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

Abstract:

Precise segmentation of fetal brain tissues in Magnetic Resonance Imaging (MRI) is vital for enhancing prenatal care and facilitating the early detection of congenital abnormalities. Nevertheless, this task is fraught with challenges, including significant anatomical variations across gestational stages, low tissue contrast, and artifacts stemming from fetal movement. These factors complicate the delineation of intricate structures, such as the brainstem, cerebellum, deep gray matter, extra-cerebrospinal fluid, lateral ventricles, white matter, and gray matter.

To address these obstacles, this study presents a novel deep learning architecture that integrates a ResNet-34 encoder with a streamlined decoder. The decoder utilizes multi-layer perceptron (MLP) modules to perform adaptive feature refinement, a design choice that specifically bolsters the model’s capacity to maintain anatomical boundaries and reduce segmentation errors associated with intensity inhomogeneities and motion artifacts.

The proposed model prioritizes computational efficiency by minimizing the parameter count and substituting transposed convolutions with bilinear upsampling. This optimization accelerates the decoder’s speed while maintaining high accuracy. Evaluated on the FeTA 2021 dataset through 5-fold cross-validation, the model surpasses baseline architectures, including UNet++, UNet, DeepLabV3+, and DeepLabV3. It achieved an average Accuracy of 97.37%, a mean Intersection over Union (IoU) of 86.93%, a Precision of 90.83%, and a mean Dice Similarity Coefficient (DSC) of 90.33%. Furthermore, its low computational demand and rapid inference times position it as an ideal candidate for integration into real-time clinical workflows.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC