Development of an Automated Tumor Spectral Artifact Removal Algorithm Using Active Learning

Motivation: Spectral artifact filtration is critical in creating accurate metabolite maps which guide high-dose radiation therapy for patients with glioblastoma. Goal(s): This project seeks to generate accurate metabolite maps for radiation therapy planning by accurately filtering spectra immediately surrounding the tumor volume. Approach: By using deep learning to automatically segment tumor infiltration, we are able to focus our model on spectra used for radiation therapy planning. Employing active learning enables the model identify uncertain spectra for targeted expert labeling. Results: We have successfully created an active learning framework and labeling platform for experts to efficiently label spectra from 89 whole-brain spectroscopy scans. Impact: Our active learning-based model streamlines expert annotation for artifact filtration in tumor spectra, optimizing radiation therapy planning in glioblastoma. This approach accelerates data labeling, paving the way toward improved spectral quality assessment for clinical spectroscopic MRI applications.