An emerging deep learning model may provide the same level of detection for clinically significant prostate cancer (csPCa) on magnetic resonance imaging (MRI) as experienced abdominal radiologists, according to new research findings.
For the retrospective study, recently published in Radiology, researchers reviewed data from 5,735 multiparametric MRI exams for a total of 5,215 patients being evaluated for prostate cancer. For external validation testing, the study authors compared a deep learning model versus four reviewing abdominal radiologists with at least 10 years of experience and a combination of deep learning and radiologist assessment.
Trained on 5,035 prostate MRI exams, the deep learning model incorporated assessment of T2-weighted images, contrast-enhanced T1 images, diffusion-weighted MRI, and apparent diffusion coefficient mapping, according to the study.
The study authors found that the deep learning model had an 86 percent area under the receiver operating characteristic curve (AUC) for diagnosing csPCa in contrast to 84 percent for experienced radiologist interpretation in external validation testing. The combination of deep learning and radiologist interpretation had an AUC of 89 percent for csPCa, according to the researchers.
“There was no evidence of a difference in image-only model performance from that of experienced radiologists …, and the (combined) model was better than the radiologist alone,” wrote study co-author Naoki Takahashi, M.D., who is affiliated with the Department of Radiology at the Mayo Clinic in Rochester, Minn.
Noting that the DL model didn’t provide tumor location, the researchers utilized gradient-weighted class activation mapping (Grad-CAM). The study authors found that GRAD-CAM highlighted the csPCa lesion in 56 of 58 true-positive exams (97 percent) in external validation testing.
Three Key Takeaways
1. Comparable diagnostic accuracy. The deep learning (DL) model demonstrated a similar level of diagnostic accuracy for clinically significant prostate cancer (csPCa) on MRI as experienced abdominal radiologists, with an 86 percent area under the curve (AUC) compared to 84 percent for experienced radiologists.
2. Enhanced performance with combined assessment. The combination of DL model and radiologist interpretation improved diagnostic performance, achieving an AUC of 89 percent, surpassing either approach alone.
3. Limitations in tumor localization with Grad-CAM. While gradient-weighted class activation mapping (Grad-CAM) reliably highlighted csPCa, the technique had limitations, including poor spatial resolution and challenges in highlighting multiple lesions, necessitating radiologists to manually draw regions of interest for accurate MRI-guided biopsy.
While the Grad-CAM provided reliable localization of csPCa, the researchers cautioned that Grad-CAM wasn’t effective in highlighting multiple lesions, had poor spatial resolution and may have limited csPCa detection to the superior, middle and inferior third regions of the prostate.
“Due to these limitations, the Grad-CAM output cannot be directly used as a region of interest for subsequent MRI-guided (fusion) biopsy, and radiologists still need to draw regions of interest in areas of suspicion,” maintained Takahashi and colleagues.
(Editor’s note: For related content, see “MRI-Based AI Radiomics Model Offers ‘Robust’ Prediction of Perineural Invasion in Prostate Cancer,” “Could MRI-Based AI Offer Better Risk Stratification for Prostate Cancer than PI-RADS?” and “MRI-Based AI Model Facilitates 50 Percent Reduction in False Positives for Prostate Cancer.”)
Beyond the inherent limitations of a single-center retrospective study, the researchers noted the reviewing radiologists specialized in prostate MRI interpretation. They also acknowledged that the deep learning model was trained on dynamic contrast-enhanced MRI sequences.
