TABLE III:
Comparison of techniques that completed the MoNuSeg challenge
| Team Name | a-AJI (95% CI) | Pre-Proc. | Data Augmentation | Model and Arch. | Loss | Post-Proc. | Additional Notes | |||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Color Norm. | Unit Var. | Range Stand. | Hist. Eq. | Rotation | Flipping | Affine deform. | Scaling | Elastic deform. | Noise addition | Color jitter | Intensity jitter | Blur/sharpen | U-Net | Mask RCNN | FCN | PANet | ResNet | VGG-Net | DenseNet | Distance Map | Cross Entropy | Dice loss | L1 loss | L2 loss | Watershed seg. | Non-max supp. | Morph. ops. | |||
| CUHK & IMSIGHT | 0.691 (0.680–0.702) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Macenko CN [42] | ||||||||||||||||||||
| BUPT.J.LI | 0.687 (0.676–0.697) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41]; Deep Layer Aggregation; geometric instance vector | ||||||||||||||||||||||
| pku.hzq | 0.685 (0.675–0.695) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Macenko CN [42]; ResNet with feature pyramid network | ||||||||||||||||
| Yunzhi | 0.679 (0.668–0.690) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Cascaded U-Net with ResNet-like arch. | |||||||||||||||
| Navid Alemi | 0.678 (0.666–0.689) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Concatenated HSV and L channels to RGB multi-headed sphagetti net; smooth Jaccard index for boundary detection; boundary map cleaned by frangi vesselness filter gave markers | ||||||||||||||||||
| xuhuaren | 0.664 (0.652–0.676) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |||||||||||||||||||
| aetherAI | 0.663 (0.653–0.673) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||||
| Shuang Yang | 0.662 (0.652–0.673) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||||||
| Bio-totem & SYSUCC | 0.662 (0.652–0.672) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Color histogram equalization [54] | |||||||||||||||||||
| Amirreza Mahbod | 0.657 (0.649–0.666) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Macenko CN [42]; markers are filtered distance maps | |||||||||||||||
| CMU-UIUC | 0.656 (0.645–0.667) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41]; in-place augmentation of segmented nuclei | ||||||||||||||
| Graham&Vu | 0.653 (0.643–0.663) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41]; combined detection and distance map pred. | ||||||||||||||
| Unblockabulls | 0.651 (0.637–0.666) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Macenko CN [42], concatenated hematoxylin channel | |||||||||||||||||||||
| Tencent AI Lab | 0.646 (0.635–0.657) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Edge enhancement used in pre-processing to separate nuclei | ||||||||||||||||
| DeepMD | 0.633 (0.619–0.647) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41]; random sharpening; TernausNet architecture | |||||||||||||||||
| Canon Medical Research Europe | 0.633 (0.604–0.661) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||||||||
| Johannes Stegmaier | 0.623 (0.603–0.643) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||||||
| Yanping | 0.623 (0.610–0.636) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41] | |||||||||||||||||
| Philipp Gruening | 0.621 (0.606–0.636) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Cross entropy and squared cosine losses | |||||||||||||||||||
| Agilent Labs | 0.618 (0.598–0.638) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |||||||||||||||||||
| Konica Minolta Lab EU | 0.611 (0.601–0.622) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||||||||||
| OnePiece | 0.606 (0.592–0.620) | ✓ | ✓ | ✓ | ✓ | SPCN [41] | ||||||||||||||||||||||||
| Junma | 0.593 (0.581–0.606) | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41], blue channel extraction | |||||||||||||||||||||||
| Biosciences R&D, TCS | 0.578 (0.538–0.619) | ✓ | ✓ | ✓ | ✓ | Smooth Jaccard index; separate nuclei using convexity | ||||||||||||||||||||||||
| Azam Khan | 0.575 (0.556–0.594) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |||||||||||||||||||
| CVBLab | 0.574 (0.560–0.588) | ✓ | ✓ | ✓ | ✓ | SPCN [41] | ||||||||||||||||||||||||
| Linmin Pei | 0.562 (0.548–0.577) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41] | |||||||||||||||||||||
| DB-KR-JU | 0.455 (0.428–0.481) | ✓ | ✓ | ✓ | ✓ | ✓ | Reinhard [43]+Macenko [42] CN; separate nuc. using circularity | |||||||||||||||||||||||
| VISILAB | 0.444 (0.425–0.463) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Macenko CN [42] | ||||||||||||||||||||||
| Sabarinathan | 0.444 (0.424–0.464) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | CLAHE [44] | |||||||||||||||||
| Silvers | 0.278 (0.228–0.328) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | Combination of DenseNet and U-Net | |||||||||||||||
| TJ | 0.130 (0.106–0.154) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | SPCN [41]; ensemble of multiple architectures | |||||||||||||||||||