Table 2.
Comparison of AI and clinical assessment methods in the diagnosis field of PCa
| Comparison | AI methods | Clinical assessment methods | |
|---|---|---|---|
| HC model | DL model | ||
| Overall performance | Relatively high | Relatively poor | |
| SROC-AUC* | 0.87 | 0.82 | |
| Pooled sensitivity* | 0.90 | 0.93 | |
| Pooled specificity* | 0.60 | 0.46 | |
| Qualitative or quantitative | Quantitative | Semi-quantitative | |
| Expert dependence | Moderate | Low | High |
| Consistency | High | Low | |
| Manual delineation | Yes | No | No |
| Features | High-throughput features extracted using specific algorithms (e.g., shape, histogram, and textural features) | Automatic extraction of deep and subtle image features using networks with substantial parameters | Features for visual assessments (e.g., location, shape, size, and intensity) and some clinical characteristics |
AI artificial intelligence, csPCa clinically significant prostate cancer, DL deep learning, HC hand-crafted, PCa prostate cancer, PI-RADS prostate imaging reporting and data system, SROC-AUC area under the summary receiver operating characteristic curves
*Performance indexes pooled across the studies on csPCa diagnoses