Table 2.
Our Ark-5 and Ark-6 outperform SOTA ImageNet pretrained models and the self-supervised domain-adapted model that utilizes even more training data, highlighting the importance of accruing and reusing knowledge in expert labels from diverse datasets for both classification and segmentation. With the best bolded and the second best underlined, a statistical analysis is conducted between the best vs. others, where green-highlighted boxes indicate no statistically significant difference at level p = 0.05.
| Initialization | Pretraining | Classification task | ||||
|---|---|---|---|---|---|---|
| 1. CXPT | 2. NIHC | 3. RSNA | 4. VINC | 5. NIHS | ||
| Random | - | 83.39±0.84 | 77.04±0.34 | 70.02±0.42 | 78.49±1.00 | 92.52±4.98 |
| Supervised | IN | 87.80±0.42 | 81.73±0.14 | 73.44±0.46 | 90.35±0.31 | 93.35±0.77 |
| SimMIM | IN | 88.16±0.31 | 81.95±0.15 | 73.66±0.34 | 90.24±0.35 | 94.12±0.96 |
| SimMIM | IN→CXR(926K) | 88.37±0.40 | 83.04±0.15 | 74.09±0.39 | 91.71±1.04 | 95.76±1.79 |
| Ark-5(ours) | IN→CXR(335K) | 88.73±0.20 | 82.87±0.13 | 74.73±0.59 | 94.67±0.33 | 98.92±0.21 |
| Ark-6(ours) | IN→CXR(704K) | 89.14±0.22 | 83.05±0.09 | 74.76±0.35 | 95.07±0.16 | 98.99±0.16 |
| Initialization | Pretraining | Segmentation task | ||||
| 6. NIHM | 7. JSRTLung | 8. JSRTHeart | 9. JSRTClavicle | 10. VINR | ||
| Random | - | 96.32±0.18 | 96.32±0.10 | 92.35±0.20 | 85.56±0.71 | 56.46±0.62 |
| Supervised | IN | 97.23±0.09 | 97.13±0.07 | 92.58±0.29 | 86.94±0.69 | 62.40±0.80 |
| SimMIM | IN | 97.12±0.14 | 96.90±0.08 | 93.53±0.11 | 87.18±0.63 | 61.64±0.69 |
| SimMIM | IN→CXR(926K) | 97.10±0.40 | 96.93±0.12 | 93.75±0.36 | 88.87±1.06 | 63.46±0.89 |
| Ark-5(ours) | IN→CXR(335K) | 97.65±0.17 | 97.41±0.04 | 94.16±0.66 | 90.01±0.35 | 63.96±0.30 |
| Ark-6(ours) | IN→CXR(704K) | 97.68±0.03 | 97.48±0.08 | 94.62±0.16 | 90.05±0.15 | 63.70±0.23 |