AdURA-Net: Adaptive Uncertainty and Region-Aware Network

Antik Aich Roy; Ujjwal Bhattacharya

cs.CVFeb 27, 2026

AdURA-Net: Adaptive Uncertainty and Region-Aware Network

Authors: Antik Aich Roy, Ujjwal Bhattacharya

Organizations: Indian Statistical Institute, Kolkata, India

Abstract

One of the common issues in clinical decision-making is the presence of uncertainty, which often arises due to ambiguity in radiology reports, which often reflect genuine diagnostic uncertainty or limitations of automated label extraction in various complex cases. Especially the case of multilabel datasets such as CheXpert, MIMIC-CXR, etc., which contain labels such as positive, negative, and uncertain. In clinical decision-making, the uncertain label plays a tricky role as the model should not be forced to provide a confident prediction in the absence of sufficient evidence. The ability of the model to say it does not understand whenever it is not confident is crucial, especially in the cases of clinical decision-making involving high risks. Here, we propose AdURA-Net, a geometry-driven adaptive uncertainty-aware framework for reliable thoracic disease classification. The key highlights of the proposed model are: a) Adaptive dilated convolution and multiscale deformable alignment coupled with the backbone Densenet architecture capturing the anatomical complexities of the medical images, and b) Dual Head Loss, which combines masked binary cross entropy with logit and a Dirichlet evidential learning objective.

Figures & tables

Figure 1: Conceptual overview contrasting conventional binary classification with the proposed uncertainty-aware framework, which enables three-way decisions (positive, negative, uncertain).

Disease	Train Split			Validation Split
	negative	positive	uncertain	negative	positive	uncertain
Cardiomegaly	7,546	22,463	1,53,367	315	929	6398
Edema	15,276	47,720	1,20,389	1907	645	5090
Consolidation	18,793	12,500	1,52,092	477	821	6344
Atelectasis	998	28,488	1,53,899	1189	47	6406
Pleural Effusion	24,254	73,788	85,343	2969	1046	3627

Table 1: Class-wise distribution of labels in the CheXpert-Small dataset across the training and validation splits. Labels are denoted as positive, negative, and uncertain.

Figure 2: Overview of the proposed AdURA-Net architecture. The Adaptive Deformable Convolution Block enhances early geometric feature extraction. DenseNet-121 serves as the backbone for hierarchical feature propagation. A dual-head prediction module produces (1) class probabilities via a sigmoid classifier, and (2) Dirichlet Evidence $(e_{i})$ for uncertainty quantification. The network is trained jointly using BCE (masked) loss, Dirichlet evidential loss, offset loss, and orthogonal regularization. During inference, the BCE head outputs the raw prediction, and the Dirichlet head outputs evidence that is used for uncertainty calculations. The abstention gate checks the uncertainty values $(u_{i})$ . If it is greater than the threshold $(\tau=0.4)$ , then it replaces it with $(-1)$ ; otherwise, it is retained.

Figure 3: The final feature representation is processed by two parallel heads. The BCE classification head produces class-wise logits and is optimized using masked binary cross-entropy, where uncertain labels ( $-1$ ) are masked. In parallel, the Dirichlet evidence head estimates class-wise positive and negative evidence for each disease. These evidences are supervised using an uncertainty-aware evidential loss. Both heads are jointly optimized through a combined uncertainty-based training objective. Here, $B$ denotes the batch size, and $D_{0},\ldots,D_{N}$ represent the disease classes, each associated with positive ( $D_{i}^{+}$ ) and negative ( $D_{i}^{-}$ ) evidence.

Backbones	Micro-AUC	Uncertainty Recall	Selective Accuracy
Densenet-121	0.9334	0.4675	0.9528
Densenet-161	0.9315	0.4582	0.9298
Densenet-201	0.9348	0.4387	0.9315

Table 2: Comparison of different DenseNet backbones under the proposed uncertainty-aware framework in terms of Micro-AUC, uncertainty recall, and selective accuracy for the 5 disease classification task (Cardiomegaly, Edema, Consolidation, Atelectasis, Pleural Effusion).

Backbones	Micro-AUC	Uncertainty Recall	Selective Accuracy
Densenet-121	0.9318	0.4245	0.9365
Densenet-161	0.9140	0.4021	0.9320
Densenet-201	0.9267	0.4130	0.9310

Table 3: Comparison of different DenseNet backbones under the proposed uncertainty-aware framework in terms of Micro-AUC, uncertainty recall, and selective accuracy for the 13-disease classification task (Enlarged Cardiomediastinum, Cardiomegaly, Lung Opacity, Lung Lesion, Edema, Consolidation, Pneumonia, Atelectasis, Pneumothorax, Pleural Effusion, Pleural Other, Fracture, and Support Devices).

Class	Densenet-121	Densenet-161	Densenet-201
Cardiomegaly	0.9169	0.9021	0.9121
Edema	0.8594	0.8440	0.8458
Consolidation	0.9417	0.9302	0.9389
Atelectasis	0.7397	0.7329	0.7187
Pleural Effusion	0.9483	0.9312	0.9421
Mean AUC	0.8812	0.8680	0.8715

Table 4: Per-class AUC comparison across different DenseNet backbones for the 5-disease classification task (Cardiomegaly, Edema, Consolidation, Atelectasis, and Pleural Effusion).

Class	Densenet-121	Densenet-161	Densenet-201	Ensemble
Enlarged Cardiomediastinum	0.8712	0.8488	0.8683	0.8751
Cardiomegaly	0.9320	0.9158	0.9331	0.9489
Lung Opacity	0.8266	0.8732	0.8578	0.8395
Lung Lesion	0.8156	0.8187	0.8054	0.8824
Edema	0.8776	0.8566	0.8872	0.8780
Consolidation	0.9426	0.9332	0.9441	0.9504

Table 5: Per-class AUC comparison across different DenseNet backbones for the 13-disease classification task (Enlarged Cardiomediastinum, Cardiomegaly, Lung Opacity, Lung Lesion, Edema, Consolidation, Pneumonia, Atelectasis, Pneumothorax, Pleural Effusion, Pleural Other, Fracture, and Support Devices).

Class	Base U-Zero	Base U-One	U-Zero(Ours)	U-One(Ours)
Cardiomegaly	0.8474	0.8264	0.8482	0.8714
Edema	0.8647	0.8957	0.8997	0.9137
Consolidation	0.9090	0.8994	0.9380	0.8781
Atelectasis	0.8448	0.8710	0.8425	0.8215
Pleural Effusion	0.9047	0.9219	0.9226	0.9452
Mean AUC	0.8741	0.8828	0.8902	0.8856

Table 6: Per-class AUC comparison for baseline DenseNet-121 and the employed adaptive deformable model under U-Zero and U-One evaluation settings for the 5 disease (Cardiomegaly, Edema, Consolidation, Atelectasis and Pleural Effusion) classification task.

Backbone	Cardiomegaly	Edema	Consolidation	Atelectasis	Pleural Effusion	Mean AUC
densenet-161	0.8543	0.9117	0.9360	0.8125	0.9288	0.8886
densenet-201	0.8420	0.9264	0.9167	0.8460	0.9343	0.8930

Table 7: Per-class AUC comparison for the adaptive deformable convolution block integrated with different Densenet backbones under U-Zero setting for the 5 disease (Cardiomegaly, Edema, Consolidation, Atelectasis, and Pleural Effusion) classification task.

Backbone	Cardiomegaly	Edema	Consolidation	Atelectasis	Pleural Effusion	Mean AUC
densenet-161	0.8545	0.9012	0.9061	0.8474	0.9214	0.8861
densenet-201	0.8537	0.9199	0.9023	0.8565	0.9267	0.8918

Table 8: Per-class AUC comparison for the adaptive deformable convolution block integrated with different Densenet backbones under U-One setting for the 5 disease (Cardiomegaly, Edema, Consolidation, Atelectasis, and Pleural Effusion) classification task.

Figure 4: Energy distribution on unseen pneumonia samples. Lower energy indicates higher confidence. The uncertainty-aware model shifts the distribution toward higher energy values, reducing overconfident predictions.

Figure 5: Energy distribution on unseen Covid-19 samples. Lower energy indicates higher confidence. The uncertainty-aware model shifts the distribution toward higher energy values, reducing overconfident predictions.

Figure 6: Per-class binary confusion matrices (positive vs. negative) for the proposed uncertainty-aware model. This analysis evaluates whether incorporating uncertainty modeling affects standard class discrimination, demonstrating that positive and negative prediction performance is preserved across all disease categories.

Chest X-Ray Classification

Chest X-ray classification uses machine learning to detect and categorize diseases and findings in radiographs for screening, triage, and clinical decision support.

See all 100 papers

Bayesian uncertainty estimation improves clinical decision making in medical AI agents
Jul 22, 2026Frederik Hauke, Patrick Wienholt, Christiane Kuhl +4Clinical Decision Support Clinical Decision-Making
Department of Diagnostic and Interventional Radiology, University Hospital RWTH Aachen, Aachen, Germany. · Department of Medical Oncology, National Center for Tumor Diseases (NCT), Heidelberg University Hospital, Heidelberg, Germany. · Else Kröner Fresenius Center for Digital Health, TU Dresden, Dresden, Germany. +1
Med-AR: Autoregressive Vision-Language Pretraining for Long-Tailed Chest X-Ray Classification and Uncertainty-Aware Evaluation
Sep 24, 2026Janhavi Prabhu, Sahil, Akshay V +3Long-Tailed Classification Chest X-Ray Classification
Qure.ai, Bangalore, India
MedFormer-UR: Uncertainty-Routed Transformer for Medical Image Classification
Apr 10, 2026Mohammed Maaz Sibhai, Abedalrhman Alkhateeb, Saad B. AhmedMedical Image Classification Vision Transformer
[1] Department of Computer Science, Lakehead University

Jul 22, 2026J/K move · Enter open · S save

Bayesian uncertainty estimation improves clinical decision making in medical AI agents

Frederik Hauke, Patrick Wienholt, Christiane Kuhl, Dyke Ferber+3

Department of Diagnostic and Interventional Radiology, University Hospital RWTH Aachen, Aachen, Germany. · Department of Medical Oncology, National Center for Tumor Diseases (NCT), Heidelberg University Hospital, Heidelberg, Germany. · Else Kröner Fresenius Center for Digital Health, TU Dresden, Dresden, Germany.+1

Clinical Decision Support Clinical Decision-Making Monte Carlo Dropout Epistemic Uncertainty

Open paper AskPDF arXiv

Machine learning models for medical image analysis typically lack a reliable measure of confidence, limiting their use in ambiguous or atypical cases. Here we show that Monte Carlo dropout, applied to a multi-task chest-radiograph classifier (eight thoracic findings, 137,593 training images), provides an epistemic uncertainty signal that tracks generalisation across training-set scales and flags confident yet error-prone predictions. Adding this signal to the point prediction raised error-detection AUROC from 0.74 to 0.77 (

Δ

AUROC +0.023, 95% CI [+0.014, +0.033]). In a controlled 2x2 factorial experiment, a clinical-decision-support agent exploited this uncertainty only when it was delivered as a binary error-risk flag rather than as raw scores, cutting confident misdiagnoses on unreliable findings from 8.5% to 2.7%. Epistemic uncertainty estimation thus carries decision-relevant information beyond point predictions, but its value for downstream agents depends on how it is communicated.

Abstract

Chest X-Ray Classification

Explore similar work

Bayesian uncertainty estimation improves clinical decision making in medical AI agents

Med-AR: Autoregressive Vision-Language Pretraining for Long-Tailed Chest X-Ray Classification and Uncertainty-Aware Evaluation

MedFormer-UR: Uncertainty-Routed Transformer for Medical Image Classification