Exhibit Hall | Forum 2
Purpose: Investigate the viability of transferred incremental learning AI auto-segmentation models on institutionally specific data for post-prostatectomy patients, accounting for variations in anatomy differences and clinical tumor volume (CTV) definitions following institutional standards.
Methods: A total of 109 prostate-bed patients were utilized towards the vendor trained (VT) models for re-training and testing. Organs at risk (OAR) and prostate bed CTV AI-models (using 30 and 60 cases for training) were built and tested (49 remaining cases). Six GU radiation oncologists participated in blinded qualitative evaluation (AI-generated vs. manual contours) of the testing datasets, evaluating clinical preference and utility scores (4-precise, 3-acceptable, 2-minor edit, and 1-manual re-draw). 30 cases with physician consensus (PC) scoring (4 or more physicians scoring precise) were used to retrain a PC-trained model and added to the 60-case set for retraining. Quantitative evaluation was performed for all models and tested for statistical significance.
Results: The 30-case OAR model had median DICE values between 0.91-0.97, improving significantly over the VT-model for all OARs (p < 0.001), except for the penile bulb (DICE, 0.50). The 60-case CTV model has a median DICE of 0.7 improving over the 30-case model. No significant difference was observed in comparing the 90-case vs. 60-case model, or PC-trained model vs. 30-case model. For the blinded clinical utility scores, the AI and manual OAR delineations were chosen as acceptable or precise 87% and 94% of the time, respectively. While the AI prostate bed CTV delineation was scored precise or acceptable 51% of the time, as compared to the manual delineation value of 72%.
Conclusion: Site-specific incremental training of the institutional data can lead to significant improvement in the clinical utility and accuracy of AI-models. Small number of datasets are sufficient for incremental retraining to build institutionally specific models, making it accessible to clinics of all sizes.
Funding Support, Disclosures, and Conflict of Interest: X. Feng and Q. Chen are co-founders of CarinaMedical LLC. University of Kentucky receives an NIHSBIR subcontract from Carina Medical LLC. X. Feng, Q. Chen, and Y. Rong are partially supported by these grants. No conflict of interest for the other authors.