Semantic Segmentation of Xenograft Tumor Tissues Imaged with Pulsed Terahertz Technology

Pulsed terahertz (THz) spectroscopy has achieved promising results in differentiating cancerous tissue from healthy adipose tissue. However, quantitatively evaluating THz spectroscopy of freshly excised tissue and its correlation to pathology presents a significant challenge. The processes involved in obtaining a pathology image cause significant deformation of the tissue. Therefore, pathology images are largely decoupled from their corresponding pulsed THz images of freshly excised tissue samples. Lacking highly correlated pathology analyzed images brings challenges in generating reliable training labels for supervised segmentation models and evaluation between segmentation and classification methods. This work proposes a novel THz image process pipeline and adopts an unsupervised image-to-image domain adaption method to generate synthetic THz scans directly from pathology images. With the generated synthetic THz scans, we train a robust segmentation neural network that effectively classifies tissue types among real-world THz scans at the pixel level.