Disclaimer: Unfold AI was previously known as iQuest
Alan Priester, PhD1,2; Joshua Shubert MS1; Sakina M. Mota, PhD1; Shyam Natarajan, PhD1,2
1Avenda Health, Inc; 2University of California, Los Angeles
Background: Accurate estimation of prostate cancer tumor size is critical for assessing risk and selecting treatment strategies. Since conventional clinical and imaging methods are weakly correlated with true tumor size, an artificial intelligence (AI) model was developed to map clinically significant prostate cancer (csPCa, i.e. Gleason grade group ≥ 2) risk in 3D. This study aimed to retrospectively compare AI mapping against conventional clinical markers for prediction of tumor volume on whole mount prostatectomy specimens.
Methods: Patients received magnetic resonance imaging (MRI) followed by radical prostatectomy. Excised prostate specimens were used to define ground-truth tumor volume by registering whole mount pathology slides to preoperative MRI and interpolating csPCa contours into 3D. Cases with at least one csPCa-bearing MRI-visible lesion and no prior ablative treatment were retrospectively selected for study inclusion. MRI and simulated biopsy information was then used to generate a 3D cancer estimation map (CEM) for each case using FDA-cleared software (iQuest, Avenda Health). Using the CEM, the sum of estimated csPCa probability for voxels throughout the prostate was correlated to true tumor volume using linear regression. The correlation of AI to tumor volume was compared to that of 6 conventional metrics derived from prostate serum antigen (PSA), biopsy results, and PI-RADS regions of interest (ROIs). The accuracies of linear regression fits were compared using Wilcoxon signed-rank tests with a Bonferroni-corrected significance threshold to account for multiple comparisons (α = 0.05/6 = 0.008).
Results: 97 patients met study eligibility criteria. AI cancer estimation maps were strongly correlated to true tumor volume with R2 = 0.81. Using a linear regression model to predict tumor volume, all conventional metrics were significantly less accurate than AI including: PSA (R2 = 0.27, p < 0.001), PSA density (R2 = 0.26, p < 0.001), number of csPCa-positive cores (R2 = 0.57, p < 0.001), maximum csPCa core length (R2 = 0.51, p < 0.001), total length of csPCa in all cores (R2 = 0.69, p = 0.005), and PI-RADS ROI volume (R2 = 0.33, p < 0.001).
Conclusion: An AI model was shown be highly predictive of true tumor volume measured on prostatectomy specimens, outperforming conventional clinical measures. More accurate tumor volume assessments may improve risk assessment and treatment strategy selection, particularly when defining margins for focal ablation or radiation dosing. This technique warrants further study in additional populations.
Source of Funding: This work was supported in part by the National Cancer Institute (R01CA218547) and by
Avenda Health, Inc.
