PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Conformal Redistribution for Inter-class Transitional Uncertainty in Medical Image Classification

Saibal Ghosh, Samarup Bhattacharya, Sanjoy Kumar Saha, Umapada Pal, Tapabrata Chakraborti

arXiv:2609.13303v1cs.CVcs.AIcs.LGstat.COstat.ML

TL;DR

Transitional disease categories create ambiguous predictions, while conformal prediction sets may leave screening decisions unresolved. AdaConRed adds label-free redistribution to adaptive conformal sets, and experiments report improved overall accuracy across score functions on OSCC and ISIC.

  • Problem

    Transitional categories overlap neighboring classes, and conformal prediction sets can remain multi-label when screening requires a single decision.

    Source: §1 · p. 2

  • Method

    AdaConRed constructs adaptive conformal sets and redistributes multi-label sets for transitional predictions to an alternative class using model outputs at inference.

    Source: §1 · p. 3, §3.6 · p. 11

  • Results

    Redistribution improves overall accuracy over no-redistribution baselines under every score function, and the proposed score attains the highest values across reported metrics.

    Source: §4.4 · p. 18

  • Takeaways & Limitations

    AdaConRed extends conformal prediction from uncertainty estimation toward uncertainty-aware decision support for transitional disease categories.

    Source: §5 · p. 26

    Redistribution applies only to multi-label sets with clinically plausible labels; the original set remains available alongside the refined decision.

    Source: §4.2 · p. 16

Abstract

from arXiv · show

Medical image classification is frequently complicated by transitional categories whose feature distributions overlap those of adjacent classes, producing ambiguous decision boundaries. Conformal prediction returns uncertainty-aware prediction sets, but these are not directly actionable in clinical screening, where a single decision is required. This work proposes adaptive conformal redistribution (AdaConRed), a label-free post-conformal decision rule that converts ambiguous prediction sets into refined class assignments. A five-stage pipeline is developed. Vision-language generative augmentation addresses minority-class scarcity; a frozen DermFoundation encoder provides embeddings; a lightweight multi-layer perceptron performs classification; an entropy-modulated, margin-aware nonconformity score constructs adaptive prediction sets; samples predicted as transitional with multi-label sets are reassigned to the most probable alternative class within the set, using only model outputs at inference. Evaluation uses the OSCC oral lesion and ISIC skin lesion benchmarks at a miscoverage level of 0.2. On the 3-class OSCC benchmark, overall accuracy improves from 73.54% to 77.38%, with oral cancer accuracy rising from 64.29% to 82.14% and benign accuracy from 56.57% to 70.20%. Reassignment of transitional samples reduces OPMD accuracy from 84.78% to 80.16%, consistent with the asymmetric cost of missed malignancy. On ISIC, overall accuracy improves from 85.83% to 87.19%, melanoma accuracy rising from 66.04% to 68.34%. AdaConRed outperforms LAC, APS and RAPS under an identical backbone and redistribution rule. Conformal prediction can be extended beyond uncertainty quantification toward actionable decision support where transitional disease categories are present, with gains concentrated in the clinically critical malignant categories. Code repository: https://github.com/saibal436ghosh/AdaConRed.

1 Introduction

Transitional disease categories overlap neighboring classes, creating ambiguous predictions whose errors can carry asymmetric clinical costs. AdaConRed uses conformal prediction sets to refine ambiguous predictions into single-label decisions.

1 Introduction

  • Transitional-category errors are asymmetric: assigning a malignant lesion to a transitional class can delay treatment, while the converse incurs additional review.Conformal predictors can return multi-label sets that are statistically valid but do not provide the single decision screening and triage require.Source: §1 · pp. 2–3
  • Malignant categories are often scarce, and underrepresentation can bias conventional supervised learning toward majority classes and degrade sensitivity.Generative augmentation and vision foundation models are presented as ways to address data scarcity, but not the separate ambiguity of transitional classes.Source: §1 · p. 2
  • Existing generative augmentation and foundation-model representations do not resolve transitional-category ambiguity, and classifiers still produce deterministic point predictions without confidence.Synthetic samples may contain hallucinated structures, so calibration and evaluation are performed exclusively on real samples.Source: §1 · p. 2
  • AdaConRed combines vision-language generative augmentation, adaptive conformal sets, and label-free redistribution to improve malignant-class sensitivity at a controlled cost to the transitional category.OSCC is the primary benchmark, with ISIC used to evaluate generalizability in a diverse multi-class setting.Source: §1 · p. 3

2 Literature Survey

Prior work explores mobile-phone oral lesion classification, pretrained representations, generative augmentation, and conformal uncertainty estimation. The paper identifies transitional-class ambiguity and uncertainty-aware decision-making as remaining gaps, and combines augmentation with conformal redistribution.

2 Literature Survey

  • Prior oral-lesion studies used photographs and deep learning, while OPMD-related ambiguity was identified as a source of confusion among benign, transitional, and malignant lesions.Mobile-phone imaging has been explored for settings with limited specialist access, and prior work reported potential misclassification among lesion classes.Source: §2 · pp. 3–4
  • Pretrained vision models and generative augmentation address data scarcity, but improved feature learning and synthetic samples do not resolve transitional-class ambiguity.The survey notes transferable representations from pretrained models and GAN-generated samples for minority and borderline classes.Source: §2 · p. 4
  • Conformal prediction has been used mainly to quantify uncertainty, while the study targets redistribution of borderline predictions into alternative categories.The paper identifies OPMD as underexplored and combines generative augmentation with conformal prediction-based redistribution for oral-lesion classification.Source: §2 · pp. 4–5

3 Methodology

The methodology combines augmentation, frozen image encoding, classification, conformal prediction, and redistribution. AdaConRed forms adaptive sets and uses model confidence to refine ambiguous predictions.

3.1 Dataset Description

  • OSCC is the primary three-class benchmark, while seven-class ISIC serves as a secondary generalization test; their OPMD and UNK categories assess transitional ambiguity.The OSCC dataset has limited OCA samples and receives augmentation; ISIC has sufficient samples across selected categories and receives no additional augmentation.Source: §3.1 · p. 5

3.2 Generative Augmentation

  • The OSCC OCA class is augmented with 750 synthetic images generated using text-to-image, image-to-image, and hybrid image-plus-text prompting.Text prompts use lesion descriptions, image prompts use reference OCA images, and the hybrid strategy combines both.Source: §3.2 · p. 6

3.3 Image Encoding

  • A frozen DermFoundation encoder extracts dense 6144-dimensional embeddings without task-specific fine-tuning; its representations transfer effectively to intraoral images despite domain shift.The encoder is pretrained predominantly on dermatological imagery.Source: §3.3 · p. 7

3.4 Classification

  • A compact, dataset-specific MLP classifies the embeddings, using class-weighted cross-entropy and label smoothing to address imbalance and overconfidence.OSCC uses one 128-neuron hidden layer; the seven-class ISIC head adds a second 32-neuron hidden layer.Source: §3.4 · p. 7

3.5 Conformal Prediction Framework

  • Split conformal prediction uses a calibration set to estimate a threshold for test prediction sets, with marginal coverage guaranteed under calibration-test exchangeability.The dataset is partitioned into training, validation, calibration, and test subsets.Source: §3.5 · pp. 7–8

3.6 Proposed Adaptive Conformal Redistribution Strategy

  • AdaConRed forms adaptive sets from cumulative probability mass and probability-distribution geometry, then refines multi-label ambiguous predictions using relative model confidence.At inference, reassignment depends only on model outputs; the method is presented as post-conformal refinement, while formal coverage remains the role of the standard split conformal set.Source: §3.6 · pp. 9–11
  • Entropy scales the influence of margin-aware weighting: flatter, more ambiguous predictions yield more cautious prediction sets.For confident, low-entropy samples, the margin term contributes less and sets are more compact.Source: §3.6 · p. 10
  • Only samples predicted as AMB with multi-label adaptive sets are reassigned to the most probable non-AMB class in the set.The rule is label-free at inference and is designed to reduce false negatives in malignant lesion detection.Source: §3.6 · p. 11

4 Results and Discussion

Across OSCC and ISIC, AdaConRed improves classification accuracy while retaining conformal uncertainty estimates; results also examine score construction, validation-sensitive trade-offs, and subgroup behavior.

4.1 Experimental Setup

  • Evaluation uses a 3-class lesion-bearing OSCC setting and a 7-class ISIC setting, with conformal miscoverage α=0.2.OSCC augmented and non-augmented test blocks contain different numbers of OCA samples and should be compared within, not across, blocks.Source: §4.1 · pp. 12–14, §4 · p. 12

4.3 Statistical Coverage Analysis

  • OSCC OCA coverage for DF+MLP rises from 44.12% to 76.19% after augmentation, while ISIC DF+MLP overall coverage is 89.90%.Per-class coverage is not guaranteed to meet the 80% marginal target; the OSCC OCA value remains below it.Source: §4.3 · p. 17, Table 3a · p. 17, Table 3b · p. 17

4.4 Comparison with Conformal Baselines, Sensitivity Analysis, and Ablation

  • Redistribution improves accuracy over no-redistribution baselines under every tested score; the proposed score exceeds RAPS by 1.92 OSCC points and 0.14 ISIC points.With the redistribution rule fixed, score construction accounts for the differences; LAC gains are marginal because it often produces singleton sets.Source: §4.4 · p. 18, Table 4 · p. 18
  • At λ=2, malignant-class accuracy and overall accuracy peak under the tested sensitivity analysis, while larger λ values increasingly redistribute genuinely transitional samples.The value is selected on validation data by maximizing malignant-class accuracy, with overall accuracy breaking ties.Source: §4.4 · p. 19, Figure 6 · p. 20
  • The component ablation attributes additional accuracy gains to both APS redistribution and entropy-modulated margin weighting on OSCC and ISIC.On OSCC, augmentation adds 3.17 points, APS redistribution adds 1.54, and margin weighting adds 2.43; on ISIC, the latter two add 0.43 and 0.93 points.Source: §4.4 · p. 19, Table 5 · p. 19

4.5 State-of-the-Art Comparison

  • In the 4-class OSCC comparison, DermFoundation AdaConRed reaches 79.64% overall accuracy, 5.22 points above the strongest reported comparator, MAVL.AUC remains unchanged because redistribution changes discrete labels without changing the softmax probabilities used to compute AUC.Source: §4.5 · pp. 19–20, Table 6a · p. 20

4.6 Subgroup Reliability Analysis

  • Subgroup analyses report variation in uncertainty and A2 accuracy across demographic, risk, lesion, and anatomical strata, with OSCC coverage relatively stable across subgroups.These are descriptive point estimates without confidence intervals or significance testing, not formal fairness evidence; OSCC OCA has the largest reported average set size, 1.44.Source: §4.6 · pp. 21, 22, 23, 26

5 Conclusion

The paper addresses ambiguous predictions from transitional disease categories by applying adaptive conformal redistribution to prediction sets. Across oral and skin lesion tasks, it reports improved classification performance, with gains concentrated in clinically critical malignant categories.

  • Standard classifiers can resolve transitional-category ambiguity arbitrarily, while conformal predictors may leave it unresolved in multi-label sets.Source: §5 · p. 26
  • AdaConRed uses an entropy-modulated, margin-aware score to construct adaptive prediction sets, then applies a label-free decision rule using only model outputs at inference.Standard split conformal sets retain the role of providing formal coverage.Source: §5 · p. 26
  • Experiments on oral and skin lesion datasets show improved classification performance, with gains concentrated in clinically critical malignant categories.Source: §5 · p. 26
  • A descriptive subgroup analysis characterizes uncertainty variation across demographic and lesion strata.Source: §5 · p. 26
  • The non-malignant, ambiguous, and malignant grouping is illustrative of the case studies; future work will examine risk-stratified conformal prediction, hierarchical uncertainty modeling, and task-adaptive scores.The redistribution mechanism requires a designated transitional category, but is not tied to the labels used in these studies.Source: §5 · p. 26
Loading 2609.13303v1…