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

• INFL and ASC-US can overlap morphologically, while benign ENDO cells can resemble HSIL, so plausible predictions need not be definitive.Source: §1 · p. 2
• Conventional multiscale fusion can suppress disagreement between spatial views, and hard predictions cannot explicitly represent conflicting or insufficient evidence.Source: §1 · p. 2
• The framework uses ordinal soft anchors, independent scale-specific predictions, and Dempster–Shafer fusion to retain uncertainty and route cases for review.Evaluation on RIVA and CRIC Cervix finds greater uncertainty for incorrect predictions and high uncertainty concentrated around equivocal categories.Source: §1 · p. 2

• Studies of isolated-cell cervical cytology differ in diagnostic groupings, datasets, and evaluation protocols, complicating direct comparison of reported results.Source: §2 · p. 2
• Multiscale architectures use complementary morphology, but prior predictions typically improve classification without explicitly characterizing evidence from each spatial view.A review of 77 studies identifies limited generalizability and insufficient representation of real-world cytological complexity as persistent challenges.Source: §2 · p. 2
• DST represents support for competing hypotheses alongside uncommitted evidence, but its use to fuse complementary spatial predictions in cervical cytology remains unexplored.Source: §2 · p. 2
• The work retains separate nuclear, cellular, and neighbourhood evidence until late fusion, allowing agreement, conflict, and insufficient evidence to affect the final decision.Source: §2 · p. 2

3. Method

• Three nucleus-centred crops of 100, 200, and 400 pixels represent nuclear, cytoplasmic, and neighbourhood contexts across both datasets.The 100-pixel crop follows RIVA's annotation convention; the larger crops add context.Source: §3 · p. 2

3.1. Multiscale Feature Extraction and Ordinal Prediction

• Independent scale-specific heads use eight clinically grounded soft anchors to retain ordinal relationships and keep ASC-US between benign and lesion regions.Anchor spacing represents ordinal positions, not raw clinical risk probabilities.Source: §3.1 · p. 3, Table 2 · p. 4

3.2. Calibrated Predictions and Evidential Fusion

• Per-scale calibrated predictions are normalized around dataset-specific decision thresholds before being converted into evidence.Evidence strength increases with distance from the decision boundary; predictions close to it provide little committed evidence.Source: §3.2 · p. 4
• Each scale assigns mass to benign, lesion, or uncertainty; reliability discounting transfers less reliable sources’ residual mass to uncertainty.Source: §3.2 · p. 4
• Sequential Dempster–Shafer fusion combines the three discounted sources, with conflict contributing to K and insufficient evidence remaining in mU.The resulting fused belief state is used by the routing policy.Source: §3.2 · p. 5

3.3. Evidence-Based Routing

• The routing policy treats mU > 0.50 as an uncertain pool, distinguishing scale agreement from disagreement rather than treating uncertainty as a diagnosis.When scale votes disagree but mU is low, fused belief can resolve the disagreement; CRIC additionally uses a dataset-specific NILM vote filter.Source: §3.3 · p. 5

4.1. Datasets

• RIVA covers all eight TBS categories; CRIC Cervix covers six and tests cross-dataset generalization under different scanning, staining, and demographic conditions.RIVA has 10,613/2,654/2,682 train/validation/test cells; CRIC has 8,073/1,730/1,731, with only 91 ASC-US cells in each of its validation and test splits.Source: §4.1 · p. 5

4.2. Metrics

• Macro-F1 is primary; lesion-F1 and routing accuracy assess binary routing, while ROC-AUC, PR-AUC, uncertainty ratio, and confidence-threshold metrics characterize fused output and uncertainty.UR > 1 indicates greater mean uncertainty for incorrect predictions; high-confidence samples have mU ≤0.40, the uncertain pool has mU >0.50, and silent wrong is the fraction of low-uncertainty samples routed to the incorrect binary branch.Source: §4.2 · pp. 5–6

4.3. Implementation Details

• Each scale-specific MLP is trained independently for 20 epochs with AdamW, learning rate 10−3, weight decay 10−4, and a frozen DINOv2 ViT-B/14 feature backbone.The checkpoint is selected by validation macro-F1 computed from ordinally supervised predictions.Source: §4.3 · p. 6

5.1. Classification and Fusion Results

• DST provides the strongest fusion performance on both datasets, with its clearest separation when spatial views disagree or evidence is insufficient.The comparison holds the calibrated predictions and scale-specific classifiers fixed, isolating the fusion strategy.Source: §5.1 · p. 6, Table 5 · p. 7

5.2. Uncertainty and Clinical Ambiguity

• NILM and SCC are routed more reliably with lower uncertainty than categories near the benign–lesion boundary.Fused lesion-transition scores retain ordinal progression across datasets, with NILM and SCC near opposite ends and ASC-US and LSIL in broader transitional regions.Source: §5.2 · p. 6, Figure 2 · p. 7
• Incorrect predictions tend to carry more uncertainty overall on both datasets, but correctly routed ASC-US samples can remain more uncertain than incorrect ones.ENDO is the main failure mode on RIVA, with poor routing accuracy and a high silent-wrong rate.Source: §5.2 · p. 6

5.3. Ablation, Sensitivity, and Cross-Dataset Stability

• The core evidential formulation transfers unchanged, while calibration and scale-specific thresholds adapt to each dataset’s validation distribution.The CRIC NILM filter is an additional refinement for obtaining a purer Genuine Uncertain pool, not a requirement for operation.Source: §5.3 · p. 6
• The ASC-US anchor value 0.60 retains stable behaviour across both datasets; reliability settings alter spatial-source influence, while λ and mhigh_U mainly affect uncertain-pool size and composition.Changing the order of sequential DST combination produces no differences on either dataset, consistent with Dempster’s rule being commutative and associative.Source: §5.3 · p. 7

Uncertainty as a first-class output

• High-confidence samples outperform the overall average, while uncertain samples underperform it, concentrating uncertainty on diagnostically difficult cases.The uncertain pool is identified as a natural target for further review.Source: §6 · p. 7, Table 8 · p. 8

Biology-driven uncertainty structure

• Correctly predicted ASC-US samples remain more uncertain than confidently misclassified ones, consistent with genuine boundary ambiguity.The passage attributes this pattern to the interaction of soft ordinal targets and DST fusion.Source: §6 · pp. 7–8

Representative routing cases

• Figure 3 illustrates five routing categories, including cases where DST resolves disagreement and cases with insufficient evidence.Each row shows a test sample’s scale-specific lesion probabilities and votes, fused evidence, and routing action.Source: Figure 3 · p. 8

Routing and design summaries

• Table 6 reports per-class routing accuracy, silent-wrong rate, mean uncertainty, and the uncertainty ratio between incorrect and correct predictions.Table 7 summarizes the cross-dataset status of principal design choices.Source: Table 6 · p. 8, Table 7 · p. 8

Uncertainty-stratified results

• The uncertainty-stratified results cover 2,682/1,731 total samples, 2,276/1,649 high-confidence samples, and 275/16 uncertain samples for RIVA/CRIC, respectively.The CRIC uncertain pool is notably small relative to its high-confidence pool.Source: §6 · p. 8

7. Conclusion

• Across RIVA and CRIC Cervix, Dempster–Shafer fusion improves over individual scales and alternative fusion strategies, especially under inter-scale conflict.The fused evidence preserves Bethesda-category ordinal progression, while incorrect predictions show substantially higher uncertainty, particularly around ambiguous boundaries.Source: §7 · p. 8

ENDO routing

• RIVA ENDO routing accuracy is 0.222, with a 46.7% silent-wrong rate and UR=0.42, so incorrect predictions carry less uncertainty than correct ones.ENDO’s benign morphology can resemble HSIL, exposing a limitation of the binary evidential frame; a richer hypothesis space is proposed as future work.Source: §8 · pp. 8–9

CRIC uncertain pool

• CRIC’s genuine uncertain pool contains only 16 test samples (0.9%), with limited class representation, making class-level routing statistics directional.The CRIC-specific NILM vote-filter threshold must be determined empirically for new datasets.Source: §8 · p. 9