Architecture & Methodology
The Science Behind Synthia
Deep learning for IHC biomarker quantification — trained on pathologist-annotated whole-slide images from academic cancer centers.
MODEL PIPELINE
20× / 40× magnification
256×256px patches
ResNet-50 + Transformer
Per ASCO/CAP guidelines
Model Architecture
Convolutional + Attention Architecture for WSI Analysis
Whole-slide images present a unique computational challenge: a single scan at 40× magnification can exceed 100,000 × 80,000 pixels. Direct processing is computationally intractable and unnecessary — relevant diagnostic information is localized at the cellular level.
Synthia's architecture decomposes the problem into two stages. First, a tile extraction and tissue segmentation pipeline identifies IHC-relevant tissue regions, excluding glass background, fat, and stroma. Second, a hybrid convolutional-attention model processes each 256×256 pixel tile at 20× magnification, performing simultaneous cell detection and staining intensity quantification.
The backbone is a ResNet-50 encoder pre-trained on histopathology data, fine-tuned on our IHC-specific annotation corpus. Transformer-based attention heads provide long-range spatial context — critical for identifying IHC hotspots in Ki-67 proliferation scoring and distinguishing membrane vs. cytoplasmic staining in HER2 assessment.
Score aggregation follows biomarker-specific logic: for HER2, membrane-positive cell counting per the modified H-score system; for PD-L1, separate tumor proportion score (TPS) and combined positive score (CPS) pathways; for Ki-67, global percentage and spatial hotspot localization.
WSI PROCESSING PIPELINE
WSI → tissue mask → 256px patches
ResNet-50 encoder (IHC pre-trained)
DAB-positive detection + intensity class
Transformer spatial heads → hotspot map
HER2 / PD-L1 / Ki-67 clinical output
Training Data
Annotation Methodology and Dataset Scale
Every training annotation went through a multi-reader adjudication protocol. Agreement below consensus threshold triggered mandatory third-reader adjudication.
Pathologist Consensus Annotation Protocol
All training annotations were produced by board-certified anatomic pathologists with subspecialty expertise in oncologic pathology. Each whole-slide image received two independent reads; cases with inter-reader kappa below 0.70 were escalated to a third reader for adjudication.
Annotation was performed at cell-level resolution for HER2 membrane staining (0/1+/2+/3+ per cell), pixel-level DAB thresholding for PD-L1 (tumor cell vs. immune cell positivity), and nucleus segmentation for Ki-67 (positive/negative count per field).
Cases from externally procured de-identified tissue archives covering breast, gastric, lung, and bladder cancer histology types. No real institution names are disclosed in this summary.
Training Dataset Summary
Internal training set. Data on file. Figures are approximate. Kappa: weighted kappa between two primary readers prior to adjudication.
Algorithmic Reproducibility
The Model Scores the Same Slide the Same Way. Every Time.
Unlike human scoring, Synthia's output is deterministic. The same WSI submitted twice returns identical results.
HER2 Test-Retest ICC
Same WSI submitted at 7-day interval. Intraclass correlation coefficient measures between-run consistency. Internal validation, n=50 WSI.
Ki-67 Test-Retest ICC
Nuclear detection consistency across repeat analyses. Includes natural variation from stochastic tile sampling. All runs exceeded ICC 0.90 threshold.
Classification Stability
0+ / 1+ / 2+ / 3+ grade assignment from the same WSI is deterministic — no stochastic classification drift between identical submissions.
Evaluate the Science on Your Data
Request a pilot study with your own de-identified WSI set. We'll deliver scored results and a full validation report.
Request Pilot Access