New research suggests that deep learning assessment of prostate biparametric magnetic resonance imaging (MRI) scans may be comparable to assessment by experienced genitourinary radiologists in detecting clinically significant prostate cancer (csPCa).
For the study, recently published in Radiology, researchers compared deep learning algorithm assessment and radiologist assessment for diagnosing csPCa in 658 men (median age of 67) with a total of 1,029 MRI-visible lesions. Patients in the cohort had a mean prostate-specific antigen (PSA) level of 6.7 ng/mL, according to the study.
The study authors found that 45 percent of the cohort (294 patients) had International Society of Urological Pathology (ISUP) grade group (GG) 2 or higher prostate lesions. They found that the deep learning algorithm detected csPCa in 282 of these patients (96 percent) in comparison to 287 cases of diagnosed csPCa (98 percent) by a genitourinary radiologist with more than 15 years of experience.
The researchers also noted that the deep learning algorithm offered comparable participant-level sensitivity rates (92 percent vs. 93 percent) and positive predictive value (PPV) (65 percent vs. 69 percent) to that of the reviewing genitourinary radiologist.
“AI has the potential to assist radiologists by standardizing intraprostatic lesion detection and reducing variability,” wrote study co-author Baris Turkbey, M.D., a senior clinician and radiologist at the National Cancer Institute and the National Institutes of Health in Rockville, Md., and colleagues.
While radiologists may consider additional factors with peripheral aspects of tumor segmentation, the researchers acknowledged that the algorithm tends to focus on the central region of lesions but also reins in false-positive findings beyond the tumor foci with a benign prostatic hyperplasia filter.
Three Key Takeaways
- Deep learning aids in standardized lesion detection. The study suggests that deep learning algorithms can aid radiologists by standardizing intraprostatic lesion detection, potentially reducing variability in diagnosis. This standardization may enhance the consistency of diagnoses across different radiologists and health-care settings.
- Comparable performance to experienced radiologists. The research suggests that deep learning assessment of prostate MRI scans can achieve comparable results to assessment by experienced genitourinary radiologists in detecting clinically significant prostate cancer (csPCa). The study findings indicate that such algorithms could serve as valuable decision support tools for radiologists, particularly in cases involving high-grade lesions.
- Efficiency and potential for automation. The deep learning algorithm demonstrated high detection rates, particularly with PI-RADS category 5 lesions, and may offer efficiencies in lesion segmentation. The exportable lesion segmentation masks provided with the algorithm could facilitate streamlined workflows that may help improve biopsy and radiation therapy planning in this patient population.
Emphasizing the deep learning algorithm’s 92 percent detection rate with PI-RADS category 5 lesions and its 0.58 Dice similarity coefficient (DSC) for segmentation of those lesions (double that of the overall DSC (0.29) for the algorithm), the study authors noted potential efficiencies in the management of patients with higher-grade prostate cancer.
“ … Not only does the algorithm offer decision support by highlighting lesions and additional areas of concern for radiologists, but its exportable lesion segmentation masks also present an opportunity for automating the labor-intensive processes of lesion segmentation (especially for PI-RADS 5 lesions), streamlining tasks related to biopsy and radiation therapy planning,” pointed out Turkbey and colleagues.
(Editor’s note: For related content, see “Predicting Clinically Significant Prostate Cancer: Can a Prostate MRI Point-Based Model Have an Impact?,” “PET/CT or mpMRI: Which is Better for Detecting Biochemical Recurrence of Prostate Cancer?” and “Going Beyond PI-RADS: A Closer Look at Emerging Prostate MRI Scoring Systems.”)
In regard to study limitations, the authors acknowledged that inconsistent MRI acquisition in the initial study population prevented incorporation of dynamic contrast-enhanced imaging in the AI model. They conceded that variations with diffusion MRI protocols may affect the performance of the model.
