Table 2.
Characteristics of exams where study radiologists changed assessment after reviewing AI information—from positive to negative, and from negative to positive
| Reader onea—changed assessments after reviewing AI information | Reader twob—changed assessments after reviewing AI information | |||||||
|---|---|---|---|---|---|---|---|---|
| Positive to negative | Truly negative | Negative to positive | Truly positive | Positive to negative | Truly negative | Negative to positive | Truly positive | |
| Total | 21% (66/321) | 62% (41/66) | 0.5% (2/437) | 50% (1/2) | 41% (102/251) | 69% (70/102) | 0.8% (4/507) | 100% (4/4) |
| Bi-RADS | ||||||||
| 3 | 30% (64/215) | 61% (39/64) | 1% (2/217) | 50% (1/2) | 48% (97/201) | 70% (68/97) | 0.5% (1/202) | 100% (1/1) |
| 4 | 3% (1/31) | 100% (1/1) | 0% | N/A | 10% (5/50) | 40% (2/5) | 4% (2/52) | 100% (2/2) |
| 5 | 1% (1/75) | 100% (1/1) | 0% | N/A | 0% | N/A | 0.6% (1/154) | 100% (1/1) |
| Type of image sign | ||||||||
| Calcification | 8% (8/100) | 38% (3/8) | 8% (9/115) | 22% (2/9) | ||||
| Mass | 16% (38/244) | 50% (19/38) | 13% (35/274) | 71% (25/35) | ||||
| Asymmetry | 23% (17/75) | 82% (14/17) | 19% (17/91) | 94% (16/17) | ||||
| Architectural distortion | 29% (29/100) | 79% (23/29) | Insufficient data | 17% (17/102) | 76% (13/17) | Insufficient data | ||
| Skin thickening | 9% (1/11) | 0% | 8% (1/12) | 0% | ||||
| Nipple retraction | 20% (2/10) | 100% (2/2) | 40% (6/15) | 83% (5/6) | ||||
| Axillary lymph node | 15% (3/20) | 100% (3/3) | 10% (2/21) | 100% (2/2) | ||||
| Number of image signs | ||||||||
| 1 | 25% (40/162) | 58% (23/40) | 24% (44/184) | 64% (28/44) | ||||
| 2 | 22% (23/105) | 70% (16/23) | Insufficient data | 17% (20/118) | 80% (16/20) | Insufficient data | ||
| 3 or more | 6% (4/58) | 75% (3/4) | 2% (1/64) | 100% (1/1) | ||||
aThe “Reader one” position was shared between two study radiologists
bThe “Reader two” position was assigned to a single study radiologist who was more experienced
Between the first assessments without AI, there was a 6-week wash-out period before the radiologists again reviewed the exams with AI assistance including image prompts and score