Artificial intelligence (AI) software may offer similar detection of clinically significant prostate cancer (csPCa) on magnetic resonance imaging (MRI) as multidisciplinary team (MDT)-supported radiologist assessment, according to new multicenter research.
For the retrospective study, recently published in European Radiology, researchers developed a deep learning software (Prostate Intelligence™ Pi-v2.4, Lucida Medical) for detection of Gleason grade group (GG) > 2 tumors on prostate MRI data for 793 patients drawn from five United Kingdom hospitals as well as the PROSTATEx dataset. The researchers subsequently evaluated the software on MRI data from 252 patients (mean age of 67.3) drawn from six facilities. Thirty-one percent of the validation cohort had GG > 2 tumors, according to the study.
At an AI Likert threshold of 3.5, researchers found that the AI model offered an area under the receiver operating characteristic curve (AUC) of 91 percent in comparison to 95 percent for radiologist assessment. In validation testing, the AI model provided 95 percent sensitivity and 67 percent specificity in contrast to 99 percent sensitivity and 73 percent specificity for radiologist evaluation, according to the study authors.
“If high specificity can be replicated in general use while maintaining sensitivity, DL-(computer-assisted diagnosis) CAD may enable reductions in biopsies and associated costs without missing a significant additional number of men with GG ≥ 2 cancers,” wrote lead study author Francesco Giganti, M.D., an associate professor in the Department of Radiology at University College London in the United Kingdom, and colleagues.
The researchers noted that the AI model detected 86 percent of GG > 2 lesions in comparison to 93 percent with radiologist interpretation. While the study authors pointed out the AI model’s high specificity at the patient level, they conceded consistently lower false positive rates with radiologist interpretation at 80 percent and 90 percent sensitivity thresholds.
Three Key Takeaways
1. AI performance close to that of radiologists. The deep learning (DL) model demonstrated strong diagnostic accuracy for clinically significant prostate cancer (csPCa), achieving an AUC of 91 percent compared to 95 percent for radiologists, with a sensitivity of 95 percent and specificity of 67 percent.
2. Potential for reducing unnecessary biopsies. If high specificity can be maintained while preserving sensitivity, AI-assisted computer-aided diagnosis (CAD) may help reduce unnecessary biopsies and associated healthcare costs without missing a significant number of GG ≥ 2 cancers.
3. AI as a decision-support tool, not a standalone solution. The study authors noted the AI software is intended to assist radiologists and support multidisciplinary team (MDT) decision-making in prostate cancer detection. Further prospective studies are needed to refine its clinical application.
Accordingly, the researchers emphasized adjunctive use of the AI software in tandem with radiologist evaluation.
“(This AL software) is not intended as a stand-alone lesion-level biopsy targeting application but is a decision-support tool to assist radiologists based on their experience as well as on clinical assessments in an MDT environment,” said Giganti and colleagues. “Prospective studies are required to determine the optimal clinical approach to additional AI-identified lesions, balancing the harm and costs associated with additional targets (potential additional detection of both clinically indolent and csPCa) based on urological preferences.”
(Editor’s note: For related content, see “Emerging Concepts and Recommendations for MRI in Prostate Cancer Screening,” “Can MRI-Based Deep Learning Improve Risk Stratification in PI-RADS 3 Cases?” and “Can Deep Learning Radiomics with bpMRI Bolster Accuracy for Prostate Cancer Prognosis?”)
In regard to study limitations, the authors acknowledged the 10 percent non-inferiority margin for comparing the adjunctive AI model and radiologist assessment. They also conceded that 46 percent of patients in the cohort did not have a biopsy and noted that datasets for AI model development and validation were drawn from the same population groups.
