Loki-OT fixes cell mimicry in pathology with tissue-aware AI reasoning
New transport-based method beats PanopTILs on breast cancer MAE with just 278 weak labels.
Lymphocyte mimicry occurs when different cell types look alike, and while pathologists use surrounding tissue context to disambiguate them, current vision models either lack contextual reasoning or can't operate at the cell level. To address this, the paper presents Loki-OT, which propagates region-level tissue reasoning to individual cell predictions via Unbalanced Optimal Transport, using MLLM-derived density priors as soft guidance for ambiguous cell reassignment. The transport plan is distilled into a lightweight student MLP classifier that learns context-aware decision boundaries within a pretrained feature space. On the independent TCGA-BRCA cohort, Loki-OT achieved lower patient-level MAE than the fully supervised in-domain PanopTILs classifier and improved F1 in epithelium-rich mimicry tissues, using 278 weak region-level MLLM estimates built on a general-domain cell foundation model. Code is available via the provided link.
- Loki-OT uses Unbalanced Optimal Transport to propagate tissue context from region-level MLLM density priors to single-cell predictions.
- Achieved lower patient-level MAE than fully supervised PanopTILs on the TCGA-BRCA cohort, using only 278 weak region-level labels.
- Distills the transport plan into a lightweight MLP student classifier, enabling context-aware decisions in pretrained cell foundation feature spaces.
Why It Matters
Brings clinically viable cell classification for pathology by cutting annotation costs and resolving mimicry errors that stump current AI.