Reproduce “Delving into Masked Autoencoders for Multi-Label Thorax Disease Classification” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junfei Xiao, Yutong Bai, Alan Yuille, Zongwei Zhou

Researchers at Johns Hopkins University developed customized pre-training and fine-tuning strategies for Masked Autoencoders (MAE) based Vision Transformers (ViTs), enabling them to achieve multi-label thorax disease classification performance comparable to or exceeding state-of-the-art Convolutional Neural Networks on chest X-rays.

Key insights

  • In-domain self-supervised pre-training using Masked Autoencoders is crucial for Vision Transformers to achieve competitive performance in medical imaging, outperforming ImageNet pre-training and contrastive learning methods.
  • Medical images benefit from a higher MAE masking ratio (90%) and a less aggressive random resized crop scale (0.5-1.0) during pre-training, which significantly improves efficiency and preserves subtle features.
  • While ViTs achieve comparable classification performance, CNNs generally exhibit better weakly-supervised localization for larger disease patterns, whereas ViTs demonstrate robustness in detecting smaller anomalies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.