Two-stage deep learning pipeline detects neonatal brain pathologies from MRI

A two-stage deep learning framework that mimics the way clinicians reason about newborn brain scans has been developed to detect pathologies from magnetic resonance imaging. The study, published in Frontiers in Neuroimaging, addresses a persistent problem in automated neonatal neuroimaging: preterm birth produces a characteristic pattern of delayed myelination that can mask concurrent structural abnormalities, while conditions such as hypoxic-ischaemic encephalopathy and hydrocephalus require rapid identification in their own right.

The research team, based at the African Institute for Mathematical Sciences in Ghana and AGH University of Science and Technology in Poland, designed the pipeline to work from paired T1-weighted and T2-weighted neonatal brain MRI. Rather than collapsing all diagnostic questions into a single flat classifier, the framework preserves the natural clinical decision hierarchy, first screening for myelination delay and then classifying structural pathology.

Stage one of the pipeline deploys a brain-age gap classifier built on a three-dimensional ResNet-18 regressor. The network estimates brain age directly from the MRI scan, and the difference between estimated and chronological age is then compared against a threshold to produce a binary decision intended to screen for myelination delay. In experiments, the age regression achieved an R-squared value of 0.941, indicating that the network’s predictions tracked true age closely across the study data.

Subjects then advance to stage two, where a second three-dimensional ResNet-18, trained specifically for structural pathology, performs a three-class classification: normal, hypoxic-ischaemic encephalopathy, and hydrocephalus. Both demand timely identification, which is why the authors argue that existing frameworks that address these questions in isolation, or that flatten them into one classifier, discard clinically meaningful structure.

The hierarchical design achieved a stage-two area under the curve of 0.919 plus or minus 0.024 in the reported experiments. This performance metric reflects the pipeline’s ability to distinguish between the three structural categories once subjects have passed through the initial myelination screen. The authors note that the two-stage approach emulates clinical reasoning, in which a clinician first considers whether a preterm infant’s brain simply shows the expected immaturity of myelination before assessing whether an additional structural abnormality is present.

The work was received by the journal on 28 August 2026 and accepted on 6 October 2026. The authors, Christian Ahwoi Ampiaha and Tina Malalanirainya of the African Institute for Mathematical Sciences Ghana, together with Alessandro Crimi of AGH University of Science and Technology, describe the framework as a way to bring automated detection closer to the sequential logic that clinicians apply when reading neonatal scans.

The study’s keywords span three-dimensional ResNet architectures, brain-age gap estimation, comorbidity detection, deep learning, hierarchical classification, hydrocephalus, hypoxic-ischaemic encephalopathy and myelination delay, underscoring the breadth of clinical questions the single pipeline is intended to address. By screening for myelination delay first and then classifying structural pathology, the framework aims to reduce the confounding that arises when preterm immaturity obscures genuine abnormalities in a single-stage model.

Subject of Research: A hierarchical two-stage deep learning framework for detecting neonatal brain pathologies from T1-weighted and T2-weighted MRI.

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

Article References: Ampiah, C. A., Malalanirainy, T., & Crimi, A. (2026). Hierarchical detection of neonatal brain pathologies using a two-stage deep learning framework from magnetic resonance imaging. Frontiers in Neuroimaging, 5, Article 1981436. https://doi.org/10.3389/fnimg.2026.1981436

Image Credits: AI Generated

DOI: 10.3389/fnimg.2026.1981436

Keywords: neonatal brain MRI, deep learning, 3-D ResNet, brain-age gap, myelination delay, hypoxic-ischaemic encephalopathy, hydrocephalus, hierarchical classification, T1-weighted imaging, T2-weighted imaging, Hierarchical, detection