Source-linked AI summary
Reliability-Aware Hybrid-K Ensemble Selection for Cervical Cytology Classification: Integrating Discrimination, Calibration, and Selective Prediction
Nisreen Albzour, Sarah S. Lam
TL;DR
Clinical cytology classifiers need reliable confidence and uncertainty estimates in addition to high accuracy. This study evaluates a reliability-aware Hybrid-K ensemble framework on SIPaKMeD, finding that a compact Hybrid-2 ensemble improved reliability point estimates and remained robust to alternative metric weightings under internal validation.
Problem
High classification accuracy alone is insufficient for clinical image analysis, where calibrated confidence and reliable uncertainty estimates are essential.
Method
Nine SIPaKMeD architectures were temperature-scaled, evaluated with an eight-metric composite score, and used to form soft-voting Hybrid-K ensembles with robustness analyses.
Results
Hybrid-2 reduced AURC by 43%, NLL by 17%, and worst-class expected calibration error by 36% relative to the best individual model, while remaining the modal ensemble in 96.8% of sampled weightings.
Takeaways & Limitations
Hybrid-2 provided a compact ensemble choice that was robust to alternative metric weightings and improved reliability point estimates under internal validation on SIPaKMeD.
Takeaways & Limitations
Calibration-dependent results are exploratory internal estimates because calibration used held-out evaluations rather than a fully independent calibration set.
Abstract
from arXiv · showhide
High classification accuracy alone is insufficient for clinical image analysis, where calibrated confidence and reliable uncertainty estimates are essential. This study proposes a reliability-aware Hybrid-K ensemble selection framework for multiclass cervical cytology classification using the SIPaKMeD dataset. Nine deep learning architectures were evaluated using a fixed stratified five-fold partition and three training seeds. After post-hoc temperature scaling, models were assessed using macro-F1, accuracy, AUROC, expected calibration error (ECE), worst-class ECE (WC-ECE), area under the risk-coverage curve (AURC), Brier score, and negative log-likelihood (NLL). Models were ranked using an equal-weight composite score, and Hybrid-K ensembles were formed from the top-ranked models using soft voting. Robustness was examined using 5,000 Dirichlet-sampled metric-weight vectors, leave-one-metric-out analysis, and corrected paired testing across 15 fold-by-seed evaluations. The final Hybrid-2 ensemble, comprising Swin-Tiny and TinyViT-5M, reduced AURC by 43%, NLL by 17%, and WC-ECE by 36% relative to the best individual model. It was selected in 96.8% of random weighting scenarios, remained unchanged across all leave-one-metric-out analyses, and improved the full composite score. However, per-metric gains were not statistically significant after Holm-Bonferroni correction (all adjusted p >= 0.168). Because post-hoc calibration did not use a fully independent calibration set, calibration-dependent results should be interpreted as exploratory internal estimates. Overall, the framework identified a compact ensemble robust to alternative metric weightings and improved reliability point estimates under internal validation on a single dataset.
1. Introduction
Cervical cytology classification has achieved strong benchmark performance, but clinical use also requires calibrated confidence and reliable uncertainty estimates. This study proposes a reliability-aware framework that jointly evaluates discrimination, calibration, selective prediction, and ensemble selection.
- Manual cytological examination is labor-intensive, subject to inter-observer variability, and difficult to scale where trained cytopathologists are scarce.
- Existing cervical cytology studies typically select models using discrimination metrics such as accuracy and macro-F1, with less emphasis on reliability.
- Calibration and selective prediction provide complementary reliability dimensions for evaluating whether confidence estimates and abstention behavior are clinically useful.
- Most existing ensemble strategies optimize a single criterion and do not jointly account for classification performance, calibration error, selective prediction, and probabilistic scoring.
- The proposed framework trains nine architectures, ranks them with an eight-metric composite score, forms soft-voting Hybrid-K ensembles, and validates selection robustness through three complementary analyses.
- The study contributes a reliability-aware ensemble selection framework because joint optimization of classification performance and reliability remains scarce in cervical cytology classification.
2. Related Work
Prior work spans cervical cytology classifiers, medical-image ensembles, calibration, selective prediction, domain generalization, and multicriteria selection. The reviewed literature motivates combining these strands because reliability and external validity remain incompletely addressed by accuracy-focused studies.
- Deep Learning for Cervical Cytology Classification: Deep learning has become dominant in automated cervical cell classification, while prior work includes classical feature-based pipelines and newer CNN, transformer, and hybrid architectures.
- Ensemble Methods in Medical Image Classification: Medical-image ensembles have primarily targeted accuracy, and prior studies rarely examine how ensemble size and composition trade off against calibration or selective-prediction quality.
- Calibration and Selective Prediction: Confidence calibration aligns predicted probabilities with empirical correctness frequencies, while temperature scaling rescales logits without changing predicted class labels.
- Calibration and Selective Prediction: Selective prediction allows uncertain cases to be abstained from and referred for review, with AURC summarizing risk–coverage behavior.
- Generalization Under Domain Shift: Performance and calibration on a training distribution do not guarantee reliability under different scanners, acquisition protocols, or patient populations.
- Multi-Criteria Model Selection: SMAA evaluates model acceptability by integrating over weight vectors on the simplex when metric preferences are uncertain.
- Positioning Relative to Prior Work: Closest prior work used explainability-weighted fuzzy ensembles and found calibration improvements after temperature scaling, but no consistent reliability benefit from larger ensembles.
3. Methodology
The methodology evaluates nine cervical cytology classifiers on SIPaKMeD using standardized cross-validation, calibration, reliability-aware scoring, and Hybrid-K soft-voting ensembles. Robustness analyses test whether ensemble selection depends on metric weighting or statistical comparison choices.
- Framework overview: The five-stage framework prepares the dataset, trains nine architectures, calibrates and evaluates them, ranks candidates, constructs Hybrid-K ensembles, and conducts final selection.The pipeline combines training, eight-metric evaluation, composite ranking, soft voting, and internal selection.
- Dataset and cross-validation: SIPaKMeD provides 4,049 expert-annotated single-cell images across five near-balanced cytomorphological classes, evaluated with stratified five-fold cross-validation.Class counts range from 787 to 831 images, limiting the influence of class imbalance on evaluation.
- Candidate models: Nine candidates span conventional CNNs, vision transformers, and lightweight hybrid CNN–transformers, including Swin-Tiny and TinyViT-5M.The shared model pool ranges from VGG16, ResNet50, DenseNet121, and EfficientNet-B0 to transformer and hybrid architectures.
- Evaluation and calibration: Models are assessed with macro-F1, accuracy, AUROC, ECE, WC-ECE, AURC, Brier score, and NLL after post-hoc temperature scaling.The calibration procedure uses leave-one-fold-out fitting on stored held-out predictions and does not change predicted class labels.
- Ranking and ensemble construction: Candidates are ranked by an equal-weight composite of normalized metrics, then Hybrid-2, Hybrid-3, and Hybrid-4 ensembles are formed from the highest-ranked models using soft voting.Hybrid-2 combines the two highest-ranked models; the selected internal pair is Swin-Tiny and TinyViT-5M.
4. Results and Discussion
The results compare nine candidate architectures under a fixed, repeated evaluation design and rank them using an equal-weight composite of classification and reliability metrics. Swin-Tiny and TinyViT-5M led the reliability-aware selection, while reliability measures separated models whose discrimination was already similar.
- Experimental design: 15 fold-by-seed evaluations used the same fixed stratified five-fold partition and three training seeds for every candidate and ensemble configuration.This enabled paired comparisons across architectures and ensemble configurations.
- Candidate ranking: Swin-Tiny achieved the highest composite score at 0.939, followed by TinyViT-5M at 0.876, ViT-Tiny at 0.787, and DenseNet121 at 0.781.Swin-Tiny and TinyViT-5M were selected for Hybrid-2.
- Candidate ranking: Macro-F1 ranged from 0.912 to 0.982 and AUROC from 0.991 to 0.999, leaving reliability metrics to separate models with similar discrimination.The ranking was driven by Brier score, NLL, AURC, and ECE.
- Candidate ranking: EfficientNet-B0 ranked last with a composite score of 0.138 despite strong discrimination, reflecting Brier score 0.129 and NLL 0.255.Its weaker probabilistic reliability, rather than a discrimination collapse, drove the low score.
4.3 Calibration and Reliability Results
Hybrid-K ensembles produced their clearest gains in reliability and selective prediction rather than discrimination, which was already near ceiling. Hybrid-2 improved several reliability point estimates relative to Swin-Tiny, subject to the internal calibration caveat.
- Calibration: Hybrid-2 reduced WC-ECE from 0.0290 to 0.0186, a 36% reduction, while ECE remained similar at 0.0039 versus Swin-Tiny's 0.0040.Overall calibration was already low across the candidate pool, making worst-class calibration more discriminating.
- Caveat: Calibration-dependent figures are internal exploratory estimates because temperature fitting used held-out predictions that were not fully independent of the evaluated fold.The limitation affects interpretation of calibration performance rather than the reported point estimates.
- Selective prediction: Hybrid-2 reduced AURC from 0.0028 to 0.0016, a 43% reduction relative to Swin-Tiny.The reduction reflects lower classification risk as uncertain predictions are removed.
- Probabilistic reliability: Hybrid-2 reduced NLL from 0.0679 to 0.0563, a 17% reduction, and Brier score from 0.0292 to 0.0258.These gains were concentrated in probabilistic scoring.
- Confusion patterns: Hybrid-2 improved Koilocytotic recall from 94.8% to 95.5% and reduced most off-diagonal errors relative to Swin-Tiny.Additional gains tapered for Hybrid-3 and Hybrid-4.
- Discrimination versus reliability: AUROC remained highly similar across configurations at 0.9988–0.9993, while meaningful gains concentrated in AURC, Brier score, and NLL.Classification metrics were near ceiling for all configurations.
4.6 Robustness Under Random Metric Weightings
The random-weighting analysis found Hybrid-2 to be the preferred ensemble under alternative metric weightings. Its selection was highly stable across the 5,000-sample robustness analysis.
- Random weighting: The robustness analysis evaluated modal Hybrid-K selection across nested Dirichlet sample sizes from 500 to 10,000 draws.The pre-specified main robustness sample size was N = 5,000.
- Interpretation: Hybrid-2 was the most stable and robust ensemble configuration under alternative metric-weighting assumptions.The analysis indicates that the preferred two-model configuration was not dependent on one fixed weighting choice.
- Random weighting: Hybrid-2 was selected in approximately 96.8% of 5,000 Dirichlet-sampled metric-weighting scenarios.The best individual model was selected in approximately 93.7% of scenarios.
4.7 Leave-One-Metric-Out Sensitivity Results
Leave-one-metric-out analysis preserved Hybrid-2 membership across all eight metric exclusions. Larger ensembles were more sensitive to metric composition, supporting the stability of the two-model configuration.
- Larger-ensemble sensitivity: Removing AURC or NLL did not change the reported Hybrid-K memberships in the leave-one-metric-out table.The table marks both exclusions as producing no change.
- Hybrid-2 stability: Hybrid-2 remained unchanged after each of the eight metrics was removed, consistently selecting Swin-Tiny and TinyViT-5M.The full eight-metric baseline was recomputed using the remaining seven metrics with equal weights.
- Larger-ensemble sensitivity: Removing macro-F1, accuracy, WC-ECE, or Brier score changed the membership of Hybrid-3 and Hybrid-4.Hybrid-3 swapped ViT-Tiny for DenseNet121, with a corresponding shift in Hybrid-4.
- Interpretation: The final Hybrid-2 recommendation was not driven by any single evaluation metric, whereas Hybrid-3 and Hybrid-4 were more sensitive to metric composition.This supports the conclusion that the two-model configuration was more stable than the larger ensembles.
4.8 Statistical Comparison Results
Hybrid-2 improved most point estimates relative to Swin-Tiny, but corrected repeated cross-validation testing found no statistically significant per-metric differences after Holm-Bonferroni correction. The evidence therefore supports consistent directional gains while requiring cautious interpretation of inferential strength.
- Corrected testing: No comparison reached significance at p < 0.05 after Holm-Bonferroni correction across the eight metrics.The analysis used corrected repeated cross-validation paired t-tests across 15 fold-by-seed evaluations.
- Point estimates: AURC decreased from 0.0028 to 0.0016, Brier from 0.0292 to 0.0258, and NLL from 0.0679 to 0.0563 for Hybrid-2 versus Swin-Tiny.Macro-F1 also increased from 0.9824 to 0.9847.
- Corrected testing: All Holm-adjusted p-values were ≥ 0.168, so the corrected analysis provided no statistical evidence that Hybrid-2 outperformed Swin-Tiny on any individual metric.The improvements were consistent in direction but remained within variability across correlated fold-by-seed evaluations.
- Effect sizes: Most observed improvements had medium-to-large descriptive effect sizes, including |d_z| = 1.46 for NLL and |d_z| = 1.07 for Brier score.These magnitudes are exploratory summaries because the fold-by-seed differences are not independent observations.
- Ensemble size: Hybrid-2 was favored over Hybrid-3 and Hybrid-4 because larger ensembles showed no statistically detectable metric improvement after correction.The smallest adjusted p-value in these comparisons was approximately 0.4305 for NLL in Hybrid-2 versus Hybrid-4.
4.9 Final Ensemble Selection
Hybrid-2 was selected as the final ensemble because it combined improved reliability point estimates with a higher composite score and strong stability under alternative selection analyses. The two-model configuration was preferred over larger ensembles on parsimony grounds.
- Final selection: Hybrid-2 reduced AURC by approximately 43%, NLL by approximately 17%, and worst-class ECE by approximately 36% relative to Swin-Tiny.It also improved the full eight-metric composite score.
- Robustness: Hybrid-2 was the modal ensemble in 96.8% of random metric-weighting scenarios and remained unchanged across all eight leave-one-metric-out ablations.Its selection therefore remained stable under alternative metric weightings and omissions.
- Parsimony: Hybrid-2, composed of Swin-Tiny and TinyViT-5M, was selected because Hybrid-3 and Hybrid-4 were less stable and added no significant corrected-test improvement.The final choice rests on consistent directional gains, robustness, selection stability, and parsimony.
4.10 Comparison with Representative SIPaKMeD Cervical Cytology Studies
The representative SIPaKMeD comparison shows that prior studies generally emphasize discrimination metrics, whereas the proposed framework additionally evaluates calibration, selective prediction, statistical testing, and robustness. Cross-study metric values are not directly comparable because experimental protocols differ.
- Comparison scope: Reported SIPaKMeD studies differ in train–test splits, preprocessing, augmentation, training protocols, and completeness of reliability reporting.The comparison is representative rather than a controlled head-to-head benchmark.
- Reliability reporting: Calibration and selective prediction were not reported for several representative studies, whereas the proposed framework reports ECE, WC-ECE, Brier, NLL, and AURC.The table marks reliability-aware selection as absent for the listed comparator studies and present for the proposed framework.
- Evaluation coverage: Prior studies primarily emphasize accuracy, F1, or AUROC, while the proposed framework jointly evaluates discrimination, calibration, probabilistic reliability, selective prediction, significance, and metric-weighting robustness.The framework uses an eight-metric composite, corrected repeated cross-validation testing, Dirichlet robustness, and leave-one-metric-out sensitivity.
Proposed study 2026
The proposed study reports strong SIPaKMeD discrimination results while framing them within a broader reliability-aware evaluation. Its reported metrics should be interpreted alongside protocol differences across studies.
- Proposed study: The proposed study reports accuracy 0.9846, macro-F1 0.9847, and AUROC 0.9992 for SIPaKMeD cervical cytology classification.These values are reported for the proposed framework in the representative study comparison.
- Interpretation: Reported values are not directly comparable across studies because splits, preprocessing, augmentation, and training protocols differ.The table presents values under each cited work’s own experimental conditions.
- Evaluation scope: The comparison includes calibration, selective prediction, statistical testing, and reliability-aware selection rather than discrimination metrics alone.The broader evaluation includes ECE, WC-ECE, Brier, NLL, AURC, corrected repeated cross-validation testing, and robustness analyses.
5. Conclusion
The reliability-aware Hybrid-2 ensemble improved several reliability point estimates over the best individual model and was robustly selected under alternative metric weightings. However, calibration-dependent findings remain exploratory because validation used a single dataset and lacked a fully independent calibration set.
- The Hybrid-2 ensemble combined Swin-Tiny and TinyViT-5M via soft voting, while matching or exceeding the best individual model’s classification performance.
- AURC fell by 43%, NLL by 17%, and worst-class ECE by 36% relative to the best individual model.
- Hybrid-2 was selected in 96.8% of 5,000 randomly sampled metric-weight scenarios and retained the same membership across all eight leave-one-metric-out analyses.
- Hybrid-3 and Hybrid-4 produced only marginally larger composite-score gains and were less stable, supporting Hybrid-2 as a parsimonious choice.
- The equal-weight composite score and imperfect reliability metrics limit interpretation of the ranking as a definitive measure of clinical trustworthiness.
- Reliability gains from internal cross-validation should not be assumed to transfer to new domains without direct external assessment.
- Future work should use external datasets, group-aware partitioning by parent image, and clinically informed or learned metric weightings.
Declarations
The study reports no external funding, uses a publicly available de-identified dataset without human participant intervention, and declares no conflicts of interest.
- The research received no external funding.
- Institutional review board approval was not applicable because the study used a publicly available, de-identified dataset without direct recruitment or intervention.
- Informed consent was not applicable because the study involved no direct interaction with human participants.
- The SIPaKMeD dataset is publicly available from its original source.
- The authors declare no conflicts of interest.