The results indicate that combining attention-enhanced convolutional models with explainable AI and LLM-based reporting can improve both performance and interpretability, positioning DeepScan as a supportive CDSS framework for breast ultrasound analysis in pre-clinical and educational settings.
Abstract
Breast cancer diagnosis using ultrasound imaging remains challenging due to image noise, class imbalance, and limited interpretability of automated systems. This study presents DeepScan, a deep learning–based Clinical Decision Support System (CDSS) designed for research and educational use, which integrates a DenseNet121 backbone with a Convolutional Block Attention Module (CBAM) to enhance feature discrimination in breast ultrasound images. The system performs three-class classification (Normal, Benign, Malignant) using a structured preprocessing pipeline that includes resizing, ImageNet-based normalization, data augmentation, and controlled oversampling to address class imbalance. To improve transparency, Grad-CAM–based visual explanations are incorporated to highlight diagnostically relevant regions influencing model predictions. Beyond image-level classification, DeepScan integrates a Large Language Model (LLM)–based reasoning engine to generate BI-RADS–aligned, structured clinical reports and provide interactive explanations for users. Experimental evaluation on the BUSI dataset demonstrates strong discriminative performance, achieving AUC values of 0.998 for Normal tissue and 0.957 for both Benign and Malignant classes, with near real-time inference latency. The results indicate that combining attention-enhanced convolutional models with explainable AI and LLM-based reporting can improve both performance and interpretability, positioning DeepScan as a supportive CDSS framework for breast ultrasound analysis in pre-clinical and educational settings.
Breast cancer is one of the leading causes of cancerrelated deaths among women worldwide, making early and accurate diagnosis essential for improving patient survival and treatment outcomes. Traditional histopathological examination is considered the gold standard for breast cancer diagnosis; however, manual analysis is time-consuming, labor-intensive, and subject to inter-observer variability. To address these challenges, this paper presents an Explainable and Efficient AI-Based System for Automated Breast Cancer Diagnosis using deep learning. The proposed methodology employs the BreakHis histopathological image dataset, where images are preprocessed through resizing and normalization before being analyzed using a fine-tuned EfficientNet-B3 model integrated with a Convolutional Block Attention Module (CBAM) to enhance feature representation. Gradient-weighted Class Activation Mapping (Grad-CAM) is incorporated to provide visual explanations of the model’s predictions, improving transparency and interpretability. The trained model achieved a classification accuracy of 97.47% on the test dataset, demonstrating high performance in distinguishing benign and malignant breast cancer images. Furthermore, a Streamlitbased web application enables real-time image upload, prediction, confidence estimation, and explainability visualization. The proposed system offers an efficient, accurate, and interpretable decision-support tool that has the potential to assist pathologists in clinical breast cancer diagnosis while promoting trust in AIdriven healthcare applications.
V. Parvathi· International Journal of Eng...· 0 citations
It is suggested that integrating advanced CNN architectures with radiomics features can significantly enhance the accuracy and reliability of both lesion segmentation and treatment prediction in breast cancer, potentially leading to better clinical outcomes.
Astha Karnwal, E. Dhamija, Pradeeba Sridar· Indian Journal of Radiology...· 0 citations
Background: HER2 status guides targeted therapy in breast cancer but is currently determined by invasive biopsy. Imaging-based HER2 prediction from dynamic contrast-enhanced MRI (DCE-MRI) could provide a non-invasive adjunct decision-support signal, but published models are typically single-center with heterogeneous preprocessing that limits reproducibility. Methods: We trained a Triple-Head Dual-Attention ResNet (THDA-ResNet) that processes three DCE phases (pre-contrast, early post-contrast, and late post-contrast) on the multicenter BreastDCEDL dataset (n = 1149, I-SPY trials), and we compared it with Vision Transformer (ViT) and Convolutional Vision Transformer (CvT) baselines, all ImageNet-pretrained. We benchmarked 14 preprocessing strategies, with and without N4 bias-field correction. External validation used the independent BreastDCEDL_AMBL cohort (43 lesions). AUC confidence intervals used stratified bootstrap; model comparisons used DeLong’s test. Results: THDA-ResNet achieved the highest AUC, 0.74 (95% CI 0.65–0.83), versus 0.66 for ViT and 0.63 for CvT, with the advantage reaching borderline significance over CvT (p=0.054) and not significant over ViT (p=0.14). At a threshold of 0.7, it retained discrimination (sensitivity 0.41, specificity 0.86), while transformers collapsed to near-trivial classifiers. External AUC was 0.66 (0.49–0.81). N4 correction did not improve performance. Conclusions: Attention-driven CNNs are a strong default for HER2 prediction from DCE-MRI on medium-sized cohorts, and N4 correction can be omitted, simplifying the pipeline.
This study evaluates the performance of transfer learning-based MobileNetV2 and ResNet50 architectures for binary breast cancer classification using mammography images to achieve reliable and competitive performance for breast cancer detection under realistic evaluation conditions and improve the diagnostic accuracy of mammogram classification.
Niketa Banyal, Aman Kumar· Frontiers in Medical Technol...· 0 citations
This study presents innovative approaches for diagnosing and classifying brain tumors from MRI images using advanced deep learning models to address
clinical data constraints
— including single-center acquisition, slice-level (not pixel-level) labeling, and real-world imaging variability — alongside extreme class imbalance. By combining convolutional neural networks (CNNs) with recurrent architectures and applying optimization algorithms like Adam and RMSprop, we improve detection accuracy and enhance interpretability. A dataset of 540 MRI images was collected from Hospital, covering various tumor types and patient backgrounds. Preprocessing steps such as normalization, segmentation, and texture-based feature extraction were applied to enhance data quality and model performance. A key innovation of this work is the use of a deep pre-trained model (VGG16), which demonstrates strong generalization potential for future clinical applications. Additionally, our use of real-world hospital data—more challenging than standard Kaggle datasets—makes our results more applicable in practice. Experimental results show that CNN-based models, especially VGG16 and ResNet v2, significantly outperform traditional methods. The VGG16 model achieved a classification accuracy of 97.1%, compared to 85.75% for Random Forest and 83.00% for Support Vector Machine. Crucially, these improvements reveal a critical trade-off: while macro-accuracy reaches 97.1%, Metastatic tumor recall remains at 75% — underscoring that clinical AI must prioritize equitable error distribution over aggregate metrics. Despite these advances, challenges remain, including imbalanced data and preprocessing limitations. Future research should focus on better data balancing and hybrid optimization strategies. This study contributes a robust framework for brain tumor detection, offering practical value for medical imaging and improved patient outcomes.
Breast density is associated with a higher risk of developing breast cancer and complicates mammographic interpretation because dense tissue can obscure suspicious findings. Deep learning (DL) models are increasingly used for automated breast density classification, yet their limited interpretability remains a concern in safety-critical medical applications. In this paper, we compare several DL backbones for four-class BI-RADS density classification and analyze their predictions using three heatmap explanation methods—GradCAM, GradCAM++, and ScoreCAM. Our results show that transformer-based models achieve marginally stronger classification performance than the evaluated convolutional neural network (CNN) baselines, with the ViT model obtaining the best overall results and showing improved recognition of extremely dense cases. Qualitative heatmap analysis suggests that ScoreCAM produces the most spatially coherent and clinically plausible explanations, particularly for transformer-based models for which gradient-based maps are often diffuse or unstable. We further present an exploratory procedure that converts ScoreCAM heatmaps into coarse binary masks through thresholding, illustrating the potential of explanation maps for weak localization of dense tissue. Rather than claiming a validated segmentation framework, we position this step as a proof of concept that may support future studies on annotation-efficient dense-tissue localization.
Salah Jabreel, Hatem A. Rashwan, N. Jebreel et al.· 2026 6th International Confe...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
