Fusion of Deep and Local Features Using Random Forests for Neonatal HIE Segmentation.

Hypoxic ischemic encephalopathy (HIE) is a prevalent brain injury in neonates, and current segmentation methods have shown limited efficacy. We present a cutting-edge 3D HIE brain image segmentation technique that combines Swin-UNETR with a random forest classifier. While Swin-UNETR, with its large parameter space, tends to overfit, the integration of the random forest classifier significantly mitigates the overfitting problem. Our deep network processes 3D HIE images using two channels (ADC and zADC) to generate a lesion probability map. This map, alongside the input channels, is segmented into 5×5 2D windows and input into the random forest for lesion probability prediction. We propose a unique log Hausdorff distance loss to regularize 3D anatomical shape regions of hypoxia and implicitly optimize surface distance metrics (MASD and NSD). In the BONBID-HIE challenge, our approach surpassed state-of-the-art methods across all metrics, offering a more scalable approach for the translational use of HIE lesion segmentation.