C-S.B.70: Deep Learning for BioImaging: What Are We learning ?
[sponser-meet-now-chat][/sponser-meet-now-chat]
Recent advances in representation learning have transformed natural image analysis, yet their impact on biological microscopy remains poorly understood. In this work, we systematically investigate what deep learning models actually learn from large-scale bioimaging data, focusing on two key biological contexts: cell culture imaging and tissue histology.
We benchmark a wide range of representations, including state-of-the-art pretrained vision models, domain-specific foundation models, randomly initialized networks, and simple handcrafted features. Across multiple datasets and tasks, we show that surprisingly simple or untrained representations can achieve performance comparable to advanced pretrained models. This suggests that current benchmarks may not reliably capture biologically meaningful representations, but instead exploit low-level visual cues or dataset-specific biases.
To better understand these behaviors, we analyze representation structure through dimensionality, layer-wise contributions, and robustness across biological tasks. We find that high-performing models often rely on low-dimensional signals and that shallow features can outperform deeper representations, challenging common assumptions derived from natural image domains. Furthermore, we demonstrate that biologically interpretable structure-only baselines, constructed without cell-level information, can remain competitive, reinforcing concerns about the validity of existing evaluation protocols.
Overall, our results highlight a critical gap between benchmark performance and biological relevance in bioimaging. We advocate for more rigorous evaluation strategies and stronger baselines to ensure that learned representations capture meaningful biological mechanisms rather than spurious correlations. This work provides practical guidelines for the development and assessment of deep learning models in computational biology and high-content imaging.
Co-authors: Ivan Svatko, Ihab Bendidi, Auguste Genovesio
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826