Segmentation Under Stress: Retraining on Noisy Data Improves Nerve Detection Generalization

Aidan Schurr, Chung Hyuk Park, Richard J. Cha, Gary Milam Jr.

Keywords: Nerve Tissue Identification, Birefringence Imaging, Transformer-based U-Net, Convolutional Neural Network (CNN), Data Enhancement

Abstract: The precise identification of nerve tissue is an essential requirement to inhibit detrimental outcomes during surgical procedures, as nerve damage has long-term adverse effects on patients. Through Birefringence imaging and a Dual U-Net architecture with a transformer-based backbone, an imaging tool was developed that could provide increased effectiveness for noninvasive intraoperative nerve identification. However, these findings were based on a training set that did not accurately represent real-time surgical procedures. In this study, we apply holistic adaptations to the DXM-TransFuse U-net and Birefringence Imaging procedures to handle clinically significant data more accurately. These enhancements are the next step towards applying a practical and significant application of this technology. Through collecting and training on more realistic data, such as nerve tissues being partially blocked by surgical tools, the neural network was able to detect nerve tissues more accurately. Using and extracting features from these noisier data, the model was able to increase the effectiveness for simpler data processing as well. In general, these advances helped develop a more robust model that could handle larger and more diverse images and represents a more clinically accurate model that aims to present an effective solution for future nerve identification.

Submission Number: 61