Interpretable science. Human impact.
We develop artificial intelligence connecting pathology, radiology and molecular information. So that every image becomes knowledge a professional can interpret.
Our purpose
Mission
We start with lung cancer — the leading cause of cancer death worldwide — because treatment selection already depends on molecular biomarkers that are today slow, costly and unevenly available, especially in low- and middle-income settings. Our work is aimed at shortening the distance between a biopsy and a therapeutic decision.
Vision
Digital pathology, MRI, CT and radiology all share the same challenge: turning an image into information that another clinical domain — genotype, prognosis, treatment response — can trust and act on. We build the methodology for that cross-domain translation, and validate it first where the data is scarcest and hardest to interpret.
Whole-slide attention
Our multiple-instance learning model divides a whole-slide image into patches and assigns each one an attention weight relative to the mutation under evaluation. Decisive regions are linked to an explanation the professional can inspect.
Conceptual attention visualization, not a patient result.
Research areas
Our most mature evidence is in digital pathology. The architecture — patch-level representation, attention-weighted fusion and knowledge-graph verification — is designed to extend to other medical imaging modalities.
PRIMARY EVIDENCE
Classification of histological growth patterns in lung adenocarcinoma using multiple-instance learning (MIL) and fuzzy angular-margin losses, trained on the public TCGA-LUAD cohort.
Flagship research
Our mutation-prediction system from H&E biopsies treats the task as an imperfect translation between two domains: tissue morphology and somatic driver mutations. Instead of reporting a single cohort-average AUROC, each gene gets its own verdict.
The reasoning behind every result
Conceptual workflow model. Not a diagnostic device.
A fuzzy angular-margin classifier (FuzzyArcLoss) identifies the six lung adenocarcinoma growth patterns per patch, reaching a macro-F1 of 92.3%.
TP53 · EGFR
AUROC between 0.70–0.72 from the image embedding alone: the morphology by itself already carries the signal.
KRAS
A 2-additive Choquet aggregator recovers a lepidic × solid interaction of +0.032, consistent with the invasive-mucinous-adenocarcinoma phenotype.
STK11 · KEAP1 · RBM10
The signal lives outside the six histological patterns — in the immune microenvironment or a biochemical state — and the system reports that instead of forcing an answer.
The platform
LCHAI does not hand back a bare probability. Every per-gene verdict is exhibited as a translation — from the source morphology, through an attention-weighted aggregation, to a knowledge graph that ties each claim back to the WHO/IASLC histological vocabulary a pathologist already reasons in.
Try it yourself
Walk a real slide through the six-level explanation stack: attention maps, the knowledge-graph explanation and the per-gene Conclusive / Inconclusive verdict.
Research platform. It does not replace professional medical assessment.
Evidence · LCHAI v2
+3.1pp
Gain at the cribriform boundary vs. SphereFace, under a paired test.
κ 0.46
Inter-pathologist agreement on ASC-US/LSIL cervical cytology.
18
Loss functions compared under one protocol of 50 runs (5 folds × 10 seeds).
2
External validation domains: prostate Gleason grading (SICAPv2) and cervical cytology (CRIC).
Build the next step