Experiments show that IRRL achieves balanced classification performance, with favorable F1-score and Matthews Correlation Coefficient results that reflect improved minority-class recognition quality, and robustness and consistency of the proposed representation learning strategy.
Abstract
Class imbalance is a prevalent issue in medical image classification that significantly degrades a model's capacity to recognize minority-class lesions, thereby restricting its applicability in real-world clinical screening scenarios. Existing studies typically address this problem through data resampling, loss re-weighting, or decision boundary adjustment strategies; however, these methods predominantly focus on compensation during the classification stage. In contrast, the representation learning process in earlier stages is often dominated by easy majority-class samples, and its impact on the feature quality of minority classes has not received adequate attention. To address this issue, we propose an Imbalance-Aware Robust Representation Learning (IRRL) framework for class-imbalanced medical image classification. IRRL prioritizes the refinement of minority-class-related local representations before global classification. Specifically, implicit local token representations are constructed from convolutional feature maps based on their receptive-field structure. Semantic confidence-guided reliability estimation, difficulty-adaptive supervised contrastive learning, and minority-class prototype regularization are then introduced to improve the learning of informative local representations and hard minority-class samples. Finally, a Transformer performs global context modeling for image-level classification. Experiments on four public datasets, including ISIC 2018, PAD-UFES-20, OCTID, and BUSI, show that IRRL achieves balanced classification performance, with favorable F1-score and Matthews Correlation Coefficient (MCC) results that reflect improved minority-class recognition quality. The results across datasets with different imaging modalities and imbalance conditions further demonstrate the robustness and consistency of the proposed representation learning strategy.
Recurrent Contrastive Learning progressively expands the support region of tail classes by recurrently reusing historical feature states across training phases and devise a Temporal Memory Queue (TMQ) to preserve corpus-level features across training phases and provide diversified global references for contrastive learning.
Zhiyuan Zhu, Xinling Meng, Junxuan Yu et al.· 0 citations
A data-driven semi-supervised framework for imbalanced binary image classification that does not depend on data augmentation, enabling reliable utilization of unlabeled data without introducing augmentation induced noise is introduced.
M. Neethu, S. S. Vinod Chandra· International Journal of Dat...· 0 citations
Skin cancer incidence is rising globally, and early accurate classification of dermoscopic lesions is critical for improving patient outcomes, particularly for melanoma where delayed detection drastically worsens prognosis. This work presents a comprehensive framework for multiclass skin lesion classification on the ISIC 2019 benchmark, which comprises eight diagnostic categories and a ninth unknown out-of-distribution (OOD) class in the test set. Our system addresses three interlinked challenges: extreme class imbalance, robust integration of patient metadata, and reliable OOD rejection. The architecture utilises an EfficientNet-B4 backbone along with a novel Adaptive Metadata Gating (AMG) module that learns imageconditioned gates bounded by an explicit ceiling in order to prevent shortcut reliance on metadata. Training incorporates an asymmetric focal loss with a double penalty for malignant false negatives, a hierarchically annealed auxiliary malignancy loss, class-aware Mixup/CutMix augmentation, Shades-of-Gray colour constancy, Exponential Moving Averaging, and Stochastic Weight Averaging. HAM10000 data supplements minority classes. Five independently trained EMA-checkpointed models are combined via softmax-temperature weighted logit-space averaging with fiveview test-time augmentation. An entropy-plus-gate composite uncertainty score drives calibrated per-fold threshold tuning for OOD rejection. On the official ISIC 2019 test set, the ensemble achieves balanced accuracy of 57.2%, macro AUC of 0.924, and macro specificity of 95.4%. On a fresh 20% stratified hold-out of ISIC training data, balanced accuracy rises to 80.3% and macro AUC to 0.975, confirming strong in-distribution discriminative capacity, with the test-set gap attributable to domain shift and the presence of unlabelled OOD samples.
S. V., Chris George Shibu, Gali Manish Kumar· 2026 International Conferenc...· 0 citations
Class imbalance is prevalent in real-world datasets. Minority samples are far fewer than majority samples. Traditional classifier design typically assumes balanced data, which causes classifiers to favor the majority class when faced with imbalanced datasets. Thus, there are high misclassification costs for minority classes in critical fields like healthcare and finance. Most existing oversampling methods for handling imbalance problems, such as SMOTE (the synthetic minority oversampling technique), suffer from limitations like noise sensitivity, failure to consider minority-class sub-cluster structures, and poor adaptability to the heterogeneity of sample distributions. This article addresses these issues by proposing a novel oversampling algorithm: the adaptive weighting–synthetic minority oversampling technique (AW-SMOTE). It consists of three progressive stages. In the first stage, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering identifies the distribution structure of the minority samples. It identifies potential sub-clusters while removing noise interference. A clear data foundation for subsequent sampling is established. In the second stage, the most representative boundary sample in each cluster is used to evaluate the weight of each cluster. The total number of synthetic samples is allocated to different clusters according to their weights. This provides global sample enhancement support. In the third stage, adaptive sample generation is performed within each cluster. It combines the two perspectives of boundary tightness and local density. The sigmoid function is used to dynamically adjust the weight ratio. Finally, new samples are synthesized in key regions to both preserve distribution characteristics and enhance discriminability in classification. Through experiments on standard datasets from the KEEL repository, the feasibility and effectiveness of this algorithm are demonstrated.
OBJECTIVE
Existing prototype learning methods for open-set recognition (OSR) often use a fixed number of prototypes to represent each class, which struggle to model the inherent intra class variations widely existing in practical scenarios. This limitation is particularly pronounced in medical image applications, where high intra-class heterogeneity makes accurate OSR challenging. To address this issue, we propose a novel framework called Adaptive Multi-Prototype Network with Pretrained Swin Transformer (PSW-AMPN) for OSR on medical images. Specifically, a sparse gated attention module is devised to compute attention scores based on prototype-sample relations, thereby adaptively suppressing redundant sub-prototypes and selectively activating discriminative ones for each class in an end to-end optimization process. By jointly optimizing classification loss and the regularization for open space risk based on multiple prototypes, PSW-AMPN effectively captures complex intra-class structures and enhances class discriminability. Furthermore, PSW-AMPN uses a Pretrained Swin Transformer and a lightweight projector as feature extractor to effectively capture both local and global features. Extensive experiments demonstrate that our approach significantly outperforms existing baselines and achieves state-of-the-art performance on multiple medical image OSR tasks.
Xingyu Cai, Haiyan Yang, Jiayi Chen et al.· IEEE transactions on bio-med...· 0 citations
Deep learning has shown strong potential in medical image analysis, but most existing methods rely on large-scale annotations and a closed-world assumption that rarely holds in clinical practice. Although Generalized Category Discovery (GCD) has advanced rapidly on natural images, it remains underexplored in medical imaging. To address this issue, we propose MedXplore, a unified framework for reliable and unbiased medical GCD, optimizing from both perceptual and decision levels. Specifically, at the perceptual level, taking a frequency domain perspective, Frequency-SNR Adaptive Attention and Consistency (FAAC) performs learnable full-spectrum filtering and global-local energy contrast activation to not only highlight local abnormal signals relative to the global context, but also provide reliable semantic anchors for patch consistency learning. At the decision level, Adaptive Cosine-Angular Margin (ACAM) adjusts angular margins using semantic difficulty and feature confidence to balance intra-class compactness and inter-class separability. Together, the two modules improve lesion-sensitive representation learning and mitigate old-class bias. Experiments on multiple benchmarks show an average \textbf{8.5\%} gain in \textit{All} accuracy over the strongest competing methods. On Kvasir, MedXplore reduces false-old errors from 14.50\% to 0.80\%, demonstrating strong robustness under severe old-new ambiguity.
