acceptodds
Under review as a conference paper at ICLR 2027

UA-CP: Uncertainty-Aware Conformal Prediction for Reliable Endoscopic Lesion Detection

Abstract

Reliable AI-assisted endoscopic lesion detection requires both accurate predictions and explicit uncertainty characterization, yet subtle appearances, ambiguous boundaries, and imbalanced category distributions challenge classification and localization. We propose UA-CP, a two-stage framework combining detector-adaptive Prior-Aware Training (PAT) with Dual Conformal Calibration (DCC). PAT incorporates category-frequency priors into classification objectives for proposal-based and dense detectors or assignment objectives for query-based detectors. DCC then constructs conformal category prediction sets using RAPS and simultaneous localization regions using conformalized quantile regression. Under exchangeability, these outputs provide finite-sample marginal coverage guarantees, with a joint lower bound obtained by the union bound. Across three complementary endoscopic benchmarks and four detector configurations, PAT generally improves detection performance, while DCC characterizes architecture-dependent trade-offs between uncertainty validity and efficiency. Results demonstrate that UA-CP generally improves detection utility while providing valid and informative category and localization uncertainty across diverse endoscopic detection settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Endoscopic lesion detection, Conformal prediction, Uncertainty quantification, Reliable AI

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5vH2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

UA-CP targets reliable endoscopic lesion detection with two stages. (1) Prior-Aware Training (PAT) adds a 1/nk1/\sqrt{n_k}-weighted, error-adaptive group-norm penalty to the classifier of Faster R-CNN/FCOS, or a prior-weighted term to DINO's Hungarian matching cost, to address class imbalance. (2) Dual Conformal Calibration (DCC) applies RAPS to class probabilities and Bonferroni-corrected CQR to box coordinates of IoU-matched detections, giving marginal label and box coverage and a union-bound joint guarantee (≥1−αL−αB\ge 1-\alpha_L-\alpha_B). Validation uses LGLDD (4 classes), Kvasir-Capsule (9 classes) and LDPolypVideo (1 class) with four detector configurations. The paper reports detection utility (AP, recall, F1, FP/image), coverage (label/box/joint), efficiency (set size, MPIW), coverage-level and RAPS-hyperparameter sweeps, and extensive qualitative examples.

Critical strengths
  1. Relevant problem and broad evaluation. Calibrated category and localization uncertainty for endoscopic detection is clinically meaningful. The paper evaluates three datasets and three detector paradigms, with calibration-replicate variance and coverage/hyperparameter sweeps. This breadth is the main point in the paper's favor, but it does not offset the issues below.
Critical weaknesses
  1. Reported coverage contradicts the claimed finite-sample guarantees.

    • At αB=0.05\alpha_B=0.05 on LGLDD (Tab. 2), DCC box coverage is 0.82 (R50), 0.85 (DINO) and 0.90 (FCOS). Joint coverage of 0.80 is below the 0.90 union bound.
    • On LDPolypVideo (Tab. 13), three of four detectors give 0.84–0.88. Figs. 6–7 show ~0.6 box coverage at nominal 0.90–0.95 (e.g., UA Faster R-CNN-Rxt101).
    • With std ≈0.01 over calibration replicates, these gaps are not finite-sample noise. Either exchangeability is violated (e.g., frame-level splits of videos, multiple correlated detections per image) or the implementation differs from the described method. Alg. 3 uses a max-over-coordinates score at level αB\alpha_B, whereas Eq. 15 uses per-coordinate αB/4\alpha_B/4.
    • The abstract nonetheless claims "valid" uncertainty, and validity is the central claim of a conformal method.
  2. Limited novelty, with results that are hard to trust.

    • DCC is an off-the-shelf combination of RAPS, CQR and Bonferroni/union bounds. PAT is standard inverse-sqrt-frequency reweighting, with no comparison to long-tailed detection methods (EQL v2, Seesaw, RFS, logit adjustment).
    • The two stages are not coupled: PAT does not target conformal efficiency.
    • Table 3's FCOS rows are identical to its Rxt101 rows (103.93±2.34, 84.14±2.80). Table 2's Rxt101 DCC row (0.95/0.94/30.58) disagrees with Table 3 (0.90/0.88/84.14).
    • GFLOPs change under PAT (Tab. 9: FCOS 188→137, DINO 137→197), although PAT should not alter inference.
    • App. A.2 theory is incorrect. Thm A.5 treats the ww-dependent coefficient as constant. Thm A.6 claims −γ×-\gamma\timesconvex stays convex.
Questions for authors

Clarifications (could change my assessment):

  1. Why does DCC box/joint coverage fall well below the nominal levels (Tab. 2, 13; Figs. 6–7)? Are train/calibration/test splits made at the patient/video level, or at the frame level?
  2. Which box calibration produced the results: Eq. 15 (per-coordinate αB/4\alpha_B/4) or Alg. 3 (max score at αB\alpha_B)? What is τmatch\tau_{match}?
  3. Is Eq. 12 marginal or class-conditional calibration? What is the per-class label coverage for rare classes (stm, can)?
  4. Please explain the duplicated FCOS/Rxt101 rows in Tab. 3, the Tab. 2 vs Tab. 3 mismatch, and the GFLOP changes in Tab. 9.

Suggestions: 5. Compare PAT with EQL v2/Seesaw/logit adjustment under the same detectors. Report an end-to-end coverage metric that counts missed lesions, since the current guarantee is conditional on GT-matched detections.

1 comment
  1. Coverage Validity

The discrepancy between theoretical and empirical coverage can arise from finite-sample variability or departures from exchangeability. Angelopoulos and Bates [1] explain that conformal prediction guarantees marginal coverage rather than coverage on every calibration/test split, while Barber et al. [2] examine the effects of exchangeability violations. In medical applications, Yasar et al. [3] report empirical coverage of 0.85 against a theoretical target of 0.90 under cross-cohort evaluation. These findings provide theoretical and empirical warrants for the plausibility of UA-CP's below-target LGLDD joint coverage of 0.80 for Faster R-CNN–R50 and 0.84 for DINO. The train–calibration–test splits were performed at the frame level, as described in Section 5.1 and Appendix C.2.

  1. Calibration specification

Box calibration. The reported results use the coordinate-wise calibration specified in Eq. 15. Algorithm 3 also defines coordinate-wise CQR nonconformity scores and maintains separate residual collections for the four bounding-box coordinates. For each coordinate, calibration computes a finite-sample corrected quantile at level 1−αB/41-\alpha_B/4. As established in Proposition A.3, the union bound guarantees simultaneous bounding-box coverage of at least 1−αB1-\alpha_B under the stated exchangeability assumptions. The matching threshold is τmatch=0.5\tau_{\mathrm{match}}=0.5; intermediate detector predictions are matched one-to-one with ground-truth lesions, retaining pairs with IoU≥0.5\mathrm{IoU}\geq0.5, consistent with the IoU threshold used in AP50 evaluation.

Category calibration. The reported results use marginal RAPS calibration with a single conformal threshold computed from all matched calibration detections, as specified in Algorithm 1 and Proposition A.1. We will clarify this distinction to ensure consistency between the formulation and implementation. Under the LGLDD setting (kreg=2,λ=0.01k_{\mathrm{reg}}=2,\lambda=0.01), the UA detectors achieve empirical matched-lesion label coverage of 94.34%/96.95% (R50), 90.70%/98.05% (RxT101), 96.00%/90.86% (DINO), and 93.35%/94.59% (FCOS) for stm/can, respectively, averaged over ten calibration runs.

  1. Result Provenance

The inconsistencies between Tables 2 and 3 are reporting errors: the RxT101-UA uncertainty row in Table 3 mistakenly duplicates the FCOS values, whereas the FCOS row is correct. For LGLDD, the corrected Faster R-CNN–RxT101-UA results achieve 0.98 label coverage, 0.95 box coverage, and 0.94 joint coverage, exceeding the nominal target of 0.90. The reporting errors will be corrected in the revised manuscript.

Regarding Table 9, the reported GFLOPs measure detector forward-pass computation at fixed input resolutions. PAT modifies the training objective, whereas DCC performs conformal calibration and post-processing based on intermediate detector predictions. These operations do not alter detector-forward GFLOPs when the inference architecture remains unchanged.

References

[1] Angelopoulos and Bates. "Conformal Prediction: A Gentle Introduction." Foundations and Trends in Machine Learning, 2023. doi:10.1561/2200000101

[2] Barber R F, Candes E J, Ramdas A, et al. Conformal prediction beyond exchangeability[J]. The Annals of Statistics, 2023, 51(2): 816-845.

[3] Yasar S, Yagin F H, Alzakari S A, et al. A Generalizable and Interpretable Framework for Molecular Subtype Classification of Pancreatic Ductal Adenocarcinoma Integrating Conformal Uncertainty Quantification and Consensus-Based Explainable Artificial Intelligence Across Multiple Cohorts[J]. International Journal of Molecular Sciences, 2026, 27(15): 6989.

Reviewer_hFdGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies class imbalance and predictive uncertainty in endoscopic lesion detection. Its two-stage framework combines PAT, which modifies detector training losses or matching costs using category frequencies and prediction confidence, with DCC, which applies RAPS to category prediction and CQR to localization. Experiments cover three endoscopic datasets and four detector configurations, reporting detection performance, coverage, and uncertainty-set efficiency. The theoretical coverage statements assume exchangeability of matched calibration and test detections.

Critical strengths

The application is meaningful: lesion detection should be assessed beyond AP, including missed lesions, false alerts, and uncertainty in category and localization. Reporting recall, FP/image, and coverage together across different detection paradigms is a sensible evaluation direction.

Critical weaknesses

1. The central reliability claim is not supported by the current theory and experiments.

My main concern is that the paper repeatedly presents its uncertainty outputs as valid and guaranteed, but the reported evidence does not establish this. The primary setting targets 95% localization coverage and at least 90% joint coverage. In Table 2, DCC achieves only 82% and 80% for R50, and 85% and 84% for DINO. For R50, Box-Std achieves 97% localization coverage; DCC reduces this to 82% while changing MPIW only from 46.22 to 45.90. This cannot support an efficiency improvement at comparable coverage. These substantial shortfalls need to be diagnosed rather than absorbed into an architecture-dependent trade-off.

The valid part of the coverage analysis applies standard split-conformal arguments and a union bound under the stipulated exchangeability assumptions. The joint bound itself is correct, but it concerns only IoU-matched detections. It does not cover missed lesions or unmatched false alerts, leaving a considerable gap to the clinical reliability motivation. The appendix also does not adequately specify patient- or video-level splits or explain how correlated detections within images and videos fit the assumed statistical unit.

The additional PAT optimization analysis contains explicit errors. Theorem A.5 treats the parameter-dependent coefficient αk(1−E[pi,k(w)])\alpha_k(1-\mathbb E[p_{i,k}(w)]) as a fixed positive weight when asserting convexity. That argument is invalid; the complete objective can be nonconvex even in a simple binary classifier with fixed features. Theorem A.6 correctly computes positive curvature for plog⁡(p+ϵ)p\log(p+\epsilon), then incorrectly states that multiplication by −γ-\gamma does not affect convexity. It reverses the curvature. The arguments therefore do not establish the claimed proximal uniqueness and stability, nor do they establish convergence of the actual Adam training and discrete matching procedure.

2. The particular PAT design is insufficiently motivated, and the key experiments do not reliably validate the combination.

I understand why both class imbalance and detection uncertainty matter, but I do not see why they motivate this particular design. Equation (8) imposes stronger classifier-norm shrinkage on rare, low-confidence classes. This is not equivalent to giving rare-class prediction errors greater weight, and the paper needs to explain and test why this shrinkage should improve rare-lesion recognition. DCC could be applied after other training schemes; the current evidence does not establish a necessary connection between PAT and conformal calibration beyond their sequential use. An on/off UA ablation does not isolate category priors, confidence adaptation, and ordinary regularization, and comparisons with standard imbalance treatments are insufficient.

More importantly, the tables supporting the main conclusions cannot be reconciled. Table 2 reports joint coverage/MPIW of 0.94/30.58 for RxT101-UA. Table 3 reports 0.88/84.14 for the same configuration and AP, with all five uncertainty statistics identical to the FCOS-UA row. The class-imbalance results also conflict: Tables 1/6 report an improvement in FCOS worst-category AP on Kvasir-Capsule from 42.8 to 43.9, whereas the minimum category AP in Table 7 decreases from 42.8 to 41.0. These discrepancies affect the principal success case and the problem the method is intended to address. Without corrected results and an unambiguous implementation, I cannot verify the central empirical claims.

Questions for authors
  1. Does the implementation use coordinate-wise Bonferroni calibration from Eq. (15), or joint maximum-residual calibration from Algorithm 3? Either can be valid, but the manuscript mixes the two. Please specify the residual collections, thresholds, and procedure used for the principal tables.

  2. Are the confidence-dependent coefficients in Eq. (8) detached or frozen between optimization steps? If so, what objective is actually optimized, and what do Theorems A.5/A.6 establish about that training procedure?

  3. Please reconcile the RxT101 results in Tables 2/3 and the worst-category AP in Tables 1/6/7. After correction, which configurations actually achieve the target coverage while improving efficiency?

  4. Please provide the patient or video-level split protocol, matching threshold, quantile-head training objective and configuration, and the anonymous repository mentioned in the reproducibility statement. These details are necessary to verify the results.

1 comment
  1. Alignment of the implemented UA-CP pipeline with its mathematical specifications

PAT operates during detector training, whereas DCC subsequently calibrates intermediate detector predictions before producing the final uncertainty-aware detection outputs. For each matched calibration detection (Appendix B.1), DCC computes four coordinate-wise CQR nonconformity scores using the maximum of the lower- and upper-endpoint errors. As specified in Eq. (15), the principal method maintains separate residual collections for the four bounding-box coordinates and computes finite-sample corrected conformal quantiles at the 1−αB/41-\alpha_B/4 level. These coordinate-specific thresholds construct simultaneous bounding-box regions with marginal coverage of at least 1−αB1-\alpha_B, as established by Proposition A.3 in Appendix A through the union bound. Thus, the intended calibration procedure is coordinate-wise Bonferroni calibration, not joint maximum-residual calibration.

  1. Experimental Consistency and Coverage-Efficiency Trade-offs

The inconsistencies between Tables 2 and 3 are reporting errors: the RxT101-UA uncertainty row in Table 3 mistakenly duplicates the FCOS values, whereas the FCOS row is correct. For LGLDD, the corrected Faster R-CNN–RxT101-UA results achieve 0.98 label coverage, 0.95 box coverage, and 0.94 joint coverage, exceeding the nominal target of 0.90. This configuration improves uncertainty efficiency by reducing MPIW from 98.68 to 30.58 and mean label-set size from 1.16 to 1.03, while increasing AP from 37.4 to 42.8. On Kvasir-Capsule, Table 12 shows that RxT101-UA, DINO, and FCOS achieve joint coverage of 0.96, 0.95, and 0.96, respectively, while reducing mean label-set size without increasing MPIW. The definition of APWorstAP_{\mathrm{Worst}} is already provided in the caption of Table 1. Thus, Faster R-CNN–RxT101-UA demonstrates both theoretical joint coverage target and improved uncertainty efficiency on LGLDD, while RxT101-UA, DINO, and FCOS achieve coverage–efficiency on Kvasir-Capsule. The reporting errors will be corrected in the revised manuscript.

  1. Reproducibility and Implementation Details

The requested methodological details are already addressed in several sections of the revised manuscript. Section 5.1 and Appendix C.2 specify the train–calibration–test partitioning protocol. Appendix B.1 (Eqs. 43–44) describes the IoU-based matching procedure, while Section 4.3 (Eqs. 14–16) and Appendix B.3 present the CQR localization calibration framework. Furthermore, our implementation is built upon the established MMDetection [1] and MMYOLO [2] frameworks, and the Reproducibility Statement explicitly identifies the accompanying anonymous code repository. These sections provide the methodological foundation and implementation context necessary to understand and reproduce the proposed framework.

[1] MMDetection: https://github.com/open-mmlab/mmdetection

[2] MMYOLO: https://github.com/open-mmlab/mmyolo

Reviewer_ifypOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes UA-CP, a two-stage pipeline for endoscopic lesion detection. UC-CP consists of two components:

  1. Prior-Aware Training (PAT): The detector is trained with category-frequency priors. For detectors with classification heads, this enters as a prior-weighted group penalty on classifier weights with scaling. For query-based detectors (e.g., DINO), it is added to the Hungarian matching cost.

  2. Dual Conformal Calibration (DCC): The authors use RAPS to get conformal label sets. They use coordinate-wise CQR with a Bonferroni split to yield conformal box regions.

Finally, experiments are run on several datasets: LGLDD (gastroscopy, 4 classes), Kvasir-Capsule (9 classes) and LDPolypVideo (1 class). The authors report AP gains from PAT in most settings, plus label-set size, MPIW and coverage tables for DCC.

Critical strengths
  1. Uncertainty for endoscopic detection (both category and localization) is clinically important and less explored than the usual domains for CP detection.

  2. The combination of imbalance-aware training (PAT) with post-hoc CP is reasonable and intuitive.

  3. The authors test different types of detectors and adapt PAT per type. They use repeated calibration subsets (10 × 75%), and perform coverage-target sweeps (Figs. 6–10), hyperparameter sensitivity analysis, and a computational complexity comparison.

Critical weaknesses
  1. The PAT component feels incremental and not actually linked tightly to CP. PAT is applied independently of the calibration stage. Its only link to the CP portion is that it changes the class probabilities RAPS consumes, so the "coupled" framing overstates the integration.

  2. There are mismatches between the theory and the algorithms. Eq. 15 uses a Bonferroni split, while Algorithm 3 uses a max-score with a single quantile, so the proof may not apply to what was run. The proof of Prop. A.4 refers to quantities the method never defines (a "maximum correction over the predicted label set" and class-specific boxes).

  3. Exchangeability is asserted without discussing correlated frames, and Appendix C.2 does not specify whether the 3:1:1 split is by frame, image, patient or video. This should be addressed explicitly in a paper of this type.

  4. At the 0.95 targets (joint bound 0.90), three of four detectors miss in Table 2 (joint coverage 0.80 for R50, 0.84 for DINO, 0.88 for FCOS; box coverage 0.82, 0.85, 0.90 against 0.95), and DCC on the raw R50 detector reaches only 0.74 box coverage in Table 3. Under exchangeability, split conformal calibration with thousands of calibration points doesn't feel like it should miss coverage by such margins.

  5. In Table 2, DCC on R50 has lower coverage than Box-Std at essentially the same MPIW (box 0.82 vs. 0.97, joint 0.80 vs. 0.95). In Table 12, DCC's MPIW is identical to the baseline values for all four detectors, with box coverage equal or within 0.01. In Table 13, DCC matches the baseline exactly on FCOS, is about 1 point lower on RxT101 and DINO, and is better only on R50 (0.95 vs. 0.89). The paper should show where DCC improves on standard box calibration and clarify whether the baselines are applied to the raw or PAT-trained detector.

  6. There are no ablations of PAT components or hyperparameters (prior weight, the error-adaptive factor, loss-level vs. assignment-level PAT, etc.).

  7. The paper has many typos ("Abalation", "Dectection", "conplementary", "Resolusion", "GLOPs", "LDPolypVedio", "Join Cov.").

Questions for authors
  1. What part of the PAT formulation is novel?

  2. Why do empirical results fall well below expected numbers?

  3. Does the framework deliver reliable uncertainty for the classes that matter clinically?

1 comment
AuthorsOpenReview ↗

Methodological and Empirical Assessment

  1. What part of the PAT formulation is novel?

Unlike conventional class reweighting, PAT incorporates category imbalance into architecture-specific optimization mechanisms. It combines category-frequency priors with the detector's current classification error through an error-adaptive, prior-weighted penalty for proposal-based and dense detectors, and a prior-aware matching cost for query-based detectors. This design is motivated by the long-tailed LGLDD dataset, where polyps account for 63.3% of annotated boxes, compared with 17.7% for ulcers and 9.5% for cancers. Its clinical relevance is supported by evidence that gastric cancers can be missed during endoscopy [1], particularly subtle early lesions [2]. Thus, PAT's contribution lies in its architecture-specific integration of category priors and prediction errors, rather than class balancing itself.

  1. Why do empirical results fall well below expected numbers?

The discrepancy between nominal and empirical coverage can arise from finite-sample variability, departures from exchangeability, or calibration implementation. Angelopoulos and Bates [3] explain that conformal prediction guarantees marginal coverage rather than the target coverage for every calibration/test split, while Barber et al. [4] analyze how departures from exchangeability affect validity. Similar discrepancies have been documented in medical applications: Yasar et al. [5] report empirical coverage of 0.847 against a nominal target of 0.90 under cross-cohort evaluation, while Vatanparast et al. [6] identify coverage shortfalls associated with limited calibration samples and approximate quantile calibration. These findings suggest that UA-CP’s below-target results are plausible, including the reported LGLDD joint coverage of 0.80 for Faster R-CNN–R50 and 0.84 for DINO, both below the nominal target of 0.90.

  1. Does the framework deliver reliable uncertainty for the classes that matter clinically?

UA-CP provides a clinically relevant approach to uncertainty quantification by combining improved detection of underrepresented lesions with category prediction sets and bounding-box uncertainty regions. For clinically important categories such as submucosal tumors (stm), ulcers (uls), and cancers (can), DCC characterizes ambiguity in lesion classification and localization, potentially helping clinicians identify findings requiring further examination or expert review [7]. The revised paper supports this potential through category-wise AP improvements, qualitative examples in Appendix E, and pooled category coverage reaching 0.98 on LGLDD.

[1] Pimenta-Melo A R, Monteiro-Soares M, Libânio D, et al. Missing rate for gastric cancer during upper gastrointestinal endoscopy: a systematic review and meta-analysis[J]. European journal of gastroenterology & hepatology, 2016, 28(9): 1041-1049.

[2] Banks M, Graham D, Jansen M, et al. British Society of Gastroenterology guidelines on the diagnosis and management of patients at risk of gastric adenocarcinoma[J]. Gut, 2019, 68(9): 1545-1575.

[3] Angelopoulos and Bates. "Conformal Prediction: A Gentle Introduction." Foundations and Trends in Machine Learning, 2023. doi:10.1561/2200000101

[4] Barber R F, Candes E J, Ramdas A, et al. Conformal prediction beyond exchangeability[J]. The Annals of Statistics, 2023, 51(2): 816-845.

[5] Yasar S, Yagin F H, Alzakari S A, et al. A Generalizable and Interpretable Framework for Molecular Subtype Classification of Pancreatic Ductal Adenocarcinoma Integrating Conformal Uncertainty Quantification and Consensus-Based Explainable Artificial Intelligence Across Multiple Cohorts[J]. International Journal of Molecular Sciences, 2026, 27(15): 6989.

[6] Vatanparast M, Elyassirad D, Ahmadzadeh A M, et al. Conformalized Semi-Supervised Learning for Medical Imaging[J]. Journal of Imaging Informatics in Medicine, 2026: 1-12.

[7] Leibig C, Allken V, Ayhan M S, et al. Leveraging uncertainty information from deep neural networks for disease detection[J]. Scientific reports, 2017, 7(1): 17816.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.