292nd LG : Computer Aided Diagnosis for the Detection of Prostate Cancer on mpMRI: A Multi-Reader Study

Friday, October 28, 2016 - 12:00
Dr. Matt Greer
Purpose: To evaluate the hypothesis that computer-aided diagnosis (CAD) of prostate mpMRI, when interpreted in a first-reader design, can improve the sensitivity and agreement of radiologists across experience levels.   Materials and Methods: 9 radiologists (with n=3 high, n=3 moderate, n=3 low experience levels in interpreting prostate MRI) from 8 institutions participated in a multi-reader study. 163 consecutive patients who underwent multi-parametric prostate MRI (T2W, ADC, b2000, and DCE MRI) with ERC at 3T. 110 patients underwent prostatectomy after mpMRI, and 53 patients were controls with no lesions on MRI with a negative TRUS biopsy. Readers were blinded to all outcomes and prospectively detected clinically relevant lesions per PI-RADSv2 on mpMRI alone. After 5 weeks, readers evaluated patients with CAD assistance in a first-reader design and scored mpMRI with PI-RADSv2. Lesions were correlated to whole mount prostatectomy specimens processed with customized 3D-printed molds. Average sensitivity, specificity, and positive predictive value (PPV) were calculated on a per-patient and per-lesion basis. Index of specific agreement (ISA) was calculated among all readers and experience levels.  Results: On lesion-based analysis, the average sensitivity for detecting index lesions of any grade was 78.2% for mpMRI and 86.3% for CAD (p=0.013) for PI-RADS≥1. With reader exclusion of false-positives, PI-RADS≥3, the sensitivity was 77.6% for MRI vs 78.1% for CAD (p=0.859). Sensitivity was equivalent for TZ lesions (78.7% vs 78.1%), but CAD improved sensitivity in PZ at PI-RADS≥1 (92.9% vs 83.9%) and PI-RADS≥3 (88.0% vs 83.3%). Index lesion sensitivity increased by 8.4% on average for each reader with CAD (p=0.02). On a per-patient basis the average sensitivity was 92.7% vs 98.0% for MRI and CAD (p=0.002). Specificity of CAD and MRI were 57% and 70% (p=0.003), respectively, and was strongly experiential dependent. Agreement was 93.7% for detection of index lesions with CAD versus 87.1% (p<0.001) for MRI. Agreement between high and low experience readers improved from 59.8% to 72.5% with CAD (p<0.001). No difference was observed in time to make reads (p=0.645).  Conclusion: With the assistance of CAD as a first-reader, readers with varying experience improved sensitivity and interreader agreement for index lesion detection on prostate mpMRI.