Deep learning reconstruction and ADC: is quantitative comparability preserved?

Deep learning (DL)-based reconstruction can improve image quality and acquisition efficiency in breast diffusion-weighted imaging (DWI), but its effects on quantitative biomarkers warrant careful consideration. Although recent work reported improved image quality without significant differences in mean apparent diffusion coefficient (ADC), statistical nonsignificance does not necessarily establish quantitative equivalence or interchangeability across reconstruction pipelines. Reconstruction related changes in noise characteristics, signal intensity, spatial resolution, and ADC distributions may affect quantitative measurements despite stable group-level mean ADC. Recent phantom evidence further suggests that DL reconstruction can preserve ADC accuracy while altering ADC distributional characteristics. Therefore, evaluation of DL-reconstructed DWI should extend beyond mean ADC comparisons to include repeatability, reproducibility, agreement, and lesion-level analyses. Reconstruction specific validation is essential for reliable cross-method application of ADC as a quantitative imaging biomarker.