Hierarchical detection of neonatal brain pathologies using a two-stage deep learning framework from magnetic resonance imaging

Automated neonatal brain pathology detection from magnetic resonance imaging (MRI) faces a dual challenge: preterm birth produces a characteristic but confounding pattern of delayed myelination that can obscure concurrent structural abnormalities, while conditions such as hypoxic-ischaemic encephalopathy (HIE) and hydrocephalus demand timely identification in their own right. Existing frameworks typically address these clinical questions in isolation or collapse them into a single flat classifier, discarding the natural clinical decision hierarchy. This study presents a hierarchical two-stage deep learning pipeline that emulates clinical reasoning from paired T1-weighted and T2-weighted neonatal brain MRI. An initial Stage-1 deploys a brain-age gap classifier, built on a 3-D ResNet-18 regressor followed by a gap-thresholded binary decision intended to screen for myelination delay. Then the same subjects advance to a Stage-2, where a second 3-D ResNet-18, trained for structural pathology, performs three-class classification: normal, HIE , and hydrocephalus . Experiments demonstrate that a regression from MRI scan to age with an R 2 of 0.941 at Stage-1, the proposed hierarchical pipeline achieves a Stage-2 area under the curve (AUC) of 0.919 ± 0.024.