Author: Tommaso Frigerio, Joshua Genender, John M. Hoffman, Catherine (Caffi) Meyer 👨🔬
Affiliation: UCLA, David Geffen School of Medicine at UCLA 🌍
Purpose: Accurate bone marrow segmentation is required for bone marrow dosimetry to monitor for dangers in PSMA-Lu177 radioligand therapy. We introduce a hybrid (AI/semantic knowledge) segmentation pipeline to address complications such as bone calcifications and metastatic infiltrations present in the patient population.
Methods: Our algorithm relies on a combination of AI applied to CT, activity information derived from the SPECT data, and semantic methods. The integrated processing comprises TotalSegmentator (a deep neural network) for bone segmentation, a bone-by-bone dynamically adjusted thresholding, morphological operations to distinguish soft tissue from cortical bone, and analysis of activity information to exclude metastasis infiltrations. On a 20 patient test set from the RESIST trial of PSMA targeted therapy, we compared our method with two alternative SOTA methods: thresholding only and DNN only. These two methods, based on pre-existing algorithms, are only based on CT data. We visually compared the quality of segmentations and computed dice scores to quantify similarity and differences in 10 ROIs presenting high activity and 10 presenting low/no activity.
Results: Our model performed consistently with our baseline models in bones that showed no activity (0.722±0.184), while it performed differently in highly active bones (0.326±0.213), where we expected the baseline models to struggle due to the metastatic infiltration changing the usual anatomy. Preliminary evaluation showed the multi-modal algorithm is able to adapt to issues specific to the patient population such as the presence of bone calcification and metastatic infiltration of bone marrow in a robust way.
Conclusion: The multi-modal segmentation method showed improvements over conventional methods in theranostics patients requiring dosimetry. Future work will utilize this approach to test whole body dosimetry’s ability to predict blood toxicity.