Gallery
Virtual IHC generated by SheafStain from H&E, as stitched 1024×1024 outputs. Drag a slider to compare the input H&E with SheafStain. Click a case to compare the ground truth, SheafStain and six prior methods.
H&E and IHC come from consecutive tissue sections, so the outputs are not expected to align with the ground truth at the pixel level.
Virtual HER2 IHC on the BCI7 test set: 977 H&E–IHC pairs held out from training.
Rows are HER2 scores, from 0 (top) to 3+ (bottom).
Virtual HER2 IHC on the MIST8 HER2 test set: 1,000 H&E–IHC pairs held out from training.
Rows are DAB-staining intensity levels of the ground-truth IHC, from weakest (top) to strongest (bottom).
Virtual ER IHC on the MIST8 ER test set: 1,000 H&E–IHC pairs held out from training.
Rows are DAB-staining intensity levels of the ground-truth IHC, from weakest (top) to strongest (bottom).
Virtual PR IHC on the MIST8 PR test set: 1,000 H&E–IHC pairs held out from training.
Rows are DAB-staining intensity levels of the ground-truth IHC, from weakest (top) to strongest (bottom).
Virtual Ki-67 IHC on the MIST8 Ki-67 test set: 1,000 H&E–IHC pairs held out from training.
Rows are DAB-staining intensity levels of the ground-truth IHC, from weakest (top) to strongest (bottom).
The sample list could not be loaded. Serve this folder over HTTP (for example python3 -m http.server) rather than opening the file directly.
References (2)
- 7S. Liu, C. Zhu, F. Xu, X. Jia, Z. Shi, and M. Jin, BCI: Breast cancer immunohistochemical image generation through pyramid pix2pix, CVPR Workshops 2022.
- 8F. Li, Z. Hu, W. Chen, and A. Kak, Adaptive supervised PatchNCE loss for learning H&E-to-IHC stain translation with inconsistent groundtruth image pairs, MICCAI 2023.