338 - Interpretable Deep Learning Predicts Metastasis-Free Survival (MFS) from Conventional Imaging for Oligometastatic Castration-Sensitive Prostate Cancer (omCSPC) Using Multi-Modality PSMA PET and CT Imaging
Presenter(s)
Y. Cao1, P. Sutera2, W. Mendes3, A. Sawant4, L. Marchionni5, N. L. Simone6, P. T. Tran7, H. C. Onal8, and L. Ren7; 1Department of Radiation Oncology, University of Maryland School of Medicine, Baltimore, MD, 2University of Rochester, Rochester, NY, 3University of Maryland School of Medicine, Baltimore, MD, USA, Baltimore, MD, 4University of Maryland, School of Medicine, Baltimore, MD, 5Department of Pathology and Laboratory Medicine, NewYork-Presbyterian/Weill Cornell Medical Center, New York, NY, 6Department of Radiation Oncology, Sidney Kimmel Medical College at Thomas Jefferson University, Philadelphia, PA, 7University of Maryland School of Medicine, Baltimore, MD, 8Department of Radiation Oncology, Baskent University Faculty of Medicine, Adana, Turkey
Purpose/Objective(s):
This study aims to predict 2-yr Metastasis-free survival (MFS) for oligometastatic castration-sensitive prostate cancer (omCSPC) patients treated by metastasis-directed therapy (MDT) by developing a novel auto-classification network using pre-treatment multi-modality imaging. This prediction will allow for treatment customization based on the prediction to improve outcomes.Materials/Methods:
This multi-institutional study involved 118 omCSPC patients treated by MDT, comprising 34 patients from an institution in US and 84 from another institution in Europe. A novel interpretable 3D convolutional-neural-network (CNN) architecture was designed with two encoding paths for CT and PET images and one encoding path for clinical parameters. Each path used a 3×3×3 kernel with the same filters to independently weigh the spatial features of each image. The Class Activation Score (CAS) was defined to reflect how much the input path contributes to the final prediction for a specific category. The Total Class Activation Score (TCAS) measures the overall strength of class activation scores across the input path. The relative percentage weights (W%) indicate the proportional contribution of each input path to the model. Five clinical parameters—Age, Gleason Score, Number of Total Lesions, Untreated Lesions, and pre-MDT Prostate-specific Antigen (PSA)—were included as inputs for the models. The model was trained to predict 2-yr MFS with the actual patient outcome as the ground-truth. The model's performance was evaluated through 10-fold cross-validation.Results:
Among the 93 patients selected from 118, 44 (47%) were confirmed to have experienced distant metastases within the two-year timeframe, while 49(53%) were confirmed without metastases for the same duration. The AI model predicted correctly for 77 (83%) patients, including 34 patients with MFS and 43 patients without MFS. Additionally, clinical parameters improved the overall prediction accuracy by 11%. The weighting for each input in the final model provides insights into its importance in the prediction. The weights for PSMA-CT and PET ranged between 41-46% and 43-51%, respectively, which are significantly higher compared to those for individual clinical parameters (6-15%). For clinical parameters, the relative percentage weights showed a slight decline from MFS (with) to MFS (without) cases, likely due to the increased emphasis on clinical parameter inputs in MFS (with) cases.Conclusion:
A novel fusion CNN model with three encoding paths was successfully developed for predicting metastasis-free survival (MFS). Our study highlighted the potential of using multi-modality imaging biomarkers (CT and PET) for 2-year MFS prediction in patients with omCSPC. This finding presents a unique opportunity for targeted treatment interventions to improve outcomes for patients identified as having a poor prognosis.