Linkedin Labtau  

News

Last updated 1 June 2022

Thursday 7 July 2022

Author: Tristan Jaouen 
Time: 14H00
Language: French
Place: LabTAU Conference Room / Visio Microsoft Teams (to register click the button above)

Abstract

We developed a region of interest-based (ROIs) computer-aided diagnosis system (CAD) to characterize International Society of Urological Pathology grade (ISUP) ≥2 prostate cancers at multiparametric MRI (mp-MRI). Image parameters from two multi-vendor datasets of 265 pre-prostatectomy and 112 pre-biopsy MRIs were combined using logistic regression. The best models used the ADC 2nd percentile (ADC2) and normalized wash-in rate (WI) in the peripheral zone (PZ) and the ADC 25th percentile (ADC25) in the transition zone (TZ). They were combined in the CAD system.

The CAD was retrospectively assessed on two multi-vendor datasets containing respectively 158 and 105 pre-biopsy MRIs from our institution (internal test dataset) and another institution (external test dataset). Two radiologists independently outlined lesions targeted at biopsy. The Prostate Imaging-Reporting and Data System version 2 (PI-RADSv2) score prospectively assigned at biopsy and the CAD score were compared to biopsy findings. At patient level, the areas under the Receiver Operating Characteristic curve (AUC) of the PI-RADSv2 score were 82% (95% CI: 74-87) and 85% (95% CI: 79-91) in the internal and external test datasets respectively. For both radiologists, the CAD score had similar AUC results in the internal (82%, 95% CI: 76-89, p=1; 84%, 95% CI: 78-91, p=1) and external (82%, 95% CI: 76-89, p=0.82; 86%, 95% CI: 79-93, p=1) test datasets. Combining PI-RADSv2 and CAD findings could have avoided 41-52% of biopsies while missing 6-10% of ISUP≥2 cancers.

The CAD system confirmed its robustness showing good discrimination of ISUP ≥2 cancers in a multicentric study involving 22 different scanners with highly heterogeneous image protocols. In per patient analysis, the CAD and the PI-RADSv2 had similar AUC values (76%, 95% CI: 70-82 vs 79%, 95% CI: 73-86; p=0.34) and sensitivities (86%, 95% CI: 76-96 vs 89%, 95% CI: 79-98 for PI-RADSv2 ≥4). The specificity of the CAD (62%, 95% CI: 53-70 vs 49%, 95% CI: 39-59 for PI-RADSv2 ≥4) could be used to complement the PI-RADSv2 score and potentially avoid 50% of biopsies, while missing 13% of ISUP ≥2 cancers. These findings were very similar to those reported in the single center test cohorts. Given its robustness, the CAD could then be exploited in more specific applications.

The CAD first provided good discrimination of ISUP ≥2 cancers in patients under Active Surveillance. Its AUC (80%, 95% CI: 74-86) was similar to that of the PI-RADS score prospectively assigned by specialized uro-radiologists at the time of biopsy (81%, 95% CI: 74-87; p=0.96). After dichotomization, the CAD was more specific than the PI-RADS ≥3 (p<0.001) and the PI-RADS ≥4 scores (p<0.001). It could offer a solution to select patients who could safely avoid confirmatory or follow-up biopsy during Active Surveillance (25%), while missing 5% of ISUP≥2 cancers.

Finally, the CAD was tested with the pre-prostatectomy mp-MRIs of 56 Japanese patients, from a population which is geographically distant from its training population and which is of interest because of its low prostate cancer incidence and mortality. The CAD obtained an AUC similar to the PI-RADSv2 score assigned by an experience radiologist in the PZ (80%, 95% CI: 71-90 vs 80%, 95% CI: 71-89; p=0.886) and in the TZ (79%, 95% CI: 66-90 vs 93%, 95%CI: 82-96; p=0.051).

These promising and robust results across heterogeneous datasets suggest that the CAD could be used in clinical routine as a second opinion reader to help select the patients who could safely avoid biopsy. This CAD may assist less experience readers in the characterization of prostate lesions.