I am thrilled to announce the preprint “Untrained Convolutional Neural Networks as Feature Extractors for Structural MRI“ is available on bioRxiv.

Description

This work shows that untrained CNN features extracted from structural MRI can match or exceed the predictive performance of state-of-the-art pretrained foundation models. The proposed un-CNN architecture combines multi-channel inputs, hierarchical encoding with multi-scale aggregation, and covariance pooling, offering reduced computational and memory costs, no need for distributing model weights, lower risk of data leakage, and improved reproducibility compared to trained models.

Authors

  • Arel Encin
  • Ines Gonzalez Pepe
  • Yohan Chatelain
  • Erin Dickie
  • Tristan Glatard