Spectrally Tunable Neural Network-Assisted Segmentation of Microneurosurgical Anatomy

Distinct tissue types are differentiated based on the surgeon's knowledge and subjective visible information, typically assisted with white-light intraoperative imaging systems. Narrow-band imaging (NBI) assists in tissue identification and enables automated classifiers, but many anatomical details moderate computational predictions and cause bias. In particular, tissues' light-source-dependent optical characteristics, anatomical location, and potentially hazardous microstructural changes such as peeling have been overlooked in previous literature.

Narrow-band images of five (n = 5) facial nerves (FNs) and internal carotid arteries (ICAs) were captured from freshly frozen temporal bones. The FNs were split into intracranial and intratemporal samples, and ICAs' adventitia was peeled from the distal end. Three-dimensional (3D) spectral data were captured by a custom-built liquid crystal tunable filter (LCTF) spectral imaging (SI) system. We investigated the normal variance between the samples and utilized descriptive and machine learning analysis on the image stack hypercubes.

Reflectance between intact and peeled arteries in lower-wavelength domains between 400 and 576 nm was significantly different (p < 0.05). Proximal FN could be differentiated from distal FN in a higher range, 490-720 nm (p < 0.001). ICA with intact tunica differed from proximal FN nearly thorough the VIS range, 412-592 nm (p < 0.001) and 664-720 nm (p < 0.05) as did its distal counterpart, 422-720 nm (p < 0.001). The availed U-Net algorithm classified 90.93% of the pixels correctly in comparison to tissue margins delineated by a specialist.

Selective NBI represents a promising method for assisting tissue identification and computational segmentation of surgical microanatomy. Further multidisciplinary research is required for its clinical applications and intraoperative integration.