Comparing and Combining Artificial Intelligence and Spectral/Statistical Approaches for Elevating Prostate Cancer Assessment in a Biparametric MRI: A Pilot Study

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Background: Prostate cancer management optimally requires non-invasive, objective, quantitative, accurate evaluation of prostate tumors. The current research applies visual inspection and quantitative approaches, such as artificial intelligence (AI) based on deep learning (DL), to evaluate MRI. Recently, a different spectral/statistical approach has been used to successfully evaluate spatially registered biparametric MRIs for prostate cancer. This study aimed to further assess and improve the spectral/statistical approach through benchmarking and combination with AI. Methods: A zonal-aware self-supervised mesh network (Z-SSMNet) was applied to the same 42-patient cohort from previous spectral/statistical studies. Using the probability of clinical significance of prostate cancer (PCsPCa) and a detection map, the affiliated tumor volume, eccentricity was computed for each patient. Linear and logistic regression were applied to the International Society of Urological Pathology (ISUP) grade and PCsPCa, respectively. The R, p-value, and area under the curve (AUROC) from the Z-SSMNet output were computed. The Z-SSMNet output was combined with the spectral/statistical output for multiple-variate regression. Results: The R (p-value)–AUROC [95% confidence interval] from the Z-SSMNet algorithm relating ISUP to PCsPCa is 0.298 (0.06), 0.50 [0.08–1.0]; relating it to the average blob volume, it is 0.51 (0.0005), 0.37 [0.0–0.91]; relating it to total tumor volume, it is 0.36 (0.02), 0.50 [0.0–1.0]. The R (p-value)–AUROC computations showed a much poorer correlation for eccentricity derived from the Z-SSMNet detection map. Overall, DL/AI showed poorer performance relative to the spectral/statistical approaches from previous studies. Multi-variable regression fitted AI average blob size and SCR results at a level of R = 0.70 (0.000003), significantly higher than the results for the univariate regression fits for AI and spectral/statistical approaches alone. Conclusions: The spectral/statistical approaches performed well relative to Z-SSMNet. Combining Z-SSMNet with spectral/statistical approaches significantly enhanced tumor grade prediction, possibly providing an alternative to current prostate tumor assessment.

This publication has 61 references indexed in Scilit: