Current deep learning algorithms have limitations

The gap

Evidence profile

Stated in the limitations and future work and cells research gap sections of the source papers, classified as general, spanning 4 journals. Those papers have been cited 1 times in total.

Supporting evidence — 5 representative gaps

• Transformers meet CNNs for insights into breast mass classification from histopathological images (2026) · Frontiers in Artificial Intelligence · cited 1× · doi

  • Investigated three deep learning approaches • Deep learning methods • Performance depends on the availability of for breast ultrasound (BUS) lesion detection: demonstrated overall superior labeled training data, including negative samples. patch-based LeNet, U-Net, and transfer detection performance across both • Evaluation is limited to 2D ultrasound lesion learning with pretrained FCN-AlexNet. datasets in terms of True Positive detection only, without downstream • Performance compared with four state-of- Fraction (TPF), false positives per segmentation or classification. the-art conventional methods: Radial Gradient image, and F-measure. • Dataset sizes are relatively small, and multi-center Index, Multifractal Filtering, Rule-based • FCN-AlexNet achieved the best validation is not explored. Yap et al. (2017) Region Ranking, and Deformable Part results on Dataset A, while patch- • The lack of a unified public benchmark remains a Models (DPM). based LeNet performed best on challenge. • Evaluated on two ultrasound datasets acquired Dataset B. using different imaging systems. • The study highlighted the adaptability and robustness of learning-based lesion detection methods across heterogeneous datasets. • Proposed a supervised breast ultrasound • Achieved competitive • The method is not fully automated, requiring (BUS) tumor segmentation method based on segmentation performance on a manual ROI selection by the operator. semantic classification and superpixel merging. BUS dataset of 320 cases, • Performance strongly depends on classifier • ROI is manually selected, followed by image producing tumor contours close to accuracy and dataset-specific training, limiting enhancement (histogram equalization, bilateral hand-labeled annotations. generalization across imaging devices. filtering, pyramid mean shift). • Reported an average F1-score of • While effective, results are inferior to fully Huang et al. (2020) • SLIC superpixels are generated, semantic 89.87% ± 4.05% and favorable TP/ supervised deep learning models (e.g., FCN) features extracted, and a Bag-of-Words (BoW) FP rates compared with five trained with pixel-level annotations. model constructed. existing approaches. • Scalability to large datasets and real-time clinical • Initial classification is performed using a • Demonstrated that semantic deployment was not evaluated. Backpropagation Neural Network (BPNN) and superpixel classification can refined using K-Nearest Neighbors (KNN) effectively guide tumor boundary reclassification. delineation. • Proposed a novel semantic segmentation • Demonstrated superior • Performance remains dependent on the quality network incorporating a channel attention segmentation performance and diversity of annotated training data. module with Multiscale Grid Average Pooling compared to existing deep learning • Like other deep learning approaches, the model (MSGRAP) to improve breast cancer approaches on a public requires pixel-level ground truth, increasing segmentation in ultrasound images. BUS dataset. annotation cost. Lee et al. (2020) • The attention module captures both global and • The proposed channel attention • Generalization to ultrasound images from local spatial information, overcoming module significantly enhanced different acquisition devices or clinical settings limitations of standard convolution. Network segmentation accuracy, showing was not extensively analyzed. architecture was optimized through ablation better semantic consistency and studies and evaluated against FCN, SegNet, boundary delineation of breast U-Net, and other state-of-the-art deep cancer regions. learning models. • Hybrid framework integrating Dense • Achieved high classification • Model interpretability remains limited due to the Convolutional Network (DenseNet) for accuracy of 99.32% on BreakHis black-box nature of deep learning. hierarchical feature extraction with Flower and 96% on BACH datasets, • Validation is restricted to benchmark datasets; Pollination Algorithm (FPA) for optimized outperforming existing state-ofreal-world clinical deployment and large-scale Wakili et al. (2025) feature selection on histopathological images the-art methods. multi-center validation are not explored. (BreakHis and BACH datasets). • Demonstrated improved • Computational cost of hybrid optimization may generalization, reduced increase training time.

  generalstated in limitationsevidence 5/5
  Keywords: learning performance deep datasets segmentation ultrasound dataset based classification semantic approaches breast detection training network
• AI-Based Classification and Prediction of Breast Cancer (2026) · International Journal of Drug Delivery Technology · doi

  combining different techniques and defining standard predictors for better results (Aberu et al., 2016). A detailed survey focusing on breast cancer screening techniques was conducted, offering a comprehensive analysis of their advantages and limitations. The study also discovered the pertinency of deep learning methods in breast cancer exposure, appraising various performance metrics and datasets used in this domain. Additionally, future research directions related to breast cancer detection were examined. The paper objects to offer a systematic understanding of field while encouraging innovative research efforts (Rautela et al., 2022). A study investigated the correctness and effectiveness of “Machine learning (ML)” and “Deep learning (DL)” methods for primary breast cancer alphanumeric mammography detection metaphors and microelectronic well-being histories. It identified 48% of false-negative cases and attained an “AUC of 0.91”, with a “specificity of 77.3%” and a al., 2019). “sensitivity of 87%” Additionally, a hybrid DL model was developed for the automatic detection of “invasive ductal carcinoma (IDC)” using the PCam Kaggle dataset. This model, which combines CNN and GRU architectures, surpassing delivered superior performance by pathologist-level addressing misclassification issues, and outperforming other models (Wang et al., 2022). An augmented “Deep recurrent neural network (RNN)” model was developed using Keras-Tuner optimization. The architecture included input, hidden, dropout, and output feature-selection techniques, the optimized RNN outperformed five layers. By accuracy, utilizing (Ballin using this et IJDDT, Volume 16 Issue 20s, 2026 Page 87 AI-Based Classification and Prediction of Breast Cancer exposure, introduced code employment, menarche traditional machine learning models (Saleh et al., 2022). A review of publicly available datasets for breast cancer diagnosis explored current deep learning approaches, repositories, and highlighted challenges and future opportunities in DL- based diagnosis (Iqbal et al., 2022).A separate study introduced a deep learning framework aimed at identifying breast anomalies by utilizing standard data. This framework incorporated pre-processing and feature extraction with a MobileNetV2 pre-trained model, achieving superior performance compared to previous methods and excelling in anomaly detection on the IN breast and MIAS datasets (Alloqmani et al., 2023). Machine learning algorithms were also employed to classify crucial risk influences for prime attacking breast cancer in the Iranian inhabitants.

  generalstated in future workevidence 5/5
  Keywords: breast cancer learning deep detection model techniques performance datasets machine using standard exposure additionally future
• Predicting the Severity of Thyroid Nodules with YOLOv8 and CA+LSR Architecture (2026) · Journal of Electronics, Electromedical Engineering, and Medical Informatics · doi

  terms of model generalization, computational efficiency, and applicability to diverse patient populations. Finally, Section V summarizes the this research, key suggesting paths future study and potential improvements to the methodology. insights and contributions of for in II. Materials and Methods The task of this study can be comprehensively divided into two main parts: model training and model evaluation. In the model training phase, as depicted in Fig. 1, we begin by collecting a diverse set of raw then meticulously ultrasound preprocessed to create a high-quality dataset. Each image is standardized and rescaled to a resolution of 640 × 640 pixels, with three color channels (R, G, and B). This standardization ensures uniformity across images, which are regions of the dataset, which is critical for effective model training. To increase the robustness of the dataset, we implement various data augmentation techniques. This involves systematically applying transformations such as flipping the images horizontally and vertically, rotating them by random angles, and cropping them to interest. These focus on specific augmentation strategies help to introduce variability in the training data, allowing the model to generalize better and perform accurately on unseen data. The model is then trained on this augmented dataset to minimize the loss function, which measures the difference between predicted outputs and actual labels. In the model evaluation phase, we scale the preprocessed ultrasound images appropriately before feeding them into the trained model. The model processes these images and generates multi-head detection results, which allow for the identification of multiple features or abnormalities within the images simultaneously. To further enhance the interpretability of the model's outputs, we apply various postprocessing techniques. While these techniques are pivotal for improving output clarity, a detailed discussion of these methods will not be covered in this paper. In the sections that follow, we will thoroughly discuss the essential steps of the proposed framework, highlighting techniques the methodologies and employed throughout the study. A. Patient's Data Preparation A well-balanced sample of benign and malignant nodules was collected before training the model to ensure reliable findings. The demographic and clinical characteristics of the study population are summarized in Table 1. The dataset comprises 105 benign nodules originating from 92 patients, including 60 women and 32 men, with an average age of 52.45±11.87 years. Conversely, the study also includes 105 malignant Fig. 1. Overview of the LabelMe-based data preparation pipeline and enhanced YOLOv8 object detection framework. Manuscript received 14 August 2025; Revised 2 December 2025; Accepted 20 March 2026; Available online 14 May 2026 Digital Object Identifier (DOI): https://doi.org/10.35882/jeeemi.v8i3.1137 Copyright © 2025 by the authors. This work is an open-access article and licensed under a Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0). 863 Journal of Electronics, Electromedical Engineering, and Medical Informatics Homepage: jeeemi.org; Vol. 8, No. 3, July 2026, pp: 861-877 e-ISSN: 2656-8632 nodules sourced from 90 patients, including 67 women and 23 men, with an average age of 49.72 ±10.95 years.

  generalstated in limitationsevidence 5/5
  Keywords: model training dataset images techniques them nodules patients diverse patient evaluation phase ultrasound preprocessed various
• 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.

  generalstated in future workevidence 5/5
  Keywords: mammography lesion segmentation pipeline recall classification based explainability cbis ddsm ultrasound busi application dice malignant
• A Novel DyDimNet-Based Dimension-Aware Deep Learning Framework for Automated Ovarian Tumor Classification Using Ultrasound Images (2026) · International Journal of Drug Delivery Technology · doi

  Current deep learning algorithms have limitations in utilizing dimension-specific spatial features in ovarian ultrasound images. - Traditional augmentation techniques can distort the original anatomy and negatively affect the model's ability to learn important features.

  generalstated in cells research gapevidence 5/5
  Keywords: current deep learning algorithms have limitations utilizing dimension-specific

Questions about this gap