Ensemble learning and test-time augmentation for the segmentation of mineralized cartilage versus bone in high-resolution microCT images

High-resolution non-destructive 3D microCT imaging allows the visualization and structural characterization of mineralized cartilage and bone. Deriving statistically relevant quantitative structural information about these tissues, however, requires automated segmentation procedures, mainly because manual contouring is user-biased and time-consuming. Despite the increased spatial resolution in microCT 3D volumes, automatic segmentation of mineralized cartilage versus bone remains non-trivial since they have similar grayscale values. Our work investigates how reliable 2D segmentation masks can be predicted automatically based on a (set of) convolutional neural network(s) trained with a limited number of manually annotated samples. To do that, we compared different strategies to select the 2D samples to annotate and considered ensemble learning and test-time augmentation (TTA) to mitigate the limited accuracy and robustness resulting from the small number of annotated training samples. We show that, for a fixed amount of annotated image samples, 2D microCT slices to annotate should preferably be selected in distinct 3D volumes, at regular intervals, rather than being grouped in adjacent slices of a same 3D volume. Two main lessons are drawn regarding the use of ensembles or TTA instead of a single model. First, ensemble learning is shown to improve segmentation accuracy and to reduce the mean and standard deviation of the absolute errors in cartilage characteristics obtained with different initializations of the neural network training process. In contrast, TTA appears to be unable to improve the model’s robustness to unlucky initializations. Second, both TTA and ensembling improved the model’s confidence in its predictions and segmentation failure detection. •For given annotation resources, better sample 2D slices in distinct 3D microCT volumes.•Ensemble learning improves model accuracy and robustness to unlucky initializations.•Ensemble learning and TTA improve the model’s confidence in its predictions.•Model failure can be located from the inconsistency of TTA or ensembling predictions.