Deep learning-based detection of murine congenital heart defects from µCT scans

AbstractCongenital heart defects (CHD) result in high morbidity and mortality rates, but their origins are poorly understood. Mouse models of heart morphogenesis are required to study the pathological mechanisms of heart development compared to normal. In mouse fetuses, CHD can be observed and detected in 3D images obtained by thoracic micro-computed tomography (μCT). However, diagnosis of CHD from μCT scans is a time-consuming process that requires the experience of senior experts. An automated alternative would thus save time, empower less experienced investigators and could broaden analysis to larger numbers of samples.Here, we describe and validate an approach based on deep learning to automatically segment the heart and screen normal from malformed hearts in mouse μCT scans. In an initial cohort, we collected 139 μCT scans from thorax and abdomen of control and mutant perinatal mice. We trained a self-configurating neural network (nnU-Net) to segment hearts from body μCT scans and validated its performance on expert segmentations, achieving a Dice coefficient of 96%. To identify malformed hearts, we developed and trained a 3D convolutional neural network (CNN) that uses segmented μCT scans as inputs. Despite the relatively small training data size, our diagnosis model achieved a sensitivity, specificity (for a 0.5 threshold), and area under the curve (AUC) of 92%, 96%, and 97% respectively, as determined by 5-fold cross-validation.As further validation, we analyzed two additional cohorts that were collected after the model was trained: a ‘prospective’ cohort, using the same experimental protocol as the initial cohort, and containing a subset of its genotypes, and a ‘divergent’ cohort in which mice were subjected to a different treatment for heart arrest (cardioplegia) and that contained a new mouse line. Performance on the prospective cohort was excellent, with a sensitivity of 92%, a specificity of 100%, and an AUC of 100%. Performance on the divergent cohort was moderate (sensitivity: 69%, specificity: 80% and AUC: 81%), but was much improved when the model was finetuned on (a subset of) the cohort (sensitivity: 79%, specificity: 88% and AUC: 91%). These results showcase our model’s robustness and adaptability to technical and biological differences in the data, highlighting its usefulness for practical applications.In order to facilitate the adoption, adaptation and further improvement of these methods, we built a user-friendlyNapari plugin(available atnapari-hub.org/plugins/mousechd-napari) that allows users without programming skills to utilize the segmentation and diagnosis models and re-train the latter on their own data and resources. The plugin also highlights the cardiac regions used for the diagnosis. Our automatic and retrainable pipeline, which can be employed in high-throughput genetic screening, will accelerate diagnosis of heart anomalies in mice and facilitate studies of the mechanisms of CHD.