Combining 4D Hyperspectral Imaging With CNN for Nerve and Ligament Differentiation

Iatrogenic nerve injuries, inadvertently caused by an oper-ating physician, occur at a significant annual rate. A major reason for such intraoperative damage is poor nerve visual-ization, which results in the transection of incorrect structures or the misidentification of nerves as vessels or ligaments. Previously, diffuse reflectance-based hyperspectral imaging (HSI) has been applied to the nerve differentiation issue in the surgical context. In contrast, our study addresses the task of differentiating between nerves and ligaments using 3D and 4D autofluorescence HSI, followed by classification using CNN and linear unmixing. Our experiments use ex-cised bovine nerve and ligament samples. The data indicate that utilizing a 4D HSI configuration increases classification performance by up to 10%, achieving nearly 98.9% accuracy in most cases. Furthermore, CNNs consistently outperform linear unmixing, at times by more than 15%. We conclude that combining multiple autofluorescence HSI datacubes acquired at several narrow-band excitation wave-lengths with a CNN approach enables superior differentiation between nerve and ligament samples. Our code is available at https://github.com/nairamatosyanlnerve-lig-4D-hsi-cnn.