Visual Enhancement and Semantic Segmentation of Murine Tissue Scans with Pulsed THz Spectroscopy

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Biomedical Engineering

Background:

• Deep learning for medical imaging faces challenges with low resolution and limited expertly-labeled datasets, particularly for emerging technologies like pulsed terahertz (THz) imaging.
• Pulsed THz imaging of excised breast tumors is hindered by image resolution limitations and domain shifts during histopathology, complicating traditional data-driven deep learning approaches.

Purpose of the Study:

• To apply Semantic Artificial Intelligence (AI) to segment tissue types in excised breast tumors using pulsed terahertz imaging.
• To address the limitations of low image resolution and lack of labeled data in THz imaging by leveraging semantic context.
• To develop a novel two-stage pipeline combining unsupervised and supervised learning for enhanced medical image analysis.

Main Methods:

• A two-stage pipeline was developed, beginning with an unsupervised image-to-image translation network.
• A supervised segmentation network was employed in the second stage of the pipeline.
• The approach utilizes synthetic THz scans generated by a bi-directional image-to-image translation network for training.

Main Results:

• The proposed Semantic AI pipeline enables enhanced near-real-time visualization of excised breast tumor tissues from THz scans.
• A supervised segmentation and classification training strategy was successfully implemented using only synthetically generated THz scans.
• The method effectively overcomes the limitations of low resolution and scarce labeled data inherent in traditional THz imaging deep learning.

Conclusions:

• Semantic AI offers a promising approach to augment deep learning in medical imaging, increasing information density and trust in results.
• The developed two-stage pipeline effectively segments tissue types in THz images of breast tumors, addressing key prior limitations.
• This work paves the way for improved diagnostic capabilities using emerging imaging technologies like THz, even with limited labeled data.