Lymphocyte Mimicry Correction via Region-Level Tissue Reasoning and Unbalanced Optimal Transport

Xiang Li; Yuqi Wang; Casey C. Heirman; Jihye Heo; Kyle J. Lafata

cs.CVAug 17, 2026

Lymphocyte Mimicry Correction via Region-Level Tissue Reasoning and Unbalanced Optimal Transport

Authors: Xiang Li, Yuqi Wang, Casey C. Heirman, Jihye Heo, Kyle J. Lafata

Organizations: Department of Electrical and Computer Engineering, Duke University, US · Medical Physics Graduate Program, Duke University, US · Department of Biomedical Engineering, Duke University, US · Department of Radiation Oncology, Duke University, US · Department of Radiology, Duke University, US · Department of Mathematics, Duke University, US

Abstract

Cell mimicry arises when different cell types appear morphologically similar. Human pathologists resolve this ambiguity using surrounding tissue context, whereas current vision models either lack contextual reasoning (cell foundation models) or cannot operate at the cell level (pathology MLLMs). We present 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. Loki-OT is motivated by the observation that pretrained cell foundation model features already encode discriminative information, including tissue context, but standard cell-level supervision fails to use tissue context effectively. The resulting transport plan is distilled into a lightweight student MLP classifier that learns context-aware decision boundaries within the 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: https://github.com/xiangli980/Lymphocyte_Mimicry_Correction_via_Loki_OT

Figures & tables

Figure 1: Loki-OT pipeline. (a) CellViT++ extracts per-cell morphological and context embeddings. (b) An MLLM provides tissue-level lymphocyte density priors. (c) The UOT arbiter balances perceptual evidence against biological priors to produce soft teacher labels. (d) A lightweight MLP is distilled via a two-stage curriculum. At inference, only the distilled MLP runs, no MLLM or OT computation required.

	Train (2 hospitals)				Test (TCGA-BRCA)
Tissue subtype	Nuclei	GT%	MLLM%	Lizard%	Nuclei	GT%	Lizard%
T1 Invasive tumor	35K	1.3	13.9	28.6	26K	2.4	24.4
T2 Tumor-assoc. stroma	18K	20.3	20.8	31.0	14K	23.2	38.3
T3 In-situ tumor	7K	0.2	11.0	35.8	2K	0.9	14.3
T4 Healthy glands	7K	1.8	2.8	31.2	0.2K	7.0	36.0
T5 Necrosis	0.6K	5.4	0.3	38.1	0.3K	1.5	52.1

Table 1: Lymphocyte density (%) by tissue subtype. Nuclei columns show train/test counts. Lizard over-predicts in epithelium-rich regions (T1, T3, T4); MLLM is coarse but directionally closer to GT than Lizard.

Method	F1 $\uparrow$	Recall $\uparrow$	Prec $\uparrow$	AUC $\uparrow$	AP $\uparrow$	MAE $\downarrow$
Lizard	0.526 [.492,.557]	0.743 [.707,.776]	0.393 [.354,.430]	0.787 [.768,.805]	0.544 [.509,.578]	83.8 [65.6,104.7]
Context-soft	0.579 [.547,.606]	0.760 [.721,.797]	0.436 [.396,.473]	0.823 [.806,.838]	0.568 [.537,.603]	68.5 [54.4,84.4]
Loki-OT	0.543 [.512,.573]	0.637 [.599,.675]	0.455 [.418,.496]	0.809 [.792,.826]	0.554 [.521,.587]	46.0 [35.5,56.9]
PanopTILs	0.474 [.435,.511]	0.430 [.380,.477]	0.561 [.516,.604]	0.816 [.795,.835]	0.578 [.544,.612]	58.3 [45.4,71.5]

Table 2: TCGA-BRCA independent cohort (124 patients). F1, recall, and precision are globally micro-averaged; MAE is computed at the patient level. Point estimate [95% bootstrap CI,

N{=}1000

]. Bold: best per column. Wilcoxon with Holm correction: Loki-OT vs. PanopTILs MAE

p{=}0.031

Figure 2: Per-ROI cell classification on representative TCGA-BRCA test cases. Columns: Ground Truth, Lizard, Context-soft, Loki-OT, PanopTIL; contours: green = FP, blue = TP, yellow = FN. (a) T3 in-situ tumor (mimics cleanly eliminated); (b) mixed T1/T2/T6 boundary; (c) T2 tumor-associated stroma; (d) T1 invasive tumor (hard mimics persist, limitation).

Method	FP	FP $\downarrow$ %	TP	TP $\downarrow$ %	Select.	F1	MAE $\downarrow$
Baseline (Lizard)	17,960	—	6,550	—	—	0.269	135.6
+ Context-soft	12,046	32.9	6,275	4.2	7.84 $\times$	0.331	87.5
+ OT-Teacher	6,017	66.5	4,120	37.1	1.79 $\times$	0.328	48.8
+ UOT-Teacher	7,469	58.4	5,024	23.3	2.50 $\times$	0.351	55.4
+ Loki-OT	7,898	56.0	5,449	16.8	3.32 $\times$	0.366	56.4

Table 3: Ablation on TIGER training set (135 ROIs, 80K nuclei). OT/UOT-Teacher rows are solver outputs (no distillation); Loki-OT is the deployable distilled model. Bold marks the training performance of the best deployable model (OT/UOT-Teacher are non-deployable).

Figure 3: Attention analysis at cell-token level (GradCAM) and feature level (channel attribution), for Stage 1 (Context-soft) vs. Stage 2 (Loki-OT). (a,b) Spatial GradCAM on a T2-stroma and a T1-invasive-tumor mimic. (c) Channel-weighted feature-importance difference map for cell (a); red: up-weighted by Stage 2, blue: by Stage 1.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Tissue subtype ( $n$ )	Lizard	PanopTILs	Context-soft	Loki-OT
T1 Invasive tumor (1108)	5.24 [4.83,5.69]	1.14 [.96,1.34]	3.69 [3.38,4.01]	2.63 [2.41,2.86]
T2 Tumor-assoc. stroma (1169)	2.51 [2.34,2.70]	2.09 [1.91,2.28]	2.35 [2.18,2.55]	1.96 [1.80,2.12]
T3 In-situ tumor (73)	3.58 [2.38,5.01]	0.67 [.25,1.22]	2.05 [1.45,2.84]	1.25 [.85,1.71]
T4 Healthy glands (19)	3.47 [1.84,5.95]	0.84 [.16,1.79]	1.84 [.47,4.05]	1.32 [.32,2.84]
T5 Necrosis (58)	2.84 [1.57,4.91]	0.28 [.14,.45]	3.21 [1.62,5.79]	1.74 [.98,2.91]
T6 Inflamed stroma (624)	6.54 [5.97,7.15]	8.68 [7.94,9.54]	6.54 [5.96,7.16]	6.70 [6.11,7.36]

Appendix

Table 4: Per-ROI MAE by tissue subtype (TCGA-BRCA test; lymphocytes per ROI), point estimate with 95% bootstrap CI [low, high] (

N{=}1000

, ROI-level resampling). Due to wide, overlapping CIs in the small-

n

subtypes (T4

n{=}19

, T5

n{=}58

, T3

n{=}73

), we do not bold a single per-tissue-subtype winner. PanopTILs attains the lowest MAE in sparse-positive epithelial subtypes (T1, T3–T5, T7) by its tendency to suppress positive predictions everywhere, whereas Loki-OT is lowest in T2 and lower than PanopTILs in the lymphocyte-dense T6, consistent with its lower patient-level MAE.

Explore similar work

CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification
Aug 17, 2026Jialu Yao, Songhao Li, Alina Yu +1Histopathology Image Classification Pathology Foundation Models
Department of Pathology and Laboratory Medicine, University of Pennsylvania, Philadelphia, PA, USA · Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA, USA · Germantown Friends School, Philadelphia, PA, USA +1
See, Measure, and Reason: Learning Visually Grounded Reasoning in Pathology
Sep 28, 2026Chengyang Zhang, Wenchuan Zhang, Bo Li +8Computational Pathology Medical VQA
College of Computer Science, Sichuan University · Department of Pathology and Institute of Clinical Pathology, West China Hospital, Sichuan University · Department of Computer Science, School of Computing, National University of Singapore +1
Retrieval-Augmented Vision Foundation Models for Robust Leukemia Cell Classification across Multiple Microscopy Datasets
Aug 11, 2026Carlos Zamora, Hiram Zuniga, Ulises Orozco-Rosas +1Retrieval-Augmented Classification Vision Foundation Model Adaptation
CETYS Universidad, Ave. CETYS Universidad No. 4, Fracc. El Lago, C.P. 22210, Tijuana, Baja California, Mexico

Aug 17, 2026J/K move · Enter open · S save

CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification

Jialu Yao, Songhao Li, Alina Yu, Zhi Huang

Department of Pathology and Laboratory Medicine, University of Pennsylvania, Philadelphia, PA, USA · Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA, USA · Germantown Friends School, Philadelphia, PA, USA+1

Histopathology Image Classification Pathology Foundation Models

Open paper AskPDF arXiv

Identifying cell types directly from routine haematoxylin and eosin (H&E) histology would enable single-cell analysis at scale, but training such models has relied on manual pathologist annotations, which are slow, expensive and unreliable for many cell types. We instead supervise morphology with molecules. Imaging-based spatial transcriptomics profiles individual cells in situ on a section that can afterwards be stained with H&E, so that molecular identity and morphology are observed for the same physical cell. We assembled 81 such paired Xenium sections spanning 16 organs, derived per-cell labels by clustering, marker-gene annotation, organ-wise human review and quality control, and mapped them onto the cell types commonly reported in each organ. This yielded 15.4 million cells, each with a paired H&E image patch and one of 23 cell types, on which we trained CytoFormer, a cell foundation model with a multi-task, per-organ classification head. On spatially held-out tissue CytoFormer reached an accuracy of 0.85 and a macro-F1 of 0.78 across all 16 organs, and its predictions reproduced the tissue architecture of an entire held-out section. The representation also transfers: with the encoder frozen, a linear head on CytoFormer features performed better than six pathology foundation models on four expert-annotated benchmarks, including on organs and cell types that were not part of pretraining. Finally, in an interactive active-learning setting, CytoFormer's embeddings are markedly more label-efficient than existing pathology foundation models, detecting normal epithelium amid look-alike tumour with an F1 of 0.82 from only a few annotations and leading the strongest baseline by 0.13 in F1. CytoFormer turns paired H&E and spatial transcriptomics into a reusable, label-efficient representation for cell-level analysis of routine histology.

Abstract

Explore similar work

CytoFormer: A Molecularly Supervised Cell Foundation Model for Histopathology Cell Classification

See, Measure, and Reason: Learning Visually Grounded Reasoning in Pathology

Retrieval-Augmented Vision Foundation Models for Robust Leukemia Cell Classification across Multiple Microscopy Datasets