutional neural networks can effectively extract features from structural MRI data without requiring pre-training on large datasets. The un-CNN architecture incorporates multi-channel inputs, hierarchical encoding with multi-scale feature aggregation, and covariance pooling to achieve competitive or superior performance compared to established pretrained foundation models. Testing across three structural MRI datasets and three downstream tasks confirmed the approach's effectiveness. The method addresses several practical limitations of trained models: it dramatically reduces computational and memory demands, eliminates the need to distribute large model weights across systems, removes risks of inadvertent data leakage during transfer learning, and enhances reproducibility. These advantages could make MRI-based feature extraction more accessible and reliable for clinical and research applications.
Limitations & open questions
The study's own limitations are not detailed in the provided abstract, such as specific dataset characteristics, the three downstream tasks evaluated, quantitative performance comparisons with baseline methods, or potential constraints of the un-CNN approach that warrant further investigation.
What different sources said
- bioRxivCenter
Untrained Convolutional Neural Networks as Feature Extractors for Structural MRI
Related
NASA's Curiosity Rover Continues Journey Toward Yardang Unit on Mars
NASA's Curiosity rover drove across sols 4913-4919 on Mars, moving between science campaigns toward a distant yardang unit—a series of wind-sculpted pale hills. The rover is in a transit phase between major science campaigns, allowing for flexible exploration of interesting geological features along the way. This ongoing exploration helps scientists understand Mars' geological history and past environmental conditions.
World on track to breach 1.5°C warming threshold by 2030 at current emission rates, major climate report warns
A comprehensive climate report by over 70 scientists finds that human-induced warming has reached 1.37°C in 2025, with the 1.5°C Paris Agreement limit likely to be crossed around 2030 if emissions continue at current levels. Global greenhouse gas emissions hit a record 56.8 billion tonnes of CO2 equivalent in 2024, driven primarily by fossil fuel burning, while the planet's carbon budget for staying below 1.5°C will be exhausted in approximately three years. The findings underscore the urgency of transitioning to renewable energy and reducing emissions to prevent the most severe climate impacts.
Study Shows Brain's Functional Architecture Shifts from Cognition to Anatomy During Loss of Consciousness
Researchers analyzed fMRI data alongside eight types of brain connectivity measures to understand what shapes how brain regions interact. They found that in awake brains, cognitive activity is the dominant driver of functional connectivity, but this relationship reverses during unconsciousness—whether induced by anesthesia or caused by disorders of consciousness. The findings suggest consciousness fundamentally reorganizes how the brain's functional architecture is organized.