Introduction
Segmenting the fetal brain from MRI is a significant challenge in prenatal care. Constant fetal motion during scans results in motion-corrupted 2D slices with low 3D resolution. While deep learning models like Convolutional Neural Networks (CNNs) have automated much of this process, they often struggle with challenging slices containing artifacts or pathologies. Traditionally, clinicians must manually review and refine these segmentations slice-by-slice, a process that is both tedious and time-consuming. This paper introduces the Uncertainty-Guided Interactive Refinement (UGIR) framework, designed to streamline this workflow by using AI confidence to guide human intervention.
The UGIR Framework
The UGIR framework optimizes the human-in-the-loop process by identifying the most problematic slices for the user to refine. Instead of a linear search through a stack of MRI slices, the system directs the user to slices with the highest estimated uncertainty.
As shown in the workflow above, the process begins with MG-Net, which performs initial segmentation and real-time uncertainty estimation. The system then fetches slices ($S_m$) based on their uncertainty levels. The user interacts with these specific slices to provide corrections, which are then processed by an interactive level-set method to refine the final segmentation.
Methods
MG-Net: Real-Time Uncertainty Estimation
Standard methods for uncertainty estimation, such as Monte Carlo Dropout or Model Ensembles, require multiple forward passes, making them too slow for real-time interactive applications. To solve this, the authors developed MG-Net (Multiple Groups Network). By utilizing Grouped Convolutions, MG-Net splits feature maps into $N$ independent groups, allowing it to generate multiple predictions in a single forward pass. The variance between these predictions provides a robust, real-time "uncertainty map."
Normalized Slice-Level Uncertainty
To ensure the system correctly prioritizes slices, the authors implemented a normalized slice-level uncertainty metric. This normalization prevents small errors in large slices from overshadowing significant errors in smaller slices, ensuring that "trouble" areas are identified regardless of the brain's volume in a particular view.
I-DRLSE: Interactive Refinement
Once a problematic slice is identified, the user provides simple interactions (like scribbles or clicks). These are incorporated into the Interactive Distance Regularized Level Set Evolution (I-DRLSE). Unlike traditional level-set methods that rely on raw pixel intensity—which can be noisy in fetal MRI—I-DRLSE utilizes the CNN’s probability maps. This makes the refinement more robust to image artifacts and ensures the boundaries respect the anatomical structure.
Experimental Results
The framework was evaluated on MRI scans from 35 fetuses (105 stacks total). Key findings include:
- Efficiency: MG-Net estimated uncertainty in just 0.29s per stack, compared to 3.12s for MC Dropout. Overall, the UGIR framework improved human refinement efficiency by approximately 30%.
- Accuracy: The final refined segmentations achieved a high Dice score of 94.83%.
As illustrated in the comparison above, the I-DRLSE method provides superior qualitative results and tighter statistical performance (lower ASSD and higher Dice scores) compared to traditional methods like Graph Cut or R-Net.
Conclusion
The UGIR framework demonstrates that AI-generated uncertainty can act as an effective "compass" for human experts, significantly reducing the time required for high-quality medical image segmentation. While the current I-DRLSE step is CPU-based, future optimizations for GPU could further enhance the responsiveness of the interactive system. This philosophy of uncertainty-guided refinement holds promise for a wide range of complex medical imaging tasks.
Summary updated