Comparative Analysis of Supervised and Self-Supervised Learning for Medical Imaging Datasets

Strategic Insights -

Researchers from the University of Florence have conducted a comprehensive study on the performance of supervised learning (SL) and self-supervised learning (SSL) on small, imbalanced medical imaging datasets. The study, sponsored by NextGenerationEU, highlights the importance of carefully selecting learning paradigms based on specific application requirements. The findings indicate that SL outperformed SSL in most experiments involving small training sets, even when a limited portion of labeled data was available.

Key Takeaways:

  • The study compared the performance of SL and SSL on four binary classification tasks: age prediction, diagnosis of Alzheimer's disease, pneumonia, and retinal diseases associated with choroidal neovascularization.
  • The experiments were conducted on four different medical imaging datasets with a mean size of training sets ranging from 843 to 33,484 images.
  • The researchers tested various combinations of label availability and class frequency distribution, repeating the training with different random seeds to assess result uncertainty.
  • In most experiments involving small training sets, SL outperformed the selected SSL paradigms, even when a limited portion of labeled data was available.
  • The study concluded that carefully selecting learning paradigms based on specific application requirements is crucial due to factors such as training set size, label availability, and class frequency distribution.
  • The authors included Chiara Marzi, Andrea Espis, and Stefano Diciotti in the research.
  • The publication of the study was sponsored by NextGenerationEU.

Statistics:

  • The study experimented with four binary classification tasks.
  • The mean size of training sets ranged from 843 to 33,484 images for the four different medical imaging datasets.
  • The researchers repeated the training with different random seeds to assess result uncertainty.
  • SL outperformed SSL in most experiments involving small training sets.
  • The study found that careful selection of learning paradigms based on specific application requirements is crucial.

Sources:

  • ScienceDirect. Comparative analysis of supervised and self-supervised learning with small and imbalanced medical imaging datasets. Scientific Reports, 2025;15(1):32345.
  • Nature Portfolio. Nature Portfolio, Heidelberger Platz 3, Berlin, 14197, Germany.
  • NewsRx. Studies in the Area of Technology Reported from University of Florence (Comparative analysis of supervised and self-supervised learning with small and imbalanced medical imaging datasets). Journal of Engineering. September 15, 2025; p 3452.