Conventional convolutional neural networks have limitations in modeling long-range contextual information

The gap

Evidence profile

Sourced from the future work and stated research gap of the source papers, classified as general, spanning 3 journals.

Supporting evidence — 4 representative gaps

• Efficient hybrid CNN–ViT framework for pulmonary nodule classification in CT images using feature-level optimization (2026) · Scientific Reports · doi

  The study proposes a hybrid deep learning framework comprising EfficientNetV2B1 and a Vision Transformer (ViT) to classify images of pulmonary nodule classification. Because of the ability of convolutional networks and ViTs to capture both local and global features, the proposed methodology captures both the fine-grained and the broad structural features of CT images and gives better results than using only one of the components. To address the trade-off between the size and speed of the model and its accuracy, model pruning was applied. The model proposed showed the best results in accuracy, precision, sensitivity, specificity, F1 score, and all of the other measurement criteria. This framework also showed its superiority to all the baseline architectures and its robustness and its reliability to all the criteria. Because of the increasing cases of lung cancer and the / ARTICLE IN PRESSARTICLE IN PRESS ACCEPTED MANUSCRIPT Table 5. Comparison of our model’s performance with DbMCA and other baseline models (CNN, DNN, SVM) On LIDC-IDRI evaluated by Shweikeh et al.56

  generalfuture work
  Keywords: model framework images features proposed accuracy criteria baseline article press proposes hybrid deep learning comprising
• Deep Learning-Based Segmentation, Classification, and Explainable Diagnosis of Breast Cancer Lesions (2026) · Journal of Intelligent Decision Making and Information Science · doi

  This study presented and evaluated two parallel, end-to-end deep learning pipelines for breast lesion segmentation, ROI-based staging, classification, and explainability — one for mammography (CBIS- DDSM) and one for ultrasound (BUSI) — and packaged both into a working web-based demonstration application. The ultrasound pipeline substantially outperformed the mammography pipeline across segmentation Dice, classification accuracy, and, most importantly, malignant-class recall, highlighting that imaging modality and lesion conspicuity have a first-order effect on attainable CAD performance 1948 Journal of Intelligent Decision Making and Information Science Volume 3(7s), (2026) 01-a20 using comparable architectures. Grad-CAM visualizations confirmed that the mammography classifiers attend to anatomically plausible regions but do so diffusely rather than with tight lesion localization, consistent with the moderate segmentation and recall results. Future work should prioritize improving mammography malignant recall through recall-oriented loss functions, ROI-cropped classification inputs (mirroring the BUSI pipeline design), and stronger or attention-augmented segmentation backbones (e.g., Attention U-Net, nnU-Net) to raise CBIS-DDSM Dice. Incorporating multi-view mammography fusion (craniocaudal and mediolateral-oblique) and, where available, prior-exam comparison would better reflect radiologist workflow. Quantitative explainability evaluation against ground-truth lesion masks, model calibration analysis, and prospective validation on external, multi-institution data are also necessary before any component of this pipeline could be considered for clinical decision support. Finally, extending the deployed application with user-facing confidence calibration and uncertainty flags would improve its suitability as a genuinely assistive — rather than autonomous — diagnostic tool.

  generalfuture work
  Keywords: mammography lesion segmentation pipeline recall classification based explainability cbis ddsm ultrasound busi application dice malignant
• A segmentation-guided CNN–Vision transformer feature fusion framework for multi-class breast ultrasound image classification (2026) · PLoS ONE · doi

  This study proposed a segmentation-guided CNN and Vision Transformer feature fusion framework for multi-class breast ultrasound image classification. The main objective was to address the low rate of classification performance and to improve the classification of normal, benign, and malignant breast ultrasound images by first localizing the lesion region and then extracting more informative and complementary deep features from the focused region of interest. For lesion localization, FET UNet was used because of its ability to combine convolutional feature learning with transformer-based contextual modelling. The segmentation stage helped reduce the influence of irrelevant background tissue, speckle noise, and surrounding anatomical structures, thereby providing cleaner and more diagnostically meaningful inputs for the classi- fication stage. After segmentation, ResNet50 and ViT-based feature extractors were used to capture complementary infor- mation from the lesion-focused ultrasound images. The ResNet50 branch extracted local texture, boundary, and structural features, while the ViT branch captured global contextual relationships and long-range spatial dependencies. These features were fused to form a more robust feature representation and were then used to develop a deep neural network classifier. The proposed framework achieved strong classification performance, with an overall accuracy of 95.11%, macro sensitivity of 95.67%, macro specificity of 97.64%, and macro F1-score of 94.46%. The ROC and precision-recall analy- ses further confirmed the robustness of the proposed approach, with high class-wise AUC and average precision values across normal, benign, and malignant categories. Although the proposed framework demonstrated strong performance under cross-validation on the benchmark BUSI dataset, its generalizability to independently collected datasets remains unverified. Differences in ultrasound acqui- sition devices, imaging protocols, image quality, patient populations, class distributions, and annotation procedures may introduce domain shifts and affect its performance on external data. Therefore, future studies should evaluate the framework on independent breast ultrasound datasets containing comparable diagnostic categories and investigate the effects of domain-related differences on classification performance. If substantial domain-related performance deg- radation is observed, domain-adaptation and domain-generalization strategies, such as domain-adversarial training, maximum mean discrepancy-based feature alignment, correlation alignment, self-supervised pretraining, and test- time adaptation, could be incorporated into the framework to improve its robustness and generalizability across inde- pendently collected datasets. Table 2. Performance of different methods of the ablation study.

  generalfuture work
  Keywords: performance domain feature framework ultrasound classification proposed segmentation class breast lesion features used based macro
• Swin-DAFU: denoising-augmented feature fusion unit for breast lesion segmentation and classification in 2D mammograms (2026) · Scientific Reports · doi

  Conventional convolutional neural networks have limitations in modeling long-range contextual information. Simple feature concatenation or direct skip-connection fusion may not sufficiently reduce noisy or weak lesion-related responses from intermediate feature maps. Accurate breast lesion segmentation still requires effective use of both local boundary details and high-level semantic information.

  generalstated research gapevidence 5/5
  Keywords: conventional convolutional neural networks have limitations modeling long-range

Questions about this gap