Hybrid quantum convolutional neural network for multi-channel image classification

Hyperspectral image classification has garnered significant attention due to its crucial applications in terrain identification and scene understanding. However, the complexity of high-dimensional spectral data, high interclass spectral similarity, and diverse spatial scales present substantial challenges for classification tasks. This paper introduces a novel multiscale dual-channel spatial attention network (MSDANet) for hyperspectral image classification, which innovatively combines multiscale feature extraction with a dual-channel spatial attention to enhance classification performance. Specifically, MSDANet implements multiscale feature extraction through parallel multibranch structures and dilated convolutions, improving the model’s adaptability to targets at different scales. Additionally, we design a dual-channel spatial attention mechanism that integrates channel and spatial attention to achieve adaptive enhancement of spectral and spatial features. The incorporation of depthwise separable convolutions and lightweight attention modules significantly reduces computational complexity. Furthermore, an innovative feature fusion strategy employing residual connections and adaptive fusion enhances feature extraction effectiveness. Experiments conducted on three benchmark datasets demonstrate the superior classification performance of the proposed MSDANet approach. The method achieves overall classification accuracies of 96.07%, 96.85%, and 94.20% on the Indian Pines, Pavia University, and Kennedy Space Center datasets, respectively, significantly outperforming existing methods. Comprehensive ablation studies validate the effectiveness of each innovative component, with results indicating strong generalization capabilities in handling complex scenes and few-shot learning scenarios. These technical strengths make MSDANet particularly valuable for real-world applications, including precision agriculture for accurate crop type identification and health monitoring, and forestry management for precise species classification and sustainable resource assessment.

Hybrid quantum inception-inspired convolutional neural network for image classification

ProtoRadNet: Prototypical patches of Convolutional Features for Radiology Image Classification Network.

Denoised generative fusion networks for noise-robust few-shot image classification

In medical imaging diagnosis, the identification of normal liver, fatty liver, and cirrhosis is often challenging due to subtle morphological and density differences. Previous studies have used CNNs, MLPs or transformers to extract lesion features. However, CNN’s global representation is limited, while MLPs and transformers have insufficient local modeling, resulting in insufficient lesion information mining. Therefore, this article proposes a hybrid network CMT-Net, which unifies the local perception of CNNs, high-dimensional mapping of MLPs, and global dependencies of Transformers into a single architecture, significantly improving the accuracy of CT liver three classification. The core components of CMT-Net include an efficient transformer (ET) module, which focuses on extracting local feature details while progressively integrating global information, significantly enhancing the model feature representation and generalization capabilities. Additionally, this paper introduces a Hybrid MLP (HM) module that combines Token-Mixing MLP and Channel-Mixing MLP to achieve deep fusion of spatial and channel information, further improving feature extraction. To validate the proposed algorithm, extensive experiments were conducted on a dataset of three liver diseases collected from the Imaging Department of Urumqi People’s Hospital. The results demonstrate that CMT-Net achieves outstanding performance in classifying normal liver, fatty liver, and cirrhosis. These findings not only provide an effective tool for precise liver disease diagnosis but also offer new directions for deep learning model design in medical image classification tasks.

Multi-scale feature enhancement and cross-dimensional attention mechanism fusion network for hyperspectral image classification

SRML: Structure-relation mutual learning network for few-shot image classification

A CNN-Transformer fusion network for Diabetic retinopathy image classification.

HSCRN: A hyperspectral image classification network based on dynamic context and feature refinement

Bloodstream infections caused by Candida spp. are among the leading hospital-acquired infections, but their diagnosis remains slow and challenging with conventional culture-based methods, which often require days to deliver results. This study aimed to develop a rapid and sensitive diagnostic strategy for the detection and quantification of Candida spp. directly from blood samples. We designed a workflow combining antibody-modified magnetic beads for pathogen isolation, magnetic SERS (surface-enhanced Raman scattering)-encoded tags for multiplexed detection, and fluorescence microscopy for rapid prescreening. Data analysis was automated using machine learning, including convolutional neural networks for image classification and self-organizing maps for spectral analysis. This method enabled the detection and quantification of seven clinically relevant Candida species (C. albicans, C. glabrata, C. tropicalis, C. auris, C. haemulonii, C. dubliniensis, and C. parapsilosis) in 7.5 mL of whole blood at septicemia-relevant concentrations as low as 2 CFU mL-1, with the results obtained in 4-5 hours. High specificity was demonstrated, with minimal cross-reactivity against bacterial controls. This integrated approach represents a rapid, sensitive, and multiplex alternative to current diagnostics, with the potential to improve the early detection and targeted treatment of candidemia, thereby enhancing the clinical outcomes and reducing the healthcare burden.

BACA-Net: a band-adaptive collaborative attention network for hyperspectral image classification

BackgroundOtitis media remains a significant global health concern, particularly in resource-limited settings where timely diagnosis is challenging. Artificial intelligence (AI) offers promising solutions to enhance diagnostic accuracy in mobile health applications.ObjectiveThis study introduces a hybrid AI framework that integrates convolutional neural networks (CNNs) for image classification with large language models (LLMs) for clinical reasoning, enabling real-time otoscopic diagnosis.MethodsWe developed a dual-path system combining CNN-based feature extraction with LLM-supported interpretation. The framework was optimized for mobile deployment, with lightweight models operating on-device and advanced reasoning performed via secure cloud APIs. A dataset of 10,465 otoendoscopic images (expanded from 2820 original clinical images through data augmentation) across 10 middle-ear conditions was used for training and validation. Diagnostic performance was benchmarked against clinicians of varying expertise.ResultsThe hybrid CNN–LLM system achieved an overall diagnostic accuracy of 97.6%, demonstrating the synergistic benefit of combining CNN-driven visual analysis with LLM-based clinical reasoning. The system delivered sub-200 ms feedback and achieved specialist-level performance in identifying common ear pathologies.ConclusionsThis hybrid AI framework substantially improves diagnostic precision and responsiveness in otoscopic evaluation. Its mobile-friendly design supports scalable deployment in telemedicine and primary care, offering a practical solution to enhance ear disease diagnosis in underserved regions.

Semi-supervised deep residual generative adversarial network for hyperspectral image classification

Spiking Neural Networks (SNNs), designed to more accurately model the brain’s neurobiological processes, have been proposed as energy-efficient alternatives to conventional Artificial Neural Networks (ANNs), which typically incur high computational and energy costs. However, the enhanced energy efficiency and computational savings incurred by using SNNs are often achieved at the expense of reduced classification performance. Recent studies have investigated the incorporation of attention mechanisms into SNNs to enhance their classification performance, but these approaches typically repurpose attention mechanisms originally developed for conventional ANNs, which fail to fully leverage the spike-based encoding characteristics intrinsic to spiking neuron dynamics. To address this challenge, we propose the Biologically Inspired Attention Spiking Neural Network (BIASNN), a novel SNN architecture designed for image classification. BIASNN introduces a biologically inspired attention mechanism that integrates adaptive leaky integrate and fire neurons with components from established attention models. Our attention mechanism is placed into an existing SNN architecture using leaky integrate and fire neurons to enhance biological fidelity by combining multiple spiking neuron models in a single network. Experiments on benchmark image classification datasets demonstrate that BIASNN achieves high classification accuracy using only four timesteps. By enabling the development of more biologically plausible attention mechanisms, BIASNN advances the capabilities of deep spiking neural networks toward more brain-like processing.

Spatial Transformer Correlation Network for natural image classification

Taxonomy-guided routing in capsule network for hierarchical image classification

The architecture design and training optimization of spiking neural network with low-latency and high-performance for classification and segmentation.

SSSAT-Net: Spectral-Spatial Self-Attention-Based Transformer Network for hyperspectral image classification

KFCNet: A Key Feature Consistency Network for microscopic urinary sediment image classification