Table 2.
PICTURE dataset CSPCa lesion detection metrics for the experienced radiologist and AutoProstate. Mean and standard deviation of false positives per patient were calculated using the 93 PICTURE dataset patients who were biopsy-negative for CSPCa, rather than over all patients, due to limitations in the ground-truth. All other metrics shown are calculated at the lesion level for the 147 CSPCa lesions and 63 nCSPCa lesions.
| Experienced Radiologist (Likert Scoring) | |
| ROC AUC | 0.64 (0.56–0.72) |
| PR AUC | 0.78 (0.71–0.84) |
| Post-thresholding (cut-off: Likert ≥4) | |
| Sensitivity/recall (%) | 78 (71–84) |
| Specificity (%) | 48 (35–60) |
| Precision (%) | 78 (71–84) |
| Mean false positives per patient | 0.3 (0.2–0.4) |
| AutoProstate | |
| ROC AUC | 0.70 (0.62–0.78) |
| PR AUC | 0.84 (0.77–0.90) |
| Post-thresholding (cut-off: ≥4.5%) | |
| Sensitivity/recall (%) | 78 (71–85) |
| Specificity (%) | 49 (37–62) |
| Precision (%) | 78 (71–85) |
| Mean false positives per patient * | 6.1 (5.5–6.8) |
| Post-thresholding (cut-off: ≥4.5%) and false-positive reduction (40 mm3) | |
| Sensitivity/recall (%) | 76 (68–82) |
| Specificity (%) | 57 (45–69) |
| Precision (%) | 80 (74–87) |
| Mean false positives per patient * | 2.5 (2.2–2.8) |
AUC: area under curve; PR: precision-recall; ROC: receiver operating characteristic. * indicates a p-value 0.001 for AutoProstate compared to the radiologist.