Tuesday, December 1 | 3:00 p.m.-4:00 p.m. | T7-SSRO02-7 | Room S401CD
For patients with metastatic prostate cancer, deep learning applied to PSMA-PET scans may be better for predicting survival than existing risk tools, according to this presentation.
Presenter Daniel Fu of Yale School of Medicine and colleagues set out to improve risk stratification in metastatic prostate cancer by developing and externally validating deep learning-derived imaging biomarkers from prostate-specific membrane antigen (PSMA) PET scans.
The group retrospectively identified 486 PSMA PET exams from patients diagnosed between 2017 and 2025. These studies were used to train an imaging model and a “fusion” model that combines imaging with clinical variables. Studies from the host institution were used to train an imaging model and a “fusion” model that combines imaging with clinical variables. On an independent test set, the researchers compared both models against six established risk tools, including the VISION nomogram and the CHAARTED and LATITUDE classifications, for predicting two-year overall survival.
In the external test set, the fusion model performed best, with an area under the curve (AUC) of 0.86, followed by the imaging model at 0.82. The comparator tools ranged from 0.43 to 0.65. The fusion model also separated patients into clear risk groups, with two-year survival of 93% versus 43%. Performance held steady across radiotracers, treatment history, and metastatic burden.
“Once validated prospectively, these biomarkers may enable earlier and more personalized treatment selection across the full metastatic disease continuum,” the researchers wrote.
Learn more by attending this Tuesday afternoon session.
