Aug 2026· Journal of King Saud University: Computer and Information Sciences· Vol 38· 0 citations· 79 references
TL;DR
Extensive experiments across three medical imaging benchmarks, two brain tumor classification datasets (SARTAJ, Br35H) and dental radiography analysis demonstrate that CENet variants achieve state-of-the-art efficiency-accuracy trade-off.
Abstract
Accurate diagnosis from medical images is critical yet challenging in resource-constrained clinical settings, where models must balance high accuracy with computational efficiency for practical deployment. While deep learning has revolutionized medical image analysis, a significant gap persists: heavyweight models (e.g., Transformers, dense CNNs) achieve strong performance but are infeasible for edge devices, whereas existing lightweight architectures often lack the sophisticated context modeling needed for robust detection of subtle and heterogeneous pathologies. To address this challenges, we introduce the Context-Enhanced Network (CENet), an efficient convolutional architecture designed for medical image classification. CENet incorporates three key modules: (1) a Two-Level Feature Fusion Module (TLFM) that performs selective, gated fusion of only the most semantically complementary feature scales to prevent noise propagation; (2) a novel Triplet Excitation Attention (TEA), a parameter-efficient channel attention module for robust feature re-calibration; and (3) Lightweight Context-Enhanced Module (LightCEM) that enables iterative, self-guided feature refinement using global context. Extensive experiments across three medical imaging benchmarks, two brain tumor classification datasets (SARTAJ, Br35H) and dental radiography analysis demonstrate that CENet variants achieve state-of-the-art efficiency-accuracy trade-off. Notably, the CENet-small variant achieves weighted F1-score of 76.08% on the multi-class SARTAJ dataset while utilizing only 2.08 M parameters and 0.232 GFLOPs, outperforming larger architectures. Furthermore, Qualitative analyses using Grad-CAM further show that CENet consistently localizes clinically meaningful regions, enhancing interpretability. Meanwhile, CENet demonstrates real-time embedded performance on the NVIDIA Jetson Orin NX, achieving up to 75.17 FPS for CENet-small (and 62.21 FPS for CENet-large), which supports practical deployment under the strict latency and power constraints. While CENet achieves strong weighted performance across all benchmarks, classes with diffuse morphology (e.g., glioma) or subtle presentation (e.g., cavity) remain challenging and require further investigation. The code for this work will be made available at https://github.com/aminabenabid/CENet
Medical image analysis has undergone transformative progress with the application of deep learning models. However, existing architectures often struggle to effectively balance local feature extraction with global contextual understanding, which is crucial for complex diagnostic tasks such as Retinopathy of Prematurity (ROP) detection. In this study, we present a pretrained lightweight Conformer model tailored for medical image classification. The model integrates convolutional layers for capturing fine-grained spatial features with transformer blocks that capture long-range dependencies, creating a unified architecture capable of robust representation learning. We evaluate the model across multiple benchmark medical imaging datasets, including ROP, BloodMNIST, RetinalMNIST and other MedMNIST benchmark datasets. With 93.61% accuracy on the ROP dataset and 99.12% accuracy on BloodMNIST, experimental results show competitive classification performance while lowering model complexity to 12.4 million parameters and 3.2 GFLOPs. Experimental results demonstrate that the comparative studies versus CNN-based and transformer-based architectures, such as ResNet50, Swin-Tiny, ConvNeXt-Tiny, Vision Transformer, and MedViT. The findings show that in clinical settings with limited resources, the suggested lightweight Conformer offers a practical and computationally efficient alternative for medical image interpretation. Furthermore, the lightweight design ensures computational efficiency, making it suitable for deployment in resource-constrained healthcare environments. These findings validate the lightweight Conformer model’s potential for scalable, accurate, and real-time medical image classification.
Sreelekshmi Vijayasree, Adithya K. Krishna, Akarsh S. Nair et al.· Journal of Imaging· 0 citations
Medical image segmentation requires high accuracy and robustness, yet practical commercial deployment also demands privacy preservation and computational efficiency. In this context, the U-Net architecture, which can be inherently decoupled into independent encoder and decoder components, serves as a natural commercial choice. However, pure Transformer-based variants like Swin-UNet often suffer from insufficient local detail capture and limited interpretability. In this paper, we propose a lightweight hybrid architecture built upon the Swin-UNet framework. Our model integrates a parallel CNN encoder to complement the shallow layer reasoning of Swin Transformers with local texture features. To bridge the semantic gap and enhance fine-grained spatial detail recovery, we design an asymmetric feature fusion strategy and introduce cross-layer skip (XSkip) connections that explicitly propagate shallow CNN features into the decoder. We further incorporate novel loss functions and an auxiliary supervision head (Aux-Head) to strengthen training stability, boundary delineation, and intermediate feature interpretability. Extensive experiments on the Synapse multi-organ segmentation dataset demonstrate that our approach achieves state-of-the-art competitive Dice scores and Hausdorff distances, offering an accurate, efficient, and interpretable solution for clinical deployment.
TvaraNet is pro-posed, an extremely lightweight segmentation network designed to preserve boundary fidelity under strict efficiency constraints and achieves competitive or superior boundary-aware performance compared to heavier architectures.
Automated ultrasound image classification is increasingly important for clinical decision support in breast, thyroid and fetal screening. However, deploying deep learning models in such safety-critical settings demands not only high predictive accuracy but also transparency, interpretability and trustworthiness-properties that existing approaches address insufficiently. Convolutional neural networks (CNNs) capture local texture patterns but struggle with global contextual dependencies, while Transformer-based models offer long-range reasoning yet require large-scale training data and remain sensitive to ultrasound-specific noise, both limiting factors for clinical deployment. We propose Multi-Scale CNN Token Transformer (MSCT-Trans), a lightweight and interpretable hybrid architecture for general-purpose ultrasound image classification. MSCT-Trans extracts multi-scale feature maps from a pre-trained CNN backbone and converts them into a unified token sequence, enabling a Transformer encoder to model global dependencies and inter-scale interactions over semantically meaningful, noise-attenuated representations. To support clinical transparency, we conducted a two-part explainability analysis-Gradient-weighted Class Activation Mapping++ spatial localisation and softmax class probability breakdown-demonstrating that MSCT-Trans consistently attends to diagnostically relevant anatomical regions, produces well-calibrated confidence estimates and associates prediction errors with model uncertainty rather than over-confident mis-classification. Here we evaluated MSCT-Trans on three ultrasound benchmarks spanning breast (BUS-BRA + BUSI + UCLM), thyroid (TN5000) and fetal imaging. MSCT-Trans consistently outperformed CNN and Transformer baselines across accuracy, macro-F1 and area under the receiver operating characteristic curve, particularly under class imbalance and limited data regimens. The combination of strong predictive performance, spatially grounded interpretability and calibrated uncertainty estimation positions MSCT-Trans as a transparent and trustworthy foundation for ultrasound-based clinical decision support. Code: https://github.com/MohsinFurkh/MSCT-Trans.
Mohsin Furkh Dar, Sayima Mukhtar· Ultrasound in Medicine and B...· 0 citations
With the rapid development of precision medicine and intelligent diagnostic technologies, automatic medical image classification has become an important tool for assisting clinical decision-making. However, substantial variations in lesion scale, complex long-range dependencies of tissue structures, and the need to capture subtle anatomical features present significant challenges to existing deep learning models. To address the limitations of the Swin Transformer in modeling multi-scale lesions and multi-level feature interactions, this study proposes a hierarchical complementary feature enhancement framework based on the Swin Transformer for medical image classification. The proposed architecture performs collaborative feature learning at three representation levels, including macro-scale lesion perception, global contextual interaction, and local detail refinement. Specifically, a Multi-scale Depthwise SE Block (MSD-SE Block) is introduced at the input of each stage of the Swin Transformer to enhance the model’s multi-scale feature representation capability. Subsequently, a Residual Convolutional Attention (RCA) module is integrated following the self-attention mechanism and the Multi-Layer Perceptron (MLP) to strengthen global contextual modeling, while a Local Detail Enhanced Residual Channel-Spatial Attention (LDERCSA) module is employed to refine subtle anatomical structures and discriminative local features. Through the coordinated interaction of these components, the proposed framework establishes a hierarchical feature enhancement mechanism that effectively improves medical image representation across multiple scales and feature levels. Comprehensive experiments were conducted on eight core subsets of MedMNIST v2, including BloodMNIST, BreastMNIST, DermaMNIST, OCTMNIST, OrganSMNIST, PathMNIST, PneumoniaMNIST, and RetinaMNIST, using an input resolution of 224
$$\times$$
224. Single-module comparison and ablation studies demonstrate that each proposed component contributes positively to the overall performance. Experimental results show that the proposed model achieves significant performance improvements on BreastMNIST, OCTMNIST, OrganSMNIST, and PneumoniaMNIST. Furthermore, comparative evaluations across all datasets confirm that the proposed framework exhibits strong robustness and generalization capability when processing multiple medical imaging modalities, including ultrasound, CT, X-ray, endoscopic, and microscopic images, thereby providing reliable technical support for computer-aided medical diagnosis systems.
Yachao Si, Yi Zhang, Ming-Zhan Zhao· Scientific Reports· 0 citations
Medical image segmentation plays a pivotal role in computer-aided diagnosis. However, the scarcity of annotated data severely hinders the deployment of deep learning models. Few-shot learning (FSL) is designed to achieve rapid adaptation to unseen classes using limited labeled samples, among which prototype-based methods have emerged as a dominant paradigm. Nevertheless, existing approaches often rely on single or coarse multi-prototype representations, failing to capture complex morphological variations and local details in medical images. Furthermore, boundary feature contamination arising from convolutional receptive fields severely degrades segmentation accuracy. To address these challenges, this paper proposes an Auxiliary Supervision-guided Edge Prototype Purification Network (EPPNet) for few-shot medical image segmentation. Specifically, the network introduces a novel Prototype Purification Module (PPM). By evaluating the semantic consistency between edge and main prototypes, the PPM selectively integrates high-confidence edge prototypes via a learnable adaptive threshold, aiming to substantially alleviate boundary feature contamination. Simultaneously, an Adaptive Prototype Generation (APG) module and an Adaptive Weight Decoding (AWD) mechanism are designed to dynamically extract semantic-aware prototypes for foreground, background, and edge regions, while assigning optimal fusion weights. Furthermore, a parallel U-Net auxiliary branch is constructed to enhance the generalization capability and spatial context modeling of the feature encoder through fully supervised dense pixel prediction. Extensive experiments on three public datasets (CHAOS-MRI, Synapse-CT, and CMRSeg) demonstrate that the proposed method outperforms existing state-of-the-art techniques across multiple organ segmentation tasks. Detailed analyses and validation further confirm significant improvements in prototype representation and boundary information utilization.
Wen-Jie Meng, Kai Liu, Minghui Wang· IEEE Transactions on Medical...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
