Deep Learning Model for Biomedical Image Analysis
D. Sage
Keynote address, Colloque Français d'Intelligence Artificielle en Imagerie Biomédicale (IABM'25), Nice, French Republic, March 17-18, 2025.
Deep learning (DL) models have profoundly reshaped biomedical imaging, transforming tasks such as image reconstruction, restoration, processing, segmentation, and quantitative analysis. Despite their apparent simplicity—being trained solely on example images—DL models present significant challenges, risks, and require advanced expertise.
In this presentation, we will explore key considerations for deploying DL models, stressing the shift from model-based approaches to data-driven methods. The first critical step is to check whether deep learning is truly necessary. If so, designing a DL image analysis pipeline involves balancing task-specific decomposition versus end-to-end systems and determining the learning strategy—supervised or self-supervised.
The well-known architectures used in imaging, including CNNs, U-Nets, GANs, and Vision Transformers, will be reviewed. Using examples from bioimaging field, we will then demonstrate how to choose between training from scratch, fine-tuning, or directly using pre-trained models, including foundation models such as Segment Anything Model.
The importance of creating valid and representative training datasets will also be highlighted, as this remains a persistent bottleneck in biomedical contexts where annotated data are rare. Finally, we will address broader challenges, including trust, ethical data usage, and the sustainable use of computational resources.
@INPROCEEDINGS(http://bigwww.epfl.ch/publications/sage2501.html,
AUTHOR="Sage, D.",
TITLE="Deep Learning Model for Biomedical Image Analysis",
BOOKTITLE="Colloque Fran{\c{c}}ais d'Intelligence Artificielle en
Imagerie Biom{\'{e}}dicale ({IABM'25})",
YEAR="2025",
editor="",
volume="",
series="",
pages="",
address="Nice, French Republic",
month="March 17-18,",
organization="",
publisher="",
note="Keynote address")