ResNet-34 with lightweight decoder achieves 97.37% accuracy on fetal brain MRI segmentation

Accurate segmentation of fetal brain tissues in MRI is critical for early congenital abnormality diagnosis but remains challenging due to fetal motion, low tissue contrast, and anatomical variability. To address this, researchers developed a novel deep learning model combining a ResNet-34 encoder with a lightweight decoder that uses multi-layer perceptron (MLP) modules for adaptive feature refinement. This design preserves anatomical boundaries and mitigates errors from motion artifacts and intensity inhomogeneities. Computational efficiency is achieved by reducing parameter count, using bilinear upsampling instead of transposed convolutions, and optimizing for speed without sacrificing accuracy.

Trained and validated on the FeTA 2021 dataset with 5-fold cross-validation, the model achieves an average accuracy of 97.37%, mean Dice Similarity Coefficient (DSC) of 90.33%, mean Intersection over Union (IoU) of 86.93%, and precision of 90.83%. It outperforms baselines such as UNet, UNet++, DeepLabV3, and DeepLabV3+ in segmenting white matter, gray matter, lateral ventricles, deep gray matter, extra-cerebrospinal fluid, cerebellum, and brainstem. Its fast inference time and reduced computational load make it well-suited for integration into real-time clinical workflows, offering a practical tool for improving prenatal care and early diagnosis.