Link between downstream diagnostic prediction performance and segmentation quality

Ascertain whether comparable downstream diagnostic prediction performance when using SynthSeg‑derived MRI features from synthetic 7T images and real 3T images reliably indicates comparable segmentation quality. Specifically, establish the relationship between segmentation accuracy and multiclass random forest prediction of cognitive status (CN, MCI, AD) based on SynthSeg cortical and subcortical derivatives, and determine conditions under which prediction performance reflects segmentation quality.

Background

Using a large independent set of 3,168 3T scans, the authors synthesized 7T images and performed multiclass random forest classification of clinical diagnosis (CN, MCI, AD) from SynthSeg‑derived cortical and subcortical features, finding similar accuracy and balanced accuracy across real 3T and synthetic 7T datasets.

The authors caution that equivalent downstream prediction performance may not necessarily imply comparable segmentation quality, noting that segmentation errors or biases can sometimes improve prediction. They explicitly state that this inference remains unresolved, motivating a need to directly link segmentation quality with downstream predictive utility.