Deep Learning Framework for Real-Time Fetal Brain Segmentation in MRI

Razieh Faghihpirayesh, Davood Karimi, Deniz Erdogmus, Ali Gholipour

Abstract

Fetal MRI is frequently compromised by unpredictable fetal movement, necessitating rapid slice acquisition and real-time processing to guide clinical imaging. We present RFBSNet (Real-time Fetal Brain Segmentation Network), a specialized deep learning architecture optimized for low-latency, "batch size 1" inference. RFBSNet achieves a state-of-the-art inference time of 3.36 ms per slice while maintaining high segmentation accuracy (97.99% Dice score). By prioritizing memory access efficiency over raw FLOP reduction, RFBSNet enables the real-time motion tracking required for automated fetal MRI navigation.

Introduction

Fetal Magnetic Resonance Imaging (MRI) provides critical diagnostic information regarding prenatal development. However, the fetus acts as a "moving target," often leading to motion artifacts that obscure anatomical details. Current clinical practice relies on fast 2D slice acquisition followed by offline 3D reconstruction. To improve this workflow, there is a significant need for "closed-loop" systems that can track the fetal head in real-time and adjust the scanner's orientation accordingly.

Such systems require instantaneous segmentation of the fetal brain from every incoming 2D slice. While numerous real-time segmentation networks exist in computer vision, they are typically optimized for high-throughput scenarios (processing large batches of images). In a clinical MRI setting, the priority is low latency—processing a single image the moment it is acquired. This paper introduces RFBSNet, designed specifically to bridge the gap between high accuracy and the low-latency requirements of real-time clinical applications.

Methodology

The RFBSNet Architecture

RFBSNet is designed to maximize computational efficiency without sacrificing the spatial resolution necessary for precise boundary detection. The architecture is built upon three primary components:

  1. Input Downsampling Module: Standard convolutional layers often incur high computational costs at the initial high-resolution stages. RFBSNet utilizes a parallel downsampling approach, combining a $3 \times 3$ convolution with a $2 \times 2$ max-pooling operation. By concatenating these outputs, the network preserves fine spatial details while rapidly reducing the feature map size.
  2. Two-Branch Structure: The network employs a dual-pathway design. A shallow branch maintains high-resolution spatial information, while a deeper, lightweight branch extracts global semantic context. This ensures the model understands both the "where" (location) and the "what" (brain tissue) of the input.
  3. U-Net Integration: RFBSNet adopts an encoder-decoder structure with skip connections. This allows for the fusion of high-level semantic features with low-level spatial details, which is essential for accurate segmentation at the brain boundaries.

Optimization for Batch Size 1

A key design choice in RFBSNet is the avoidance of depthwise-separable convolutions. Although these are popular in mobile-oriented networks for reducing floating-point operations (FLOPs), they often suffer from high memory access latency on GPUs when processing a single image (batch size 1). RFBSNet prioritizes parallelism and memory efficiency to ensure true real-time performance in a live scanning environment.

Experimental Setup

The model was trained and evaluated using a dataset of 3,496 2D fetal MRI slices acquired on 3T Siemens scanners using HASTE sequences. The data was partitioned into:

  • Training/Validation (80/20%): Used for model development.
  • Normal Test Set: 840 slices with minimal motion artifacts.
  • Challenging Test Set: 136 slices containing significant noise and motion artifacts to test the model's robustness.

Ground truth masks were manually generated by an experienced annotator. Performance was measured using the Dice similarity coefficient and inference latency on an NVIDIA GeForce RTX 2080 Ti.

Results

RFBSNet demonstrated superior performance compared to eight state-of-the-art models, including U-Net, ENet, and Fast-SCNN.

Speed and Efficiency

RFBSNet achieved an inference time of 3.36 ms per slice. This is significantly faster than the standard U-Net (9.82 ms) and even specialized real-time models like Fast-SCNN (5.74 ms). Furthermore, RFBSNet is highly parameter-efficient, utilizing approximately 14 times fewer parameters and FLOPs than a standard U-Net.

Segmentation Accuracy

On the normal test set, RFBSNet achieved a 97.99% Dice score, matching the accuracy of much larger, non-real-time models. On the challenging test set, the model maintained a Dice score of 84.04%. While this represents a decrease under extreme conditions, RFBSNet remained significantly more robust than other real-time architectures, which often failed to localize the brain entirely in the presence of heavy artifacts.

Discussion and Conclusion

The development of RFBSNet addresses a critical bottleneck in fetal MRI. By optimizing for the "batch size 1 paradox," the researchers have created a tool that provides the low latency required for real-time scanner guidance.

While performance on highly artifact-prone images remains a challenge, the speed and general accuracy of RFBSNet make it a viable candidate for integration into automated MRI pipelines. Future work will focus on using these real-time segmentations to drive automated slice re-acquisition and 3D super-resolution reconstruction, potentially transforming fetal MRI into a more reliable and streamlined diagnostic process.

Summary updated

BibTeX
@inproceedings{faghihpirayesh2022deep,
  author = {Razieh Faghihpirayesh and Davood Karimi and Deniz Erdogmus and Ali Gholipour},
  title = {{Deep Learning Framework for Real-Time Fetal Brain Segmentation in MRI}},
  booktitle = {PIPPI @ MICCAI},
  year = {2022},
  doi = {10.1007/978-3-031-17117-8\_6},
}
Figure