Fig. 3: Evaluation of deep learning-based heart segmentation from μCT scans.

a Heart segmentation results (evaluated by an expert as perfect), shown on a single slice of a μCT scan: original image (no overlay), manual segmentation (blue overlay), automated segmentation by the nnU-Net (yellow overlay), and overlap between automated and manual segmentations (red overlay). b Segmentation performance metrics for the five models trained by nnU-Net on five data folds. Dice coefficient, recall, and precision are shown (see Methods). Ensemble corresponds to the pixel-wise average over the five folds. c Horizontal bars show the nnU-net segmentation performance evaluated by an expert: an embryologist visually assessed heart segmentations from n = 111 mice (28 hearts used for training were excluded) and labelled 96 hearts (86%) as perfectly segmented. d, e Vertical bars show the number of perfectly (pale green) or imperfectly (dark grey) segmented hearts according to stage (d) or label (CHD or normal) (e) in the same data set of n = 111 mice. The proportion of perfectly segmented hearts was not significantly associated with developmental stage (Fisher test p = 1), and was marginally significantly associated with the presence of CHD (Fisher test p = 0.048). f Boxplots compare the heart volumes as measured from automated segmentations in n = 139 mice for the two developmental stages. Green dots correspond to normal hearts, red dots to hearts with CHD. Scale bar in (a): 4 mm. Source data is available in Supplementary Data 1.