Fig 1.
Osteocytes and dendrites manually labelled to train deep learning models.
(A) Confocal laser scanning microscopy images of osteocytes (red, phalloidin, cytoskeleton). (B) Images were greyscaled for training, (C) followed by manual annotation and labelling of osteocytes and dendrites. Raw images contained features that presented significant challenges to deep learning models, including (D) shrunken osteocytes possibly fixed during apoptosis,(E) transiting blood vessels, and (F) locations of fused dendrites or the edge of an out-of-plane osteocyte. Scale bar = 10 µm and applies to all panels.
Table 1.
Segmentation time, Dice score, mean IoU and IoU for individual classes. The methods M1 and M2 are derived from (Kerschnitzki, 2013; Mabilleau et al., 2016) [30,40], and (Ashique et al., 2017) [41] respectively, Otsu and Canny represent our non-DL segmentation methods.
Table 2.
Performance metrics of implemented models on two different training strategies.
Table 3.
Performance metrics of experiments attempting to improve dendrite segmentation accuracy (L-Reg: Local affine registration method).
Fig 2.
Connectomics used as training strategy for deep learning models.
(A) Annotated labels in an image, and (B) the corresponding connectomics map. After training, model is tested to produce a (C) prediction and (D) corresponding connectomics map, which can be compared to the annotated map to assess accuracy. Red squares in the prediction highlight instances of disconnected dendrite segments, while the blue square indicates a misclassification where a dendrite has been erroneously labeled as an osteocyte. Scale bar = 10 µm and applies to all panels.
Table 4.
Connectivity metrics computed from the connectomics analysis. The average is taken from the validation set consisting of 7 images.
Fig 3.
Deep learning model can accurately detect and identify the lower osteocyte connectivity of aged bone, when compared to young bone.
These results are similar to changes we measured manually in a previous study [19]. (A) Confocal laser scanning microscopy images of osteocytes (red, phalloidin, cytoskeleton) from young (2 month old) and (B) aged (36 month old) mice. Running the trained deep learning model on these datasets correctly measured that aging causes significant decreases in (C) average network connectivity, (D) osteocyte density, and (E) average dendrite length. (F) The model also correctly predicted significant increases in blunted dendrites or dead ends in the network. **** denotes p < 0.0001 by Mann-Whitney U test compared to young control. Scale bar = 10 µm.
Fig 4.
Deep learning model can detect and identify the lower osteocyte connectivity of bones in which TGF-β Receptor II has been ablated from osteocytes, compared with control mice.
Some of these results (decreased connectivity and number of dead ends) are similar to those measured by us previously, while others (decreased osteocytes and dendrite length) are not fully captured [19]. (A) Confocal laser scanning microscopy images of osteocytes (red, phalloidin, cytoskeleton) from 2 month old control mice (TβRIIctrl) and (B) their litter-mates in which TGF-β Receptor II was ablated from osteocytes (TβRIIocy-/-). Running the trained deep learning model on these datasets correctly measured that aging causes significant decreases in average network connectivity and average number of blunted canaliculi in the network, but not previously observed decreases in osteocyte density and average dendrite length. * denotes p < 0.05 by Mann-Whitney U test compared to control, n.s. denotes no statistical significance. Scale bar = 10 µm and applies to all panels.
Fig 5.
Deep learning model predicts changes in osteocyte network parameters due to age with similar accuracy to measurements performed manually.
Direct comparison of measurements of percentage decrease in (A) osteocyte density and (B) osteocyte network connectivity as calculated manually by us previously [19], and by our deep learning model. Very similar decreases in osteocyte density were measured (38.03% vs. 34.44%, respectively) while a trend of similar results were found in connectivity of the osteocyte network (43.3% vs. 55.57%, respectively).