Chest Radiology Report Binary Classification on MIMIC-CXR (test)
98.4AP (LO)CXR-BERT
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CXR-BERTFine-tuning strategy=all layers (FT-All), Model type=black-box2025.04 | 98.4 | 99.2 | 97 | 96.4 | 96.2 | 64.1 | 98.7 | 99.1 | 97.8 | 98.2 | 95.3 | 54.1 | |
| IP-CRRModel type=interpretable, Confidence threshold=0.85, Max queries=2002025.04 | 97.2 | 57.8 | 95.9 | 89.2 | 92.5 | 46.8 | 91.8 | 35 | 88.9 | 81.1 | 86 | 45.1 | |
| CBMModel type=interpretable by design2025.04 | 94.7 | 34.5 | 93.4 | 79.1 | 87.4 | 43.2 | 88.4 | 24.1 | 85.3 | 73.8 | 80.1 | 43.1 | |
| CXR-BERTFine-tuning strategy=last layer (FT-Last), Model type=black-box2025.04 | 90 | 36.1 | 96.9 | 86.4 | 94.5 | 44.9 | 82.9 | 22.3 | 91.2 | 78.9 | 88.7 | 44.9 | |
| Flan-T5-largeFine-tuning strategy=without fine-tuning, Model type=black-box2025.04 | 52.7 | 7.3 | 44.5 | 38 | 61.6 | 19 | 66.3 | 13.9 | 32.1 | 54.3 | 75.4 | 29.9 |