--- language: en tags: - immunohistochemistry - histopathology - image-feature-extraction - vision-transformer --- # SEPIA: A Marker-Aware Foundation Model for Immunohistochemistry SEPIA learns marker-aware representations of immunohistochemistry (IHC) images. It combines hematoxylin and DAB image representations with protein embeddings that encode the identity of the stained marker. SEPIA is initialized from Hibou-L and trained with H/D inpainting. This repository contains the model weights. ## Model weights | Model | Architecture | Embedding dimension | Weights | |---|---|---|---| | SEPIA | ViT-L/14 | 1024 | sepia.pth | ## Model overview SEPIA takes an RGB IHC tile and its marker identity as inputs. The standard tile size is 224×224 pixels. Images are separated into hematoxylin-only and DAB-only renderings, then normalized using the checkpoint's image statistics. The two renderings share a patch projection and produce paired H/D tokens. A projected ESM protein embedding conditions the DAB tokens on the stained marker. The encoder produces a 1024-dimensional image representation and separate H and D patch representations. ## Benchmarking Test-set AUROC (%) reported as mean ± standard error over five repeats on seven IHC tasks, alongside the mean rank over all 44 IHC-relevant tasks. The pan-cancer DAB localization task uses 13 held-out markers. Best mean values are bold. Lower mean rank is better. | Model | Breast cancer
Histological subtype | Breast cancer
Luminal vs non-luminal | DLBCL
HANS subtype | DLBCL
Double expressor | NSCLC
Intratumoral CD8 | Type-2 diabetes
Diabetic status | Pan-cancer
DAB location (held out) | Overall
Mean rank ↓ (44 tasks) | |---|---:|---:|---:|---:|---:|---:|---:|---:| | DINOv2 | 54.9 ± 0.4 | 81.1 ± 0.7 | 73.2 ± 0.8 | 64.9 ± 1.5 | 71.0 ± 1.4 | 73.6 ± 0.8 | 77.4 ± 0.3 | 7.30 ± 0.19 | | ResNet-50 | 56.0 ± 0.7 | 86.3 ± 0.3 | 77.6 ± 0.6 | 76.6 ± 0.9 | 75.2 ± 0.2 | 72.2 ± 1.1 | 81.4 ± 0.8 | 6.64 ± 0.20 | | PathoDuet | 50.9 ± 1.7 | 85.3 ± 0.5 | 82.8 ± 0.7 | 83.0 ± 0.9 | 77.7 ± 0.6 | 74.9 ± 0.8 | 88.2 ± 0.6 | 5.68 ± 0.25 | | StainNet | 54.9 ± 1.6 | 88.4 ± 0.5 | 87.5 ± 0.9 | 90.2 ± 0.7 | 80.5 ± 0.5 | 79.6 ± 1.2 | 88.4 ± 0.7 | 3.70 ± 0.26 | | UNI | 56.2 ± 1.9 | 87.2 ± 0.7 | 84.9 ± 0.8 | 85.2 ± 0.6 | 82.6 ± 0.4 | 76.1 ± 0.7 | 88.2 ± 0.4 | 3.50 ± 0.18 | | Hibou | 57.0 ± 1.4 | 87.6 ± 0.4 | 86.7 ± 0.4 | 89.1 ± 0.8 | 82.3 ± 0.3 | 73.1 ± 0.7 | 91.1 ± 0.5 | 3.43 ± 0.24 | | UNI2-h | 57.2 ± 0.5 | 87.2 ± 0.6 | 86.3 ± 0.8 | 81.2 ± 1.2 | **83.7** ± 0.3 | 74.0 ± 0.6 | 83.6 ± 0.8 | 3.09 ± 0.32 | | SEPIA | **61.6** ± 1.7 | **89.4** ± 0.3 | **90.9** ± 0.4 | **91.9** ± 0.4 | 83.6 ± 0.2 | **89.3** ± 0.4 | **96.4** ± 0.4 | **2.66** ± 0.27 | ## Acknowledgements We thank the teams behind Hibou, DINOv2, and ESM for their publicly available models and implementations.