Romain Andres

Original photo Eigenvalue decomposition Grad-CAM (CLIP) Binary digit rendering Edge detection Watershed segmentation Semantic segmentation Object detection Neural style transfer Depth estimation Face mesh ASCII art ORB keypoints Voronoi diagram Fractal quadtree Feature-based morph

Caen, Normandy, France

I’m Romain, an AI engineer passionate about applying deep learning to solve complex challenges in oncology and personalized medicine.

research focus

Over the past year, as an AI Engineer at the Centre François Baclesse (Caen), I contributed to a national federated learning project aimed at improving survival prediction for glioblastoma patients. I built the technical pipeline for this multi-center initiative — distributed training with Fed-BioMed, multi-center data engineering and anonymization, and ML Ops (Docker, deployment) — in collaboration with Inria and other leading French research institutions.

This role built directly on my Master’s thesis on the explainability (XAI) of a deep learning model for predicting treatment response in glioblastoma, where I explored methods such as Grad-CAM, LRP, and mechanistic interpretability to make AI predictions more transparent for clinicians. This work is now the subject of a first-author manuscript currently under review.

Earlier, I developed a deep learning algorithm UI based on OHIF Viewer for the segmentation of brain metastases, which resulted in a co-first author publication in NeuroImage.

I hold a Master’s degree in AI, Data Analysis and Health Access from the University of Caen, as part of the SATIN project (Health, Territory, Innovation and Digital Technology).

what’s next

I’m now looking for my next step: a CIFRE PhD or a research / AI engineer position. I’m always open to discussing opportunities aligned with my interests in explainable AI, federated learning, computer vision, and their clinical applications — see the contact links below.

beyond research

Outside of work, I maintain a self-hosted FreeBSD NAS/cloud server, tinker with Docker/Kubernetes clusters, and keep a running technology watch on new developments in AI and medicine.

latest posts

Aug 01, 2025	Explainability and Clinical Trust of Deep Learning in Glioblastoma Treatment Response
Jun 01, 2024	MetIA — Deploying a Deep Learning Interface for Brain Metastases Segmentation
Nov 01, 2023	GeneticPedigreeChartToPed — Digitizing Family Trees for Hereditary Cancer Risk

selected publications

NeuroImage

Development and routine implementation of a deep learning algorithm for automatic brain metastases segmentation on MRI for RANO-BM criteria follow-up

Loïse Dessoude, Raphaëlle Lemaire, Romain Andres, Thomas Leleu , and 10 more authors

NeuroImage, Feb 2025

Dessoude, Lemaire and Andres contributed equally (co-first authors).

Abs Bib HTML

@article{dessoude_development_2025,
  title = {Development and routine implementation of a deep learning algorithm for automatic brain metastases segmentation on {MRI} for {RANO-BM} criteria follow-up},
  author = {Dessoude, Loïse and Lemaire, Raphaëlle and Andres, Romain and Leleu, Thomas and Leclercq, Alexandre G. and Desmonts, Alexis and Corroller, Typhaine and Orou-Guidou, Amirath Fara and Laduree, Luca and Le Henaff, Loïc and Lacroix, Joëlle and Lechervy, Alexis and Stefan, Dinu and Corroyer-Dulmont, Aurélien},
  year = {2025},
  month = feb,
  journal = {NeuroImage},
  volume = {306},
  pages = {121002},
  doi = {10.1016/j.neuroimage.2025.121002},
  note = {Dessoude, Lemaire and Andres contributed equally (co-first authors).}
}

Under review
Explainability and Clinical Trust of Deep Learning in Glioblastoma Treatment Efficacy Prediction: A Comprehensive Framework

Romain Andres, Noémie Moreau, Loïse Dessoude, Loïc Le Henaff , and 5 more authors

Feb 2025

Abs Bib HTML

First-author manuscript presenting a multi-layered XAI pipeline (Grad-CAM, LRP, LIME, linear probing, causal activation patching) to audit a ResNet-51q model predicting glioblastoma treatment response. The study uncovers that the model relies on an unexpected heuristic based on a proxy of the surgical resection status, anchoring its long- versus short-survival predictions on the detection of the surgical site rather than on tumoral features. This finding was confirmed through targeted ablation studies and high-resolution saliency maps, and the overall framework was validated in a multi-reader, multi-case clinical study.
@unpublished{andres_explainability_2025, title = {Explainability and Clinical Trust of Deep Learning in Glioblastoma Treatment Efficacy Prediction: A Comprehensive Framework}, author = {Andres, Romain and Moreau, Noémie and Dessoude, Loïse and Le Henaff, Loïc and Missohou, Fernand and Stefan, Dinu and Desmonts, Alexis and Herault, Romain and Corroyer-Dulmont, Aurélien}, year = {2025}, journal = {Manuscript under review}, }