For the detection of clinically significant prostate cancer (csPCa) on magnetic resonance imaging (MRI), the use of artificial intelligence (AI) offers a higher area under the receiver operating characteristic curve (AUROC) and significantly lower false positives than radiologist assessment, according to emerging research.
For the study, recently published in Lancet Oncology, researchers trained the AI system on 9,207 MRI exams from three different facilities and tested the system on 1,000 MRI exams from four different medical centers.1 The study authors also performed a parallel observational study, using 400 MRI exams from the aforementioned testing cohort, to compare the AI model to the performance of 62 radiologists in interpreting prostate MRIs with the PI-RADS 2.1 classification system. According to the study, the radiologists had a median of seven years of prostate MRI experience and were drawn from 45 centers in 20 countries.
The researchers found the AI system had a 91 percent AUROC in comparison to an 86 percent pooled AUROC for reviewing radiologists. In contrast to radiologists for the 400-case subset, the AI model had 50.4 percent fewer false positive results at the same specificity and facilitated a 20 percent reduction in indolent cancer detection at the same sensitivity.1
“The (research) showed that a state-of-the-art AI system was superior in discriminating patients with clinically significant prostate cancer at biparametric MRI compared with the mean of 62 radiologists using PI-RADS (2.1) within an international reader study,” wrote lead author Anindo Saha, MSc, PhD(c), who is affiliated with the Diagnostic Image Analysis Group and the Minimally Invasive Image-Guided Intervention Center at the Radboud University Medical Center in Nijmegen, Netherlands, and colleagues.
In the aforementioned 400-case subset, the study authors found the AL model had a superior positive predictive value (PPV) (68 percent vs. 53.2 percent) and a higher negative predictive value (NPV) (93.8 percent vs. 90.2 percent) in comparison to radiologists. However, when utilizing a PI-RADS 3 or greater operating point in the entire 1,000-case testing cohort, the researchers noted that radiologists had comparable PPVs (60.6 percent vs. 60.5 percent) and NPVs (97.3 percent vs. 97.3 percent) to the AI model.1
Three Key Takeaways
1. Higher diagnostic accuracy. The AI system demonstrated a higher area under the receiver operating characteristic curve (AUROC) of 91 percent, compared to the pooled AUROC of 86 percent for the 62 radiologists. This indicates that the AI system is more accurate in distinguishing between patients with and without csPCa.
2. Reduction in false positives. The AI model achieved 50.4 percent fewer false positive results at the same specificity compared to radiologists. This reduction in false positives can decrease unnecessary follow-up procedures and patient anxiety.
3. Improved predictive values. In a subset of 400 MRI exams, the AI system had a superior positive predictive value (PPV) of 68 percent versus 53.2 percent for radiologists, and a higher negative predictive value (NPV) of 93.8 percent compared to 90.2 percent for radiologists. These improved predictive values suggest that the AI model is better at correctly identifying both the presence and absence of csPCa.
“We hypothesize that this difference in performance between the radiologists participating in the reader study and the radiologists reporting in practice was due to those reporting in practice having access to patient history (including previous prostate-specific antigen levels and imaging and biopsy outcomes), peer consultation (or multidisciplinary team meetings), and protocol familiarity,” added Saha and colleagues.
In comparison to radiologist interpretation of multiparametric prostate MRI in the PROMIS study, the researchers said the AI model offered a 34 percent higher specificity rate (45 percent vs. 79.1 percent) and over a 17 percent higher NPV (76 percent vs. 93.8 percent) at similar sensitivity rates (88 percent vs. 89.5 percent).1,2
(Editor’s note: For related content, see “Study: Adjunctive AI Imaging Software Enhances Contouring of Prostate Cancer,” “Researchers Unveil PI-QUAL v2 for Prostate MRI Quality Assessments” and “Study: PSMA PET/CT Agent May Rule Out 93 Percent of PI-RADS 3 Lesions.”)
In regard to study limitations, the authors acknowledged that over 93 percent of the reviewed MRI exams were performed with one MRI device manufacturer. They also noted that reviewing radiologist assessments were provided in a controlled, online reading environment as opposed to native workstations.
References
1. Saha A, Bosma JS, Twilt JJ, et al. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study. Lancet Oncol. 2024 Jun 11:S1470-2045(24)00220-1. Doi: 10.1016/S1470-2045(24)00220-1. Online ahead of print.
2. Ahmed HU, El-Shater AB, Brown LC, et al. Diagnostic accuracy of multi-parametric MRI and TRUS biopsy in prostate cancer (PROMIS): a paired validating confirmatory study. Lancet. 2017;389(10071):815-822.
