Learning to Reason Like a Cytopathologist:Clinically Grounded Multiscale Evidential Learning for Cervical Cytology

Ajeet Kumar Singh; Nehal Mishra; Anurag Tripathi; Santanu Chaudhury; Dorien Herremans

cs.CVOct 21, 2026

Learning to Reason Like a Cytopathologist:Clinically Grounded Multiscale Evidential Learning for Cervical Cytology

Authors: Ajeet Kumar Singh, Nehal Mishra, Anurag Tripathi, Santanu Chaudhury, Dorien Herremans

Organizations: IIT Delhi, India · Info Origin INC. · Singapore University of Technology and Design

Abstract

Cervical cytology presents a fundamentally ambiguous diagnostic landscape, where equivocal categories such as Atypical Squamous Cells of Undetermined Significance (ASC-US) occupy a morphological grey zone that is definitional. For clinical decision support, the goal extends beyond assigning a definitive label to every sample: cases must also be distinguished by whether the available evidence supports a confident decision or warrants further consideration, a distinction conventional classification models are not designed to explicitly represent. We propose a multiscale evidential framework that models cervical cytology along an ordinal disease continuum using clinically grounded soft anchors, representing progression from benign reactive changes to precancerous lesions and malignancy. Three biologically motivated spatial contexts--nuclear morphology, cytoplasmic characteristics, and neighbourhood cellular architecture--are independently encoded using a shared frozen DINOv2 backbone, preserving inter-scale disagreement that early fusion may obscure. Calibrated predictions are transformed into Dempster--Shafer belief masses and fused with reliability discounting, allowing conflicting evidence and insufficient support to remain explicitly uncommitted rather than forcing a decision. The fused belief state drives interpretable routing that separates confidently classified cases from those with irreducible evidential insufficiency, enabling clinically meaningful triage. Across two independent publicly available conventional cytology datasets, RIVA and CRIC Cervix, the framework generalises consistently, with uncertainty strongly associated with diagnostic difficulty: incorrect predictions carry 1.94x-4.13x higher uncertainty than correct ones.

Figures & tables

Class	Clinical position	Interpretation
NILM	Benign	Definitive
ENDO	Benign	Definitive
INFL	Benign	Definitive
ASC-US	Lesion	Equivocal
LSIL	Lesion	Definitive
ASC-H	Lesion	Equivocal

Table 1: The eight TBS diagnostic categories considered in this work.

Figure 1: Overview of the proposed framework. Multiscale patches capturing nuclear, cytoplasmic, and neighbourhood context are processed through a shared frozen DINOv2 backbone and independent scale-specific heads. Calibrated predictions are transformed into Dempster–Shafer evidence and fused with reliability discounting to support explicit uncertainty for ambiguity-aware routing.

Class	Clinical position	CIN3+ risk	$\mathbf{y_{c}}$
NILM	Clearly benign	Negligible	0.00
ENDO	Benign; mimics HSIL	Negligible	0.05
INFL	Reactive; overlaps ASC-US	Very low	0.25
ASC-US	Definitionally equivocal	Low	0.60
LSIL	Low-grade lesion	Moderate	0.85
ASC-H	Cannot exclude HSIL	High	0.90

Table 2: Clinically grounded, non-uniform ordinal soft anchors for the eight TBS categories. The ordering and relative spacing encode clinically motivated proximity along the benign-to-malignant continuum; anchor values are ordinal positions rather than raw clinical risk probabilities.

Category	Action	Condition
Confident Benign	Benign sample	All 3 votes benign, $m_{U}\leq 0.40$
Confident Lesion	Lesion sample	All 3 votes lesion, $m_{U}\leq 0.40$
Genuine Uncertain	Further review	Split vote, $m_{U}>0.50$
Ambiguous (agree)	Further review	Unanimous vote, $m_{U}>0.40$
Disagree (confident)	DST-resolved	Split vote, $m_{U}\leq 0.50$

Table 3: Five-way evidence-based routing.

m_{U}

denotes the fused uncertainty mass.

Metric	RIVA	CRIC	Metric	RIVA	CRIC
Accuracy	0.750	0.927	Benign-F1	0.808	0.939
Macro-F1	0.726	0.924	Lesion-F1	0.644	0.909
Macro-Precision	0.721	0.927	ROC-AUC	0.814	0.971
Macro-Recall	0.736	0.921	PR-AUC	0.709	0.969

Table 4: Overall benign–lesion routing performance using fused DST evidence. PR-AUC is reported due to the naturally imbalanced screening distribution.

		RIVA				CRIC Cervix
		Overall		Conflict		Overall		Conflict
	Method	Acc.	MF1	Acc.	MF1	Acc.	MF1	Acc.	MF1
scales	100px (nucleus)	.7293	.7112	.5476	.5467	.9012	.8982	.5510	.5510
	200px (cytoplasm)	.7163	.6947	.5153	.5120	.9185	.9159	.6531	.6529
	400px (neighbourhood)	.7260	.6895	.5393	.5022	.9006	.8974	.5476	.5475
fusion	Majority vote	.7210	.7071	.5822	.5815	.9053	.9031	.7109	.7125

Table 5: Comparison of individual scales and late-fusion strategies using the same calibrated scale-specific predictions. Accuracy and Macro-F1 are reported separately. Conflict results are computed only on samples for which the scale predictions disagree.

	RIVA				CRIC Cervix
Class	Acc.	SW $\downarrow$	$\boldsymbol{\mu_{U}}$	$\boldsymbol{\mathrm{UR}\uparrow}$	Acc.	SW $\downarrow$	$\boldsymbol{\mu_{U}}$	$\boldsymbol{\mathrm{UR}\uparrow}$
NILM	.913	1.4	.091	4.13	.943	1.7	.074	5.11
ENDO	.222	46.7	.101	.42	–	–	–	–
INFL	.742	5.5	.218	1.69	–	–	–	–
ASC-US	.405	40.5	.165	.57	.637	9.9	.215	1.66
LSIL	.519	12.6	.278	1.10	.863	4.4	.138	1.05

Table 6: Per-class routing behaviour across datasets. Acc.: routing accuracy; SW: silent wrong (%);

\mu_{U}

: mean uncertainty mass; UR: uncertainty ratio between incorrect and correct predictions.

Panel of Figure 2 — Figure 2: Per-class distributions of fused lesion-transition scores across the RIVA and CRIC Cervix test sets. The distributions retain the ordinal progression from benign to malignant categories after multiscale evidence fusion.

Panel of Figure 3 — Figure 3: Representative multiscale evidential routing cases from the RIVA and CRIC Cervix test sets. Each row shows an actual test sample, its scale-specific lesion probabilities and votes, the DST-fused evidence, and the resulting routing action. The examples collectively illustrate all five routing categories: confident benign, confident lesion, genuine uncertain, ambiguous agreement, and disagreement resolved by DST.

Component	Setting	Status
Multiscale crops	100/200/400 px	Shared
Ordinal structure	ASC-US $=0.60$	Shared
Binary evidential frame	$\Theta=\{B,L\}$	Shared
Reliability discounting	$\alpha\in[0.40,1.00]$	Adapted
DST fusion	Sequential	Shared
BPA sharpness	$\lambda=3$	Shared

Table 7: Cross-dataset status of principal design choices.

Stratum	RIVA		CRIC
	Macro-F1	Acc.	Macro-F1	Acc.
All	0.7259	0.7502	0.9238	0.9266
High confidence	0.7568	0.7830	0.9403	0.9424
Uncertain	0.5250	0.5345	0.5897	0.7500

Table 8: Uncertainty-stratified routing performance. High-confidence predictions (

m_{U}\leq 0.40

) substantially outperform uncertain predictions (

m_{U}>0.50

), indicating that DST uncertainty mass is associated with prediction reliability.

Computational Pathology

Computational pathology uses algorithms to analyze digitized tissue images for diagnosis, prognosis, biomarker discovery, and clinical decision support.

See all 277 papers

Refining Cytology Predictions with Conditional Random Fields
Sep 25, 2026Manon Dausort, Tiffanie Godelaine, Karim El Khoury +3Medical Image Classification Conditional Random Fields
ICTEAM, Universit´e Catholique de Louvain, Belgium
Reliability-Aware Hybrid-K Ensemble Selection for Cervical Cytology Classification: Integrating Discrimination, Calibration, and Selective Prediction
Sep 1, 2026Nisreen Albzour, Sarah S. LamComputational Pathology Ensemble Learning
School of Systems Science and Industrial Engineering, Binghamton University, Binghamton, NY 13902, USA
Geometry-aware Gaussian Prior and Axial Attention for Cervical Cytology Image Classification
Jul 11, 2026Yating Li, Cheng Ye, Nenan Lyu +2Computational Pathology Medical Image Classification
School of Cyber Science and Technology, University of Science and Technology of China, Hefei 230026, China · School of Information Science and Technology, University of Science and Technology of China, Hefei 230026, China · Department of Gynecologic Oncology, National Cancer Center, National Clinical Research Center for Cancer, Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100730, China

Sep 25, 2026J/K move · Enter open · S save

Refining Cytology Predictions with Conditional Random Fields

Manon Dausort, Tiffanie Godelaine, Karim El Khoury, Maxime Zanella+2

ICTEAM, Universit´e Catholique de Louvain, Belgium

Medical Image Classification Conditional Random Fields Medical VLMs Zero-Shot Image Classification

Open paper AskPDF arXiv

Vision-language models (VLMs) achieve strong zero-shot (ZS) classification on histology images but do not perform as well on cytology, whose stains and cell morphology differ markedly compared to histology. Conditional random fields (CRFs) can refine noisy VLM predictions by propagating information across patches, but existing CRF frameworks were designed for histopathology and do not transfer to cytology datasets, released as independent patch pools spanning multiple staining protocols. We introduce CytoCRF, which adapts the pairwise terms to cytology by targeting chromatin and cytology-specific staining, and further enrich the neighborhood of each potential term by combining multiple backbones. Across ten cytology datasets, CytoCRF outperforms existing CRF frameworks at every annotation budget, reaching +13.6 percentage points over the best baseline and +33.7 over ZS with only 50 annotations. Combining information from multiple backbones brings further gains, showing that the neighborhood topology matters more than the pairwise potential computed over it.

Abstract

Computational Pathology

Explore similar work

Refining Cytology Predictions with Conditional Random Fields

Reliability-Aware Hybrid-K Ensemble Selection for Cervical Cytology Classification: Integrating Discrimination, Calibration, and Selective Prediction

Geometry-aware Gaussian Prior and Axial Attention for Cervical Cytology Image Classification