We are pleased to announce an upcoming talk by Dr. Enzo Ferrante, taking place on April 15th at 14:00 in the seminar room on the 9th floor, Martensstr. 3, Erlangen. The talk will also be accessible online via Zoom: https://fau.zoom-x.de/j/69178930707?pwd=nkSabGusROQifRrXZ9sEtWpuciGyl1.1
Talk Title
Towards robust anatomical segmentation of medical images
Bio
Dr. Ferrante completed his PhD in Computer Science at Université Paris-Saclay and INRIA (Paris, France), carried out postdoctoral research at Imperial College London (UK), and earned a Systems Engineering degree from UNICEN University (Tandil, Argentina). He has also been a visiting PhD student at Stanford University, a Fulbright Visiting Researcher at Harvard Medical School in Boston, and an Invited Professor at Université Paris-Saclay, France.
His research interests span machine learning for computer vision and NLP, with a current focus on fairness and robustness in healthcare applications and biomedical imaging.
He is currently a faculty researcher at leading a group at the Applied Artificial Intelligence Lab of Argentina’s National Research Council (CONICET) and University of Buenos Aires; the Academic Leader of AnyoneAI, an EdTech startup focusing on creating AI talent in LATAM; and the Head of Machine Learning at Apolo Biotech, a startup focusing on RNA design to boost crop production. His research has been recognized with several awards, such as the Google Award for Inclusion Research, the Distinguished International Associate Award from the UK Royal Academy of Engineering, and the Friederich Wilhelm Bessel Award from the Von Humboldt Foundation, among others.
Abstract
Deep learning models for medical image segmentation typically optimize pixel-level objectives, making it difficult to enforce global constraints on shape, topology, and spatial coherence. In this talk, I will present our efforts towards building robust and anatomically plausible segmentation models along two interconnected research axes.
First, I will introduce hybrid graph neural network architectures — HybridGNet (Gaggion et al., MICCAI, 2021; IEEE TMI, 2022) and its 3D extension HybridVNet (Gaggion et al., Medical Image Analysis, 2025) — that combine convolutional encoders with graph-based decoders to produce landmark-based segmentations with built-in topological guarantees. I will discuss image-to-graph skip connections for improved robustness under occlusion, differentiable rasterization for training with pixel-level annotations, the emergence of implicit anatomical correspondences, and uncertainty quantification via variational formulations (Cosarinsky et al., MIDL, 2026). I will also show how mesh-based cardiac representations enable imaging genetics, where unsupervised geometric deep learning on left-ventricular meshes from the UK Biobank led to the discovery of novel genetic loci associated with cardiac morphology (Bonazzola et al., Nature Machine Intelligence, 2024).
Second, I will address fairness and bias in medical image analysis. Building on CheXmask, our dataset of over 676,000 chest X-ray segmentation masks (Gaggion et al., Scientific Data, 2024), I will present methods for unsupervised bias discovery based on Reverse Classification Accuracy (Gaggion et al., MICCAI FAIMI Workshop, 2023) that anticipate performance disparities across demographic subgroups without ground-truth annotations, including a recent extension using in-context segmentation models and conformal prediction (Cosarinsky et al., IEEE TMI 2026)
References
[1] Gaggion, N., Mansilla, L., Milone, D.H. & Ferrante, E. “Hybrid graph convolutional neural networks for landmark-based anatomical segmentation.” MICCAI, 2021.
[2] Gaggion, N., Mansilla, L., Mosquera, C., Milone, D.H. & Ferrante, E. “Improving anatomical plausibility in medical image segmentation via hybrid graph neural networks.” IEEE Trans. Med. Imaging, 42(2):546–556, 2022.
[3] Gaggion, N.. “Multi-view Hybrid Graph Convolutional Network for Volume-to-mesh Reconstruction in Cardiovascular MRI.” Medical Image Analysis, 2025.
[4] Cosarinsky, M. et al. “CheXmask-U: Quantifying uncertainty in graph-based anatomical segmentation for X-ray images.” MIDL, 2026.
[5] Bonazzola, R. et al. “Unsupervised ensemble-based phenotyping enhances discoverability of genes related to left-ventricular morphology.” Nature Machine Intelligence, 6:291–306, 2024.
[6] Gaggion, N. et al. “CheXmask: a large-scale dataset of anatomical segmentation masks for multi-center chest x-ray images.” Scientific Data, Nature 11(1):511, 2024.
[7] Gaggion, N., Echeveste, R., Mansilla, L., Milone, D.H. & Ferrante, E. “Unsupervised bias discovery in medical image segmentation.” MICCAI FAIMI Workshop, 2023.
[8] Cosarinsky, M. et al. “ConfIC-RCA: Statistically Grounded Efficient Estimation of Segmentation Quality”. IEEE TMI 2026. .