Table 1 The sensitivity of GMIC and GLAM models with pre-trained and transfer learning modes for three categories of cancers in four concordance levels in Lifepool dataset.
Interpretation | Missed cancers | Prior vis cancers | Prior invis cancers | |||
|---|---|---|---|---|---|---|
Pre-trained | Transfer learning | Pre-trained | Transfer learning | Pre-trained | Transfer learning | |
GMIC | ||||||
Almost perfect perfect perfect | 72.6% (37) | 86.3% (44) | 84.3% (43) | 94.1% (48) | 81.1% (86) | 92.5% (98) |
Substantial | 70.2% (33) | 85.1% (40) | 82.1% (32) | 92.3% (36) | 79.8% (83) | 90.4% (94) |
Moderate | 66.7% (32) | 81.3% (39) | 80.0% (28) | 88.6% (31) | 77.3% (75) | 87.6% (85) |
Poor | 64.3% (27) | 78.6% (33) | 76.7% (23) | 85.3% (29) | 74.7% (59) | 84.8% (67) |
GLAM | ||||||
Almost perfect | 70.6% (36) | 84.3% (43) | 82.4% (42) | 92.2% (47) | 80.2% (85) | 91.5% (97) |
Substantial | 68.1% (32) | 83.0% (39) | 79.4% (31) | 89.7% (35) | 76.9% (80) | 87.5% (91) |
Moderate | 64.6% (31) | 77.1% (37) | 77.1% (27) | 85.7% (30) | 75.3% (73) | 85.6% (83) |
Poor | 61.9% (26) | 73.8% (31) | 73.3% (22) | 83.3% (25) | 72.2% (57) | 81.0% (64) |
