Supervised Semantic Segmentation of Murine THz Spectroscopy Images with Imprecise Annotations.

Semantic Artificial Intelligence possesses attributes that are particularly beneficial for deep learning tasks in medical imaging. By infusing semantic context into the fundamental classification process, the richness of data in medical images can be enhanced, leading to a potential increase in the reliability of the outcomes. In this research, we explore the use of semantic AI for distinguishing different tissue types within breast tumors that have been excised and imaged using pulsed terahertz (THz) technology, which is a cutting-edge method in the imaging field. Prior work has demonstrated traditional data driven methodology for deep learning on THz images has been challenging due to lacking of precise pixel-to-pixel labels, which is caused by domain transformation and tissue changes during histopathology. This work seeks to address this limitation through a noisy label handling enabled semantic AI mechanism. Specifically, we introduce a three stage training pipeline using supervised deep learning networks, contrastive feature reduction, and metric learning. The combination of these contributions enables better cancerous tissue differentiation. The recall for cancer tissue regions is significantly improved to 90 % from 55 % using the proposed semantic segmentation network.