University of Bamberg
open search form
You are here

Information Systems and Applied Computer Sciences

Chair of Explainable Machine Learning

  1.  University of bamberg
  2. Faculties
  3. Information Systems and Applied Computer Sciences
  4. Subject groups
  5. Applied Computer Science
  6. Chair of Explainable Machine Learning
  7. News
  8. Rethinking Model Prototyping: Our Latest Paper in Scientific Reports
2025 Springer Nature Limited

Rethinking Model Prototyping: Our Latest Paper in Scientific Reports

The integration of deep learning in clinical practice remains a challenge due to the limited and heterogeneous nature of medical datasets. Our latest paper, Rethinking model prototyping through the MedMNIST+ dataset collection, published in Scientific Reports, addresses this issue by introducing a comprehensive benchmarking framework designed to enhance the evaluation of machine learning models in medical imaging.

A New Benchmark for Medical Image Classification

In our latest paper, we introduce an extensive benchmarking framework that systematically evaluates deep learning architectures across a diverse set of medical imaging datasets. Our benchmark spans 12 datasets, covering nine imaging modalities and anatomical regions, while incorporating multiple classification tasks, input resolutions, and varied sample sizes. By assessing ten commonly used Convolutional Neural Network (CNN) and Vision Transformer (ViT) architectures under different training methodologies, we provide a well-rounded perspective on model effectiveness and development strategies.

Our findings challenge several prevailing assumptions in the field. Notably, CNNs continue to demonstrate strong performance, holding their own against ViTs in medical image classification. Additionally, we show that computationally efficient training schemes and modern foundation models present promising alternatives to costly end-to-end training. Interestingly, we also observe that increasing image resolution does not always translate to better performance; in many cases, lower resolutions suffice for prototyping while significantly reducing computational overhead. These insights emphasize the need for broad and standardized benchmarking approaches to drive clinically relevant innovations and ensure model robustness.

For more details, you can access the paper here and the associated code repository, including benchmarks, here.

About Scientific Reports

Our work is published in Scientific Reports, a prestigious open-access journal that covers original research across the natural sciences, psychology, medicine, and engineering. As the 5th most-cited journal in the world, with over 734,000 citations in 2023, Scientific Reports provides a highly respected platform for impactful research. The journal ensures high visibility through open access, making studies immediately available to a global audience. It receives approximately 2.7 million visitors per month and is indexed in major databases such as Web of Science, PubMed, Scopus, and Google Scholar. The Journal Impact Factor stands at 3.8, with a five-year impact factor of 4.3, underscoring its influence in the scientific community.

 

xAILab Bamberg