ABSTRACT Recent medical image segmentation networks, such as SwinUNet and AttentionUNet, achieve high accuracy through complex architectures containing hundreds of millions of parameters. However, due to expensive annotation and limited data (e.g., the MSD‐Pancreas dataset contains only 281 samples), these models often suffer from overfitting and poor generalization. To address these challenges, we propose a lightweight 3D medical image segmentation network, AMPPNet, with only 1.97 M parameters and three major innovations: the Adaptive Multi‐stage Polarization Perception (AMPP) module, which integrates two key submodules: dual‐bond efficient polarization multi‐scale channel self‐attention (DPMC) and dual‐bond multi‐scale polarization hybrid spatial attention (MSPHA). The DPMC module achieves cross‐layer channel semantic alignment and local–global polarization fusion through a dual‐bond design, while the MSPHA module captures multi‐scale spatial dependencies by jointly attending to local texture details and global structural cues. These modules, combined with the Adaptive Multi‐stage Feature Fusion (AMFF) mechanism, enable hierarchical semantic aggregation with minimal computational overhead. The Pyramid Adaptive Matrix Multi‐Level Loss (PAMMLoss) dynamically combines Dice, cross‐entropy, and Tversky losses at shallow, mid, and deep layers with learnable adaptive weights, enabling fine‐grained optimization and mitigating gradient conflicts. AMPPNet is highly parameter‐efficient and accurate—despite its compact design, it achieves excellent Dice scores of 81.17% (NIH), MSD (10 datasets) of 72.71%, demonstrating its strong generalization capabilities under data‐scarce conditions. Our code is publicly available: https://github.com/XiangruW/AMPPNet .
ABSTRACT While the Segment Anything Model (SAM) has revolutionized medical image segmentation with its zero‐shot capabilities, its reliance on manual, per‐image prompting (e.g., points or boxes) imposes a prohibitive interaction burden and creates a profound intent gap between sparse spatial cues and complex anatomical semantics. To address these challenges, we propose OneRef‐Seg, a unified task‐level single‐reference conditioning framework. OneRef‐Seg combines the flexibility of reference‐guided interaction with single‐example task‐level conditioning; at deployment, a unified OneRef‐Seg model, jointly trained across predefined binary segmentation tasks, receives a single annotated reference image‐mask pair that specifies the current anatomical target and automatically segments unseen target images from the same task without per‐target spatial prompting. Technically, we introduce a Reference‐Target Feature Fusion Module (RTFM) designed to explicitly bridge the semantic gap through dense contextual alignment. The RTFM utilizes a cascaded reasoning mechanism to migrate reference priors into the target feature space, effectively bypassing the fragility of heuristic prompt generation. Extensive evaluations on the REFUGE, ACDC, and Fluoroscopy benchmarks demonstrate that OneRef‐Seg achieves performance comparable to single‐round prompt‐driven SAM variants while significantly reducing user intervention. Notably, OneRef‐Seg remains highly competitive against fully supervised specialist models, offering a practical and efficient solution for high‐throughput medical AI workflows.
ABSTRACT Multilabel retinal disease classification from fundus images remains challenging due to severe long‐tailed label distributions, heterogeneous lesion scales, and patient‐specific disease comorbidity. To address these issues, we propose LPC‐Net, a unified framework that improves the recognition of rare diseases while preserving interpretability. LPC‐Net introduces: (i) an anatomically guided lesion‐aware multi‐instance learning module that aggregates multiscale patches from global, macula‐centered, and optic‐disc‐centered regions without requiring lesion annotations; (ii) a prototype‐balanced margin regularization that enforces geometric separation for tail‐class prototypes to mitigate feature collapse under imbalance; and (iii) a sample‐specific comorbidity graph that blends dataset‐level co‐occurrence priors with instance‐level predictions, enabling a lightweight label GNN to refine logits via residual corrections. Experiments on ODIR‐5K and RFMiD demonstrate consistent gains over strong baselines. LPC‐Net achieves 0.901 mAP on ODIR‐5K and 0.768 mAP on RFMiD, with notable improvements on tail classes. Under prespecified six‐output mappings and without target‐domain fine‐tuning, cross‐dataset evaluation yielded 0.846/0.618/0.681 (Macro‐AUC/Macro‐F1/mAP) for ODIR‐5K‐to‐RFMiD and 0.836/0.598/0.659 for RFMiD‐to‐BRSET. These results are interpreted as transfer under harmonized label definitions rather than universal diagnostic equivalence or clinical generalizability.
ABSTRACT Sickle Cell Disease (SCD), the most prevalent inherited disease worldwide, requires the rapid and accurate detection of sickle‐shaped red blood cells. In Hematology, manual examination of sickle cells in blood smears is time‐consuming and prone to qualitative errors. Artificial intelligence has transformed healthcare, offering revolutionary solutions to diagnostic problems. The current experiment proposes a hybrid feature fusion framework combining ResNet‐50 and Vision Transformer (ViT) to overcome the limitations of individual convolutional or transformer‐based models, known as ViResNet‐SCD. The model integrates the local features of microscopic images from ResNet‐50 and distinguishes global spatial relationships through ViT. The model is trained on a well‐documented dataset, and the dataset has been augmented with state‐of‐the‐art data augmentation strategies to handle the class imbalance issue. The final dataset consists of 10 000 images in training (70%), validation (15%), and testing (15%). For facilitating clinical confidence and model interpretability, Gradient weighted Class Activation Mapping (Grad‐CAM) was utilized to highlight the important morphological regions that influence the model's classification. The hybrid model attained a test set accuracy of 98.80%, outperforming the baseline models, namely ResNet‐50 and ViT, across all assessment metrics. For statistical significance, the architecture also attained a Cohen's Kappa and Matthews Correlation Coefficient (MCC) value of 0.976, indicating wide agreement and balanced classification. Comparative analysis with state‐of‐the‐art methods shows a significant improvement in the results produced by the proposed method.
ABSTRACT In recent years, segmentation models based on deep learning have provided new perspectives and methods for automated skin lesion diagnosis. However, when dealing with dermoscopic images, existing methods often ignore uncertain lesion boundaries and ambiguous regions, leading to limited segmentation stability and accuracy. To solve this problem, this paper proposes a novel network for dermoscopic image segmentation based on fuzzy learning, named SFUNet. By integrating the Transformer backbone into the U‐Net architecture, the model extracts semantic information from input medical images and yields a pixel‐level representation for segmentation. Since boundary uncertainty and region ambiguity are widespread in dermoscopic image segmentation, a new adaptive fuzzy module is integrated into the U‐Net skip connection path. This module constructs several membership functions for the input multi‐scale features to characterize gradual transitions at uncertain lesion boundaries, and applies an adaptive gated fusion mechanism to the reconstructed features, thereby improving boundary accuracy and suppressing noise. In addition, an attention mechanism is introduced in the decoder to focus on the lesion‐related regions. The effectiveness of SFUNet was evaluated through comparisons with existing advanced segmentation networks on the ISIC 2017 and ISIC 2018 datasets. Cross‐dataset evaluation was performed on the PH dataset to verify the model's generalization ability. Experimental results showed that SFUNet improved overlap and boundary metrics, demonstrating its strong segmentation performance and computational efficiency.
