←
Aitor González-Marfil43 posts
Aitor González-Marfil avatar

Aitor González-Marfil

@AAitorG

PhD student | Deep Learning | Computer Vision | Biomedical Images

◷ Joined February 2022▧ 6 Photos and videos
Aitor González-Marfil avatar
Aitor González-Marfil@AAitorG·Nov 10

Here it is, PhenoMe in action! 🎬 Load, process and explore your data in seconds. Spot outliers, patterns, and relationships, even among the metadata you have (dose, time, temperature). Analyze faster, test smarter, and build your next hypothesis with confidence.

0:20
Estibaliz Gómez de Mariscal, PhD avatar
⟳ Aitor González-Marfil retweeted
Estibaliz Gómez de Mariscal, PhD@gomez_mariscal·Nov 4

PhenoMe is model-agnostic, supporting the use of Foundation Models and representation learning in life sciences research. It integrates ready-to-use pretrained DINOv2 and STED-FM, and functionality to process images with any PyTorch model

Post attachment
Estibaliz Gómez de Mariscal, PhD avatar
⟳ Aitor González-Marfil retweeted
Estibaliz Gómez de Mariscal, PhD@gomez_mariscal·Nov 4

♥️ Connecting all these features, it is then possible to run cluster analysis, cross-correlation analysis (supporting interpretation of the results), analysis of dynamic processes, and retrieve image prototypes that represent conditions, replicas, clusters or detected phenotypes!

Post attachment
Estibaliz Gómez de Mariscal, PhD avatar
⟳ Aitor González-Marfil retweeted
Estibaliz Gómez de Mariscal, PhD@gomez_mariscal·Nov 4

#PhenoMe processes🔬 images (2D, multichannel, time-lapse, with/without masks) with a pretrained vision model (e.g. DINOv2) and displays the latent space representation interactively (Jupyter&Napari). This lets you explore new data acquisitions and identify existing phenotypes😎

Post attachment
Estibaliz Gómez de Mariscal, PhD avatar
⟳ Aitor González-Marfil retweeted
Estibaliz Gómez de Mariscal, PhD@gomez_mariscal·Nov 4

🚀Our latest work on AI embeddings and phenotyping is out! #PhenoMe is a new interactive tool bridging image representations derived from #FoundationModels with explainable features and biological readouts🕸️ Congrats to @AAitorG on this exciting work!🥳 biorxiv.org/content/10.64898…

Post attachmentPost attachment

What if you could explore thousands of microscopy images as an interactive map, without training a single model? 🔬 I'm really excited to share our new framework and preprint: "PhenoMe: An interactive AI framework for interpretable, microscopy-driven phenotyping" 🧵👇

0:20
Aitor González-Marfil avatar
Aitor González-Marfil@AAitorG·Nov 3

Together we built a tool I feel very proud of, and I hope the bioimaging community finds it useful! 😊 try it out: pip install phenome or using @napari_imaging: pip install phenome-gui

Aitor González-Marfil avatar
Aitor González-Marfil@AAitorG·Nov 3

Across 7 public datasets, these embeddings captured cell types, mechanisms of action, dose responses, antibiotic dynamics, and cell-cycle trajectories without task-specific retraining.

Aitor González-Marfil avatar
Aitor González-Marfil@AAitorG·Nov 3

Map your dataset in 2D and click any point or cluster to inspect the images behind it. Correlate embedding axes with intensity, texture, shape or metadata (dose, time, cell line) to see why groups separate. Spot acquisition artefacts and outliers before they skew the analysis.

Aitor González-Marfil avatar
Aitor González-Marfil@AAitorG·Nov 3

Pretrained vision models (DINOv2, Cell-DINO, STED-FM) already capture rich visual patterns in microscopy images. Turning those embeddings into biological insight is still hard. PhenoMe links them directly to your experimental metadata and classical morphological measurements.

Aitor González-Marfil avatar
Aitor González-Marfil@AAitorG·Nov 3

What if you could explore thousands of microscopy images as an interactive map, without training a single model? 🔬 I'm really excited to share our new framework and preprint: "PhenoMe: An interactive AI framework for interpretable, microscopy-driven phenotyping" 🧵👇

0:20
Lukas Kuhn avatar
⟳ Aitor González-Marfil retweeted
Lukas Kuhn@lukaskuhn77·Nov 2

Introducing LeVJEPA 1.5 Large, the first step in scaling our method. It closes the gap to V-JEPA on both image and video benchmarks, trained with 5.6x less compute and over an order of magnitude less data.

0:35