Table 2.
icobrain-dl consistently achieves high overlap in segmenting different brain structures across subject age ranges, while only minimally sacrificing accuracy and sometimes even outperforming models that are tailored for specific age ranges (icobrain-dl-p for pediatric data and icobrain-dl-a for adult data).
| Dice similarity coefficient (DSC) | ||||||
|---|---|---|---|---|---|---|
| Pediatric dataset 2.p | Adult dataset 2.a | |||||
| GT vs | icobrain-dl-p vs | GT vs | icobrain-dl-a vs | |||
| icobrain-dl | icobrain-dl-p | icobrain-dl | icobrain-dl | icobrain-dl-a | icobrain-dl | |
| WM | 84.8 (2.0) | 85.0 (2.1) | 92.3 (2.9) | 88.5 (1.1) | 85.5 (1.0) | 91.8 (1.7) |
| CGM | 83.0 (2.5) | 79.7 (3.7) | 88.2 (3.2) | 83.2 (1.6) | 82.5 (1.6) | 89.9 (2.3) |
| Lateral ventricles | 83.0 (4.6) | 83.4 (4.7) | 88.5 (6.4) | 88.7 (3.6) | 88.9 (3.7) | 93.3 (2.7) |
| Hippocampus | 78.8 (2.7) | 71.4 (11.3) | 82.1 (11.3) | 76.7 (1.8) | 77.2 (1.9) | 91.3 (1.7) |
| Caudate | 83.5 (3.5) | 82.5 (8.3) | 83.5 (9.3) | 81.8 (3.1) | 80.5 (3.2) | 91.5 (2.2) |
| Putamen | 84.8 (1.5) | 83.9 (3.1) | 89.8 (3.6) | 83.4 (2.4) | 80.7 (1.8) | 92.1 (1.6) |
| Cerebellar GM | 83.9 (2.5) | 86.9 (1.9) | 89.9 (1.9) | 78.4 (2.8) | 79.5 (2.4) | 94.1 (1.3) |
| Cerebellar WM | 76.1 (4.1) | 75.9 (4.4) | 81.1 (2.8) | 77.8 (2.8) | 77.4 (2.3) | 92.4 (1.6) |
| Thalamus | 82.6 (2.4) | 78.7 (7.0) | 90.7 (6.7) | 85.0 (1.0) | 84.5 (1.3) | 94.7 (1.0) |
We compare segmentation accuracy as measured by the Dice similarity coefficient expressed as a percentage between three deep learning models that are trained using different subsets of training data against manually created ground truth (GT) from pediatric (CANDIShare, dataset 2.p) and adult (MICCAI2012, dataset 2.a) data. The Dice similarity coefficient between the models’ predictions is also shown. The Dice similarity coefficient is reported as: mean value (standard deviation) across subjects. WM = White Matter, CGM = Cortical Gray Matter, GM = Gray Matter.