Multimodal MRI segmentation of key structures for microvascular decompression via knowledge-driven mutual distillation and topological constraints

Abstract

Purpose

Microvascular decompression (MVD) is a widely used neurosurgical intervention for the treatment of cranial nerves compression. Segmentation of MVD-related structures, including the brainstem, nerves, arteries, and veins, is critical for preoperative planning and intraoperative decision-making. Automatically segmenting structures related to MVD is still challenging for current methods due to the limited information from a single modality and the complex topology of vessels and nerves.

Methods

Considering that it is hard to distinguish MVD-related structures, especially for nerve and vessels with similar topology, we design a multimodal segmentation network with a shared encoder–dual decoder structure and propose a clinical knowledge-driven distillation scheme, allowing reliable knowledge transferred from each decoder to the other. Besides, we introduce a class-wise contrastive module to learn the discriminative representations by maximizing the distance among classes across modalities. Then, a projected topological loss based on persistent homology is proposed to constrain topological continuity.

Results

We evaluate the performance of our method on in-house dataset consisting of 100 paired HR-T2WI and 3D TOF-MRA volumes. Experiments indicate that our model outperforms the SOTA in DSC by 1.9% for artery, 3.3% for vein and 0.5% for nerve. Visualization results show our method attains improved continuity and less breakage, which is also consistent with intraoperative images.

Conclusion

Our method can comprehensively extract the distinct features from multimodal data to segment the MVD-related key structures and preserve the topological continuity, allowing surgeons precisely perceiving the patient-specific target anatomy and substantially reducing the workload of surgeons in the preoperative planning stage. Our resources will be publicly available at https://github.com/JaronTu/Multimodal_MVD_Seg.

Appendix A Appendix

A.1 Data collection

All scans were acquired from one hospital and annotated by a doctor with five years of experience, followed by double-checking the accuracy and consistency by a doctor with eight years of experience. The dataset was collected from June 2020 to June 2022 at the largest center that performs MVD in our region. The annotation of a case involves delineating approximately 150 slices for two modalities from one patient, which takes over 1 hour for a clinician with five years of experience. Surgeries before June 2020 used 1.5T TOF-MRA instead of 3T TOF-MRA. In clinic, 3T TOF-MRA combined with HR-T2WI sequence can show a clearer relationship between cranial nerves and vessels. Consequently, we have tried our best to collect our dataset in a single-center study.

A.2 Discussion

The PH in 3.3 is calculated based on patches; however, since our target structures are thin tubular structures, the scale difference between target size and patch size ensures that the local connectivity within each patch is maintained. By evaluating the connectivity within each patch, we can reconstruct the global topology from localized observations, effectively compensating for any potential disruptions caused by the patch boundaries.

In Table. 2, we observe that $L_{IMD}$ and $L_{topo}$ cause a slight decrease in Dice of brainstem; we speculate that the brainstem, being a relatively large and stable structure, may not benefit much from the proposed loss functions that aim to enhance the segmentation of thin and curvilinear structures such as arteries, veins, and nerves. Besides, the proposed losses may introduce some noise or interference to the brainstem segmentation, especially when the brainstem is adjacent or overlapped with other structures.