Abstract
Cranial nerve (CN) segmentation is a critical component of neurosurgical planning and neurological diagnosis. However, the task is complicated by the minute size of these structures and the frequent unavailability of complete multimodal imaging in clinical environments. This paper introduces CNTSeg-v2, a deep learning framework designed for robust, modality-agnostic segmentation. By utilizing T1-weighted (T1w) imaging as a primary anatomical guide to filter auxiliary diffusion MRI data, the model maintains high accuracy even when specific imaging modalities are missing. Evaluated on both research and clinical datasets, CNTSeg-v2 achieves state-of-the-art performance, demonstrating significant resilience to the data irregularities common in real-world medical settings.
Introduction
Cranial nerves are delicate, tube-like structures responsible for essential functions such as vision, hearing, and facial movement. For neurosurgeons, mapping these nerves is vital to avoid accidental injury during complex procedures. Despite their importance, segmenting CNs is notoriously difficult; they are often only a few voxels wide and reside within complex anatomical regions where contrast is low.
While multimodal imaging—combining structural MRI (like T1w) with diffusion MRI (such as Fractional Anisotropy or FOD Peaks)—improves visibility, it introduces a practical challenge: missing data. In clinical practice, patients rarely receive a full suite of diffusion scans due to time constraints, motion artifacts, or hardware limitations. Most existing AI models are rigid, requiring a fixed set of inputs to function. CNTSeg-v2 addresses this by providing a "modality-agnostic" architecture capable of handling any combination of available auxiliary scans without requiring retraining.
Methodology
CNTSeg-v2 utilizes a two-branch encoder-decoder architecture optimized for feature fusion and geometric consistency. The framework is built upon three innovative components:
1. Arbitrary-Modal Collaboration Module (ACM)
The ACM serves as the core of the fusion process. Recognizing that T1w images are the clinical standard and provide high anatomical contrast, the model treats T1w as the "primary" modality. The ACM uses T1w features to "score" and select the most relevant information from whatever auxiliary modalities (e.g., T2w, FA, DEC) are present. By employing a Bernoulli indicator, the module remains robust even when certain inputs are zeroed out or missing.
2. Dual-Attention Feature-Interactive Module (DFM)
Once features are selected, the DFM fuses them using a combination of spatial and channel attention. This ensures the network prioritizes the most informative data channels and focuses its "attention" on the specific spatial regions where the nerves are located, effectively filtering out background noise.
3. Deep Distance-guided Multi-stage (DDM) Decoder
A common failure in nerve segmentation is "fragmentation," where the model predicts broken or discontinuous segments. To solve this, the authors integrated Signed Distance Maps (SDM) into the decoder. Rather than performing simple binary classification (nerve vs. background), the model learns to predict the distance of each voxel to the nearest nerve boundary. This geometric supervision encourages the model to maintain structural continuity and recover fine details of the nerve tracts.
Experimental Setup
The model was trained and benchmarked using two distinct datasets:
- Human Connectome Project (HCP): 102 high-quality cases were used for training and rigorous performance benchmarking.
- Clinical Multi-shell Diffusion MRI (MDM) Dataset: 10 cases were used to validate the model's ability to generalize to real-world clinical data.
Because no "perfect" ground truth exists for these microscopic structures, reference labels were generated using advanced Unscented Kalman Filter (UKF) tractography, followed by meticulous expert filtering and Region of Interest (ROI) screening to ensure accuracy.
Results and Discussion
CNTSeg-v2 demonstrated superior performance across several metrics:
- State-of-the-Art Accuracy: On the HCP dataset, the model achieved a Mean Dice score of 73.16%, outperforming competitive methods such as MMFnet (71.94%) and the original CNTSeg (71.59%).
- Robustness to Missing Modalities: When compared against other missing-modality frameworks like RFNet and mmFormer, CNTSeg-v2 showed significantly higher stability. The model produced usable segmentations using only T1w images, with performance scaling upward as more auxiliary modalities were added.
- Clinical Generalization: The model maintained high performance on the MDM clinical dataset, proving it is not overfitted to the "clean" data found in research repositories.
- Computational Efficiency: The entire prediction pipeline takes less than 15 minutes, with the neural network inference itself requiring less than 50 seconds, making it feasible for near-real-time surgical planning.
Conclusion
CNTSeg-v2 represents a significant step forward in making automated cranial nerve segmentation practical for the clinic. By anchoring the fusion process in the T1w modality and utilizing geometric priors through Signed Distance Maps, the framework overcomes the dual hurdles of missing data and structural complexity. Future research will focus on further refining the model for low-quality, noisy clinical scans to ensure accessibility for a wider range of neurosurgical facilities.
Summary updated