Abstract
We explored effects of (1) training with various sample sizes of multi-site vs. single-site training data, (2) cross-site domain adaptation, and (3) data sources and features on the performance of algorithms segmenting cerebral infarcts on Magnetic Resonance Imaging (MRI). We used 10,820 annotated diffusion-weighted images (DWIs) from 10 university hospitals. Algorithms based on 3D U-net were trained using progressively larger subsamples (ranging from 217 to 8661), while internal testing employed a distinct set of 2159 DWIs. External validation was conducted using three unrelated datasets (n = 2777, 50, and 250). For domain adaptation, we utilized 50 to 1000 subsamples from the 2777-image external target dataset. As the size of the multi-site training data increased from 217 to 1732, the Dice similarity coefficient (DSC) and average Hausdorff distance (AHD) improved from 0.58 to 0.65 and from 16.1 to 3.75 mm, respectively. Further increases in sample size to 4330 and 8661 led to marginal gains in DSC (to 0.68 and 0.70, respectively) and in AHD (to 2.92 and 1.73). Similar outcomes were observed in external testing. Notably, performance was relatively poor for segmenting brainstem or hyperacute (< 3 h) infarcts. Domain adaptation, even with a small subsample (n = 50) of external data, conditioned the algorithm trained with 217 images to perform comparably to an algorithm trained with 8661 images. In conclusion, the use of multi-site data (approximately 2000 DWIs) and domain adaptation significantly enhances the performance and generalizability of deep learning algorithms for infarct segmentation.