ResNet-34 with Lightweight Decoder for Accurate and Efficient Segmentation of Fetal Brain MRI

Researchers have developed a ResNet-34-based deep learning model with a lightweight decoder for segmenting fetal brain tissues in MRI scans, achieving 97.37% accuracy and 90.33% mean Dice Similarity Coefficient. The model addresses critical challenges in prenatal diagnosis by handling fetal motion artifacts and anatomical variability while maintaining computational efficiency suitable for real-time clinical use.

This research addresses a genuine clinical need in prenatal care by advancing automated segmentation of fetal brain MRI scans. Accurate tissue identification in fetal brains is essential for early detection of congenital abnormalities, yet remains technically challenging due to fetal motion, low tissue contrast, and significant anatomical variations across gestational stages. The proposed architecture combines a ResNet-34 encoder with a lightweight decoder using multi-layer perceptron modules, achieving superior performance metrics compared to established baselines like UNet and DeepLabV3+.

The technical approach represents an incremental but meaningful advancement in medical image segmentation. Rather than pursuing maximum model complexity, the researchers prioritized efficiency through parameter reduction and bilinear upsampling, demonstrating that computational optimization need not compromise accuracy. This design philosophy aligns with growing industry recognition that clinical AI tools must balance performance with practical deployment constraints in hospital environments.

The clinical impact centers on enabling faster, more consistent prenatal diagnoses. Real-time inference capability allows integration into existing clinical workflows without requiring specialized infrastructure, potentially improving access to advanced diagnostic tools across varying healthcare settings. The 90.33% Dice coefficient and 86.93% IoU metrics represent solid performance for complex anatomical segmentation tasks.

Future development opportunities include validation on diverse ethnic populations and gestational ages beyond the FeTA 2021 dataset, assessment of real-world clinical performance against radiologist assessments, and potential extension to 3D volumetric analysis. The work contributes to the broader trend of specialized neural architectures designed for specific medical imaging tasks rather than universal off-the-shelf models.

• →ResNet-34 with lightweight MLP-based decoder achieves 97.37% accuracy on fetal brain MRI segmentation, outperforming UNet, UNet++, and DeepLabV3+ baselines.
• →Model design prioritizes computational efficiency through reduced parameters and bilinear upsampling, enabling real-time clinical deployment.
• →The architecture successfully handles technical challenges including fetal motion artifacts, low tissue contrast, and anatomical variability across gestational ages.
• →Fast inference time and reduced computational load make the model suitable for integration into existing prenatal care workflows.
• →5-fold cross-validation on FeTA 2021 dataset demonstrates robust generalization across diverse fetal brain anatomy cases.