YOLOv8-Seg for Multi-Tissue Segmentation of Fetal Brain MRI: A FeTA Benchmark and Comparative Study with U-Net...

Accurate segmentation of fetal brain tissues from Magnetic Resonance Imaging (MRI) is essential for quantitative assessment of neurodevelopment during gestation. The Fetal Tissue Annotation Challenge (FeTA) has established multi-class segmentation of seven brain tissues as a benchmark task. State-of-the-art approaches, such as U-Net and its variants, achieve high accuracy but typically require volumetric reconstruction, extensive preprocessing, and considerable computational resources, limiting their applicability for real-time or embedded scenarios. In contrast, You Only Look Once YOLOv8-Seg is a recent, lightweight instance segmentation model that offers high inference speed and minimal preprocessing requirements. To the best of our knowledge, no previous study has systematically evaluated YOLOv8-Seg for multi-tissue segmentation of fetal brain MRI. In this work, we present the first adaptation and benchmark of YOLOv8-Seg on the FeTA dataset, comparing its performance against two-Dimensional (2D) and three-Dimensional (3D) Convolutional Networks for Biomedical Image Segmentation (U-Net) baselines. Our evaluation covers both segmentation accuracy and computational efficiency. We further investigate the impact of task complexity by highlighting class-specific challenges of multi-tissue segmentation performance, such as tissue boundary ambiguities. Results show that YOLOv8-Seg, though designed for natural image analysis, achieves competitive accuracy primarily on large tissue structures, while exhibiting performance trade-offs on smaller or less well-defined tissues, and significantly outperforming U-Net-based methods in inference speed. Specifically, YOLOv8-Seg attains a mean Dice score of 0.7479 and an average HD95 of 2.3 mm across seven tissue classes on the FeTA benchmark, while enabling near real-time inference at 0.12 s per volume with only 2.8 GB of Graphical Processing Unit memory usage.