Author: Rex Carden, Carlos E. Cardenas, Ho-hsin Rita Chang, John B Fiveash, Heinzman A. Katherine, Yogesh Kumar, Gaurav Nitin Rathi, Richard A. Popple, Kayla Lewis Steed 👨🔬
Affiliation: University of Alabama at Birmingham 🌍
Purpose: Brain metastases (BMs) often require multiple radiotherapy (RT) courses as new lesions appear. Comparing follow-up imaging with prior RT plans is time-intensive. We developed an AI tool that aligns pre-treatment tumor contours with follow-up MRI to streamline post-RT interpretation. We evaluated its performance, feasibility, and efficiency in automating contour propagation to reduce clinician workload and improve post-treatment assessments.
Methods: Our approach automatically registers planning CTs to follow-up MRIs using segmented brain structures for accurate transformation. Clinical gross tumor volume (GTV) contours are rigidly overlaid on follow-up MRI and stored as DICOM for Picture Archiving and Communication System (PACS) compatibility. Using AAPM TG-132 guidelines, we performed a quantitative analysis comparing manually vs automatically registered contours in 20 BMs patients. We then propose a qualitative review in 20 patients (two cohorts of 10 each, with or without propagated GTVs) by six observers to confirm accuracy and utility.
Results: The automated workflow successfully registered simulation CT and follow-up MRI for all 20 patients. Comparing propagated contours from manual vs automated rigid registrations, average (±SD) DSC values were 0.97±0.01 for brain, 0.89±0.03 for brainstem, and 0.65±0.18 for GTVs; mean MSD values were 0.0±0.0 mm, 0.1±0.1 mm, and 0.6±0.4 mm, respectively. Preliminary data (one observer) indicated that follow-up MR review time per GTV decreased from 1.0 minute (manual) to 0.6 minute (automated).
Conclusion: We developed and preliminarily validated an AI-based solution to overlay RT-plan information on follow-up MRI in BMs patients, offering functionality not currently available in standard PACS systems. Quantitative metrics show good contour alignment and a 40% reduction in review time per GTV. A multi-observer qualitative study is in progress for final commissioning, and we plan to integrate additional RT data to broaden the tool’s clinical utility.