Table 2. Challenge Resultsa.
| Ranking | Team name | Affiliations | Methods | Training set | Mean sensitivity for biopsied lesions (95% CI) | Phase where team achieved best performance | Code available |
|---|---|---|---|---|---|---|---|
| 1 | NYU B-Team | New York University—Langone Health | Phase 1: EfficientDet, Max-Slice-Selection, and Augmentation and Ensembled Perturbations; phase 2: phase 1 methods with cancer cell prediction head and multilocation crop | Phases 1 and 2: DBTex1 and internal data set | 0.957 (0.924-0.984) | 2 | No |
| 2 | ZeDuS | IBM Research—Haifa | Phase 1: RetinaNet ensemble with heatmap NMS; phase 2: phase 1 methods with SWIN46 and NFNet47 | Phases 1 and 2: DBTex1 with internal data set | 0.926 (0.881-0.964) | 2 | Yes, both phases48 |
| 3 | VICOROB | VICOROB—University of Girona | Phase 1: Fast R-CNN, ensembled; phase 2: phase 1 methods with FP reduction (no ensemble) | Phases 1 and 2: DBTex1 with OPTIMAM/OMI-DB | 0.886 (0.836-0.930) | 2 | Yes, both phases49,50 |
| 4 | Prarit | Queen Mary University of London—CRST and School of Physics and Astronomy | Unknown | Unknown | 0.822 (0.754-0.884) | 1 | No |
| 5 | UCLA-MII | UCLA Medical & Imaging Informatics | Phase 1: Faster R-CNN, FPN,51 IoSIB, and Blob Detector | Phase 1: DBTex1 | 0.814 (0.751-0.875) | 1 | Yes, phase 152 |
| 6 | Pranjalsahu | Stony Brook—Department of Computer Science | Phase 1: Faster R-CNN with Confidence Peak Finder | Phase 1: DBTex1 | 0.790 (0.717-0.854) | 1 | Yes; phase 153 |
| 7 | Team-PittRad | University of Pittsburgh—Department of Radiology | Phase 1: YOLOv554 and Cross Stage Partial Networks | Phase 1: DBTex1 | 0.786 (0.720-0.852) | 1 | Yes, phase 155 |
| 8 | Coolwulf | Unknown | Unknown | Unknown | 0.390 (0.301-0.475) | 1 | No |
| NA | Baseline modelb | NA | Faster R-CNN | DBTex1 | 0.379 (0.304-0.456) | NA | Yes56 |
| NA | Data set baseline modelb | NA | DenseNet32 | DBTex1 | 0.444 (0.366-0.523) | NA | Yes45 |
Abbreviations: FP, false positive; FPN, feature pyramind network; IoSIB, intersection over the smaller intersecting box; NA, not applicable; NFNet, Normalizer-Free-ResNets; NMS, nonmaximal suppression; R-CNN, region-based convolutional neural network; SWIN, shifted window transformer.
a
95% confidence intervals (CI) were computed using bootstrapping, with 5000 bootstraps.
b
Not submitted for challenge.