AdURA-Net: Adaptive Uncertainty and Region-Aware Network

TL;DR

The key highlights of the proposed AdURA-Net 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.

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 (6)

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