Deep learning-based detection of murine congenital heart defects from µCT scans.
Authors
Affiliations (8)
Affiliations (8)
- Institut Pasteur, Université Paris Cité, Imaging and Modeling Unit, Paris, France.
- Integrated Drug Discovery, Sanofi R&D, Paris, France.
- Université Paris Cité, Imagine - Institut Pasteur Unit of Heart Morphogenesis, INSERM UMR1163, Paris, France.
- Unité Médico-Chirurgicale de Cardiologie Congénitale et Pédiatrique, M3C-Necker, Hôpital Universitaire Necker-Enfant-Malades, APHP, Paris, France.
- Université Paris Cité, Imagine - Institut Pasteur Unit of Heart Morphogenesis, INSERM UMR1163, Paris, France. [email protected].
- Institut Pasteur, Université Paris Cité, Imaging and Modeling Unit, Paris, France. [email protected].
- University of Würzburg, Rudolf Virchow Center for Integrative and Translational Bioimaging, Würzburg, Germany. [email protected].
- University of Würzburg, Center for Artificial Intelligence and Data Science, Würzburg, Germany. [email protected].
Abstract
Micro-computed tomography (μCT) provides 3D images of congenital heart defects (CHD) in mice. However, diagnosing CHD from μCT scans is time-consuming and requires clinical expertise. Here, we present a deep learning approach to automatically segment and screen normal from malformed hearts. On a cohort of 139 μCT scans of control and mutant mice, our diagnosis model achieves an area-under-the-curve (AUC) of 97%. For further validation, we acquired two additional cohorts after model training. Performance on a similar 'prospective' cohort is excellent (AUC: 100%). Performance on a 'divergent' cohort containing novel genotypes is moderate (AUC: 81%), but improves markedly after model finetuning (AUC: 91%), showcasing robustness and adaptability to technical and biological differences in the data. A user-friendly Napari plugin allows researchers without coding expertise to utilize and retrain the model. Our pipeline will accelerate diagnosis of heart anomalies in mice and facilitate mechanistic studies of CHD.