Organizations: Indian Statistical Institute, Kolkata 700108, India · Jadavpur University, Kolkata 700032, India · University College London, London WC1E 7JE, United Kingdom
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.
Figures & tables
Figure 1: Overview of the proposed AdaConRed pipeline with stages: (1) vision-language generative augmentation, (2) frozen DermFoundation encoding, (3) MLP-based classification, (4) entropy-modulated, margin-aware conformal prediction, and (5) adaptive conformal redistribution (AdaConRed).
Figure 2: Illustration of different prompting modalities used to generate OCA samples.
Figure 3: An illustration of conformal prediction set construction using calibration-based quantile thresholding, resulting in uncertainty-aware overlapping class regions.
Figure 4: Overview of the proposed redistribution strategy, where ambiguous samples from conformal prediction sets are reassigned to non-malignant or malignant categories, based on relative model confidence (N: non-malignant, A: ambiguous, M: malignant).
Method
All
Benign
OPMD
OCA
ResNet-50
65.33
25.25
91.85
11.76
ResNet-50 (Aug.)
58.00
39.39
76.36
21.43
InceptionV3
66.83
51.52
78.26
32.35
InceptionV3 (Aug.)
61.08
38.38
78.80
36.90
MobileNet
68.67
54.04
79.62
35.29
MobileNet (Aug.)
61.69
36.36
80.43
39.29
Table 1a: Class-wise accuracy (%) comparison on the OSCC dataset (3-class)
Method
All
AK
BCC
BKL
MEL
NV
SCC
UNK
ResNet-50
73.37
54.78
73.30
38.55
53.59
65.14
54.00
86.92
InceptionV3
73.46
51.59
63.72
53.25
47.77
68.49
67.00
86.05
MobileNet
75.52
61.78
55.25
41.69
48.31
78.32
52.00
87.52
MSL+MLP
85.78
50.32
84.90
63.86
65.90
86.92
62.00
93.92
MSL+MLP+AdaConRed
85.82
50.96
86.56
73.73
67.39
88.21
64.00
91.59
DF+MLP
85.83
59.87
84.71
58.07
66.04
86.51
73.00
94.19
Table 1b: Class-wise accuracy (%) comparison on the ISIC dataset (7-class)
Figure 5: Confusion matrices (row-normalised) for DF+MLP (Aug.) on the 3-class OSCC dataset (a) before and (b) after adaptive conformal redistribution.
Table 2a: Illustrative examples of conformal redistribution on the OSCC dataset
Table 2b: Illustrative examples of conformal redistribution on the ISIC dataset
Method
All
Benign
OPMD
OCA
MSL+MLP
80.33
73.23
86.14
58.82
MSL+MLP (Aug.)
78.31
63.64
88.32
69.05
DF+MLP
84.00
83.84
87.77
44.12
DF+MLP (Aug.)
82.31
67.17
91.85
76.19
Table 3a: Conformal prediction coverage (%) for the standard conformal set Γ(x) on the OSCC dataset (3-class)
Method
All
AK
BCC
BKL
MEL
NV
SCC
UNK
MSL+MLP
80.73
45.86
77.53
65.06
53.59
77.39
50.00
92.48
DF+MLP
89.90
68.15
85.45
67.95
77.40
91.81
77.00
95.79
Table 3b: Conformal prediction coverage (%) for the standard conformal set Γ(x) on the ISIC dataset (7-class)
Dataset
Score
Malig. Acc. (%)
All Acc. (%)
All F1 (%)
All Spec. (%)
OSCC
LAC
67.86
76.06
73.86
90.42
APS
72.62
77.21
75.46
91.62
RAPS
75.00
77.72
75.80
92.10
Proposed
79.76
79.64
77.18
93.79
ISIC
LAC
66.31
86.03
70.47
94.27
APS
66.85
86.26
70.88
95.36
Table 4: Comparison of conformal score functions under a fixed DF+MLP backbone at α=0.2 . The redistribution rule is applied unchanged in all cases, so differences are attributable to set construction alone. Malignant-class accuracy refers to OCA for OSCC and MEL for ISIC. OSCC results use the 4-class setting, ISIC the 7-class setting.
Dataset
DF+MLP
+ Aug.
+ APS
+ Entropy-margin (AdaConRed)
OSCC
72.50
75.67
77.21
79.64
ISIC
85.83
—
86.26
87.19
Table 5: Component-wise ablation under a fixed DermFoundation backbone, reported as overall accuracy (%). Augmentation applies to OSCC only. Redistribution is applied in the two rightmost configurations. OSCC uses the 4-class setting, ISIC the 7-class setting.
Figure 6: Effect of the entropy scaling parameter λ on malignant-class and overall classification accuracy for the OSCC and ISIC datasets under the DF+MLP backbone at α=0.2 .
Framework
Embedding
Acc.
F1
Spec.
AUC
BEiT3 [ 37 ]
ViT-G
71.24
69.37
72.56
76.31
PubMedCLIP [ 38 ]
ViT-B/32
67.13
65.42
68.35
72.46
BMCA-CLIP [ 39 ]
ViT-L/14
73.67
71.94
74.82
78.42
ViP [ 40 ]
ViT-B/16
72.44
70.63
73.58
77.17
MAVL [ 41 ]
ResNet-50
74.42
72.84
75.53
79.13
eCLIP [ 42 ]
Swin-T
73.61
71.87
74.73
78.33
Table 6a: Comparative study with state-of-the-art methods reported in the literature on the 4-class OSCC dataset
Framework
Segmentation
Classes
Acc.
F1
DenseNet201-Cubic SVM [ 44 ]
×
MEL, NV, BCC, AK, BKL, DF, VASC, SCC
80.87
-
ResNet18+VGG19-SVM [ 45 ]
✓
BEN, MEL
89.09
90.16
DSCCNet [ 46 ]
×
BCC, MEL, NV, SCC
83.20
58.09
ResNet50-Inception [ 47 ]
×
BEN, MEL
84.10
-
UNet [ 48 ]
×
BEN, MEL
79.90
-
GLCM-SVM+RF [ 49 ]
✓
SK, BCC, MEL
89.31
44.09
Table 6b: Comparative study with state-of-the-art methods reported in the literature on the multi-class ISIC dataset
Figure 7: Subgroup-wise comparison of conformal coverage, average prediction set size, and A2 accuracy on the OSCC dataset across (a) age group, (b) gender, (c) smoking habit, (d) betel quid chewing habit, and (e) lesion category.
Figure 8: Subgroup-wise comparison of conformal coverage, average prediction set size, and A2 accuracy on the ISIC dataset across (a) age group, (b) gender, (c) anatomical region, and (d) lesion category.
Figure 9: Comprehensive heatmap visualization of subgroup-wise conformal prediction metrics across the (a) OSCC and (b) ISIC datasets.
Figure 10: Distribution of conformal prediction set cardinalities across different strata on the OSCC dataset across (a) age group, (b) gender, (c) smoking habit, (d) betel quid chewing habit, and (e) lesion category.
Figure 11: Distribution of conformal prediction set cardinalities across different strata on the ISIC dataset (a) age group, (b) gender, (c) anatomical region, and (d) lesion category.
Figure 12: Age-wise variation of conformal coverage, average prediction set size, and A2 accuracy across the (a) OSCC and (b) ISIC datasets.
Deep learning models for medical imaging often exhibit overconfidence, creating safety risks in ambiguous diagnostic scenarios. While Conformal Prediction (CP) provides distribution-free statistical guarantees, standard methods such as Regularized Adaptive Prediction Sets (RAPS) optimize for average efficiency and can mask severe failures on difficult inputs. We propose an Adaptive Lambda Criterion for RAPS that minimizes the worst-case coverage violation across prediction set size strata. On OrganAMNIST (58,850 abdominal CT images, 11 classes), standard size-optimized RAPS converges to near-deterministic behavior with stratified undercoverage on uncertain samples, while our method achieves 95.72 percent global coverage with average set size 1.09 and at least 90 percent coverage across all strata. Cross-domain validation on PathMNIST (107,180 pathology images, 9 classes) confirms generalizability. Quantitative Grad-CAM analysis (rho = -0.30, p < 1e-22) shows that multi-label predictions correspond to focused attention on anatomically ambiguous regions. These results demonstrate that the proposed method improves reliability while maintaining efficiency, making it suitable for safety-critical medical AI applications.
One Octadion, Novanto Yudistira, Lailil Muflikhah
Faculty of Computer Science, Universitas Brawijaya, Malang, Indonesia
In high-stakes healthcare applications, machine learning models are frequently trained on data from one patient population and deployed on another, creating a distribution shift that degrades both accuracy and reliability. Semi-Supervised Domain Adaptation (SSDA) addresses this by leveraging labeled data from some source domain to improve model performance on a target domain where labels are scarce. However, existing SSDA methods optimize primarily for point-prediction accuracy and offer no principled uncertainty quantification --- a prerequisite for clinical trust. Conformal Prediction (CP) can address this limitation by providing prediction sets with rigorous, distribution-free coverage guarantees. However, applying CP post-hoc to a pre-trained model can yield prohibitively large prediction sets, as SSDA pre-training methods do not account for the nonconformity score geometry that determines conformal set size. Conformal Risk Minimization (CRM) has been used to resolve this issue in the fully supervised setting by integrating the CP objective directly into model training, but it requires a large labeled dataset to compute nonconformity thresholds during training, precisely the data that is scarce in the SSDA regime. We propose an end-to-end framework that integrates CRM into the SSDA training objective, enabling effective CRM in the limited-labeled-target-data regime. The key idea is to utilize Optimal Transport (OT) to generate pseudolabels for unlabeled target instances, providing the additional training signal needed by CRM to operate using only a small labeled target set. This results in a model jointly optimized for domain invariance and conformal efficiency, producing prediction sets that are compact, coverage-valid, and support domain-specific constraints such as excluding mutually contradictory diagnoses in skin lesion classification.
Conformal predictions have attracted significant attention in the field of uncertainty quantification, mainly because of their strong marginal coverage guarantees. Full conditional guarantee is not an attainable goal, a well known fact in conformal predictions literature. As a result, several approaches have tried to approximate this behavior by adapting the conformal sets of test-time samples according to their similarity to calibration examples. Although the latter has gained traction and shown impressive performances for regression problems, its application to image classification remains under-explored. We conduct an extensive benchmarking on natural image classification tasks with vision-language models (VLMs), using our open source implementation of a recent localized conformal prediction algorithm. We show that straightforward usage of the cosine similarity between test-time and calibration visual features, an intuitive choice for VLMs, is not sufficient to improve over the non-local baselines. In response, we propose a simple non-linear transformation of the cosine similarities, which conserves marginal coverage guarantees and achieves statistically significant mean set sizes reduction. Code is available at https://github.com/cfuchs2023/lcp-vlm/.