Publication

Abstract : Neonatal Hypoxic-Ischemic Encephalopathy (HIE) is a major brain injury in newborns resulting from insufficient supply of oxygen during the perinatal period, often leading to long-term neurological dysfunctions and delayed development . Timely and accurate diagnosis is essential for enhancing clinical outcomes. In this work, we propose a multimodal approach for automatically segmenting HIE lesions and severity classification using MRI scans. A SegResNet architecture was employed to segment lesions from Apparent Diffusion Coefficient (ADC) and Z-scored ADC (Z-ADC) maps. In the following stage, features derived from predicted lesion masks and clinical meta data were combined using an early fusion approach. The fused feature set was used for classifying the HIE lesions according to severity as mild, moderate and severe. Experiments were conducted on the Boston Neonatal Brain Injury Dataset (BONBID-HIE). The segmentation model achieved a Dice score of 0.749, while the severity classification obtained an accuracy of 88.64%. These findings show the effectiveness of the proposed structure in both clinically relevant severity classification and accurate lesion spot identification, indicating its ability to support early clinical decision-making in neonatal clinical care..