MagViT, an interpretable multi-magnification transformer framework with scale-gated fusion and patient-level model selection, is presented, an interpretable multi-magnification transformer framework with scale-gated fusion and patient-level model selection relative to prior ViT-centered BreakHis work.
Abstract
Breast cancer is one of the most common types of cancer among women around the world. Rapid detection and early treatment can hinder its progress to more complex stages and can impede its spread to other parts of the body. Histopathological image classification is the most common task in cancer detection due to its robustness in analyzing cellular data. Breast histopathology classification requires handling both multi-scale tissue morphology and clinically relevant generalization beyond the source domain. This paper presents MagViT, an interpretable multi-magnification transformer framework with scale-gated fusion and patient-level model selection. The model uses four BreakHis magnifications (40X, 100X, 200X, 400X) and extracts per-scale representations with a ViT backbone, and combines them via a learnable gate that masks missing scales. Patient-level five-fold cross-validation with a fixed seed has been run and compared with three architectural branches. The most accurate branch is then selected as the final model due to the strongest patient-level accuracy while retaining the simplest fusion pathway. On BreakHis, our architecture achieves a mean image accuracy of 0.9191, a mean patient accuracy of 0.9643, and a mean macro-F1 of 0.9042. External transfer experiments provide preliminary evidence of cross-dataset generalization under controlled adaptation settings on BUSI (image accuracy 0.8306, macro-F1 0.7480, patient accuracy 0.8291) and IDC (image accuracy 0.8577, macro-F1 0.8191, patient accuracy 0.8372). Grad-CAM visualization indicates that the model focuses on diagnostically significant and meaningful regions across magnifications. Relative to prior ViT-centered BreakHis work, this study emphasizes patient-level selection and cross-dataset robustness under a reproducible protocol.
Breast cancer is a significant cause of cancer-related deaths among women worldwide. Its early identification and screening are essential for improved patient outcomes and reduced mortality rates. Histopathological image analysis is considered as the gold standard for the diagnosis and prognosis of breast cancer. Nevertheless, the complexity of Whole-Slide Images (WSI) and their manual examination make this task time consuming, and prone to pathologist subjectivity. Recently, Deep Learning (DL) technology has achieved remarkable success in computer vision. However, their application still faces critical challenges in pathology analysis, including Region-of-Interest (RoI) scale variations, inter- and intra-class heterogeneity, diverse staining protocols, and the scarcity of annotated datasets. Furthermore, DL model’s findings are opaque and lack decision-level transparency. This study proposes a novel explainable multi-model DL framework for breast cancer classification leveraging histopathological images. The framework integrates Contrast Limited Adaptive Histogram Equalization (CLAHE) for image contrast enhancement, and diverse data augmentation to mitigate class imbalance and overfitting. Proposed architecture ensembles two branches, one employes DenseNet201 benefiting from Transfer Learning (TL) via ImageNet weights, while other utilizes a custom light weight attention based Hierarchal Feature Fusion (HFF) Network. DenseNet201 utilizes multilevel features to effectively tackle gradient vanishing issues and capture intricate feature representations, while HFF-Net, designed specifically for biomedical imaging, leverages HFF stem and multiscale feature extraction with Swish activation to enhance learning stability. Attention mechanism introduced within HFF-Net further refines the output features. Final feature vectors from the two branches are fused at the Global Average Pooling (GAP) layer, consolidating discriminative information. Experimental results on BRACS dataset demonstrate the proposed framework achieves 97.15% accuracy, 92.59% precision, 93.81% recall, and a 93.19% F1-score in screening tasks, while for grading tasks, it attains 84.08% accuracy, 83.39% precision, 83.64% recall, and 83.44% F1-score on 4391 test samples. Additionally, Gradient-Class Activation Mapping (Grad-CAM) saliency heatmap are generated for visual representation of proposed model’s choices, thereby increased transparency. The integration of these advanced techniques significantly enhances diagnostic reliability, addressing the challenges in histopathological image analysis.
Muhammad Nabeel Mehmood, Muhammad Hassaan Ashraf· Informatica· 0 citations
Breast cancer remains a global health challenge, representing one of the most prevalent and deadly cancers worldwide. Histopathological analysis of biopsy specimens is the gold standard for diagnosis, but the increasing demand for assessments often surpasses the capacity of available pathologists. To address this challenge, this study proposes a lightweight framework for breast cancer classification from whole-slide images (WSIs) that decouples feature extraction, tile-to-slide aggregation, and classification. Tile-level descriptors are extracted using a pre-trained VGG19 network modified with a configurable max pooling layer that controls descriptor dimensionality. The resulting tile embeddings are summarized into compact slide-level representations by Incremental Uniform Attention (IUA), a non-parametric, permutation-invariant aggregation operator with linear time complexity and no trainable parameters. The slide descriptors are then classified using six conventional machine learning models. The framework was evaluated on $3{,}241$ WSIs from the TCGA-BRCA and HCMI-CMDC cohorts using a nested cross-validation protocol. In its most compact configuration, the pipeline reduced the stored feature representation from approximately 1.7 TB to 64 MB, corresponding to a reduction of more than $26{,}000\times $ , while preserving high predictive performance. Support Vector Machine yielded the most stable results across pooling configurations, whereas Logistic Regression achieved the highest overall performance in the most compact setting, with 99.7% accuracy, 99.8% F1-score, and 99.99% AUC. A theoretical analysis further shows that IUA has linear time and auxiliary-memory complexity with respect to the number of tiles per slide, in contrast to the quadratic complexity of self-attention-based Multiple Instance Learning (MIL) aggregators. These results indicate that compact, parameter-free aggregation can substantially reduce computational and storage costs, enabling large-scale WSI cohort analysis on conventional hardware.
Wysterlânya K. P. Barros, Raquel de Melo Barbosa, Marcelo A. C. Fernandes· IEEE Access· 0 citations
The proposed IMIL-Net provides a high-accuracy, interpretable, and clinically-aligned diagnostic solution by analyzing complete patient examinations and providing a statistically validated decision-making process, and represents a trustworthy tool for integration into the clinical otolaryngology workflow.
Biao Xu, Miao Zhang, Shuai Jiang et al.· European archives of oto-rhi...· 0 citations
Background/Objectives: Deep learning approaches have demonstrated strong performance in breast histopathology image classification; however, reliable generalization across heterogeneous acquisition environments remains challenging due to domain shift. In clinical practice, missed malignant cases are particularly critical because they may directly affect diagnostic decisions and patient outcomes. This study systematically investigates the behavior of modern deep learning architectures and adaptation strategies under realistic cross-domain conditions, with particular emphasis on malignant case detection and false-negative reduction. Methods: Three modern architectures—ConvNeXt-Tiny, Swin-Tiny, and MaxViT-Tiny—were initially trained on a large-scale breast histopathology dataset and subsequently evaluated on the BreaKHis dataset using strict patient-level separation to avoid information leakage. Three transfer settings were investigated: direct zero-shot transfer, head-only adaptation, and full fine-tuning. Performance was evaluated independently across four magnification levels (40×, 100×, 200×, and 400×) using accuracy, precision, sensitivity, F1-score, ROC–AUC, PR–AUC, and false-negative rates. Results: Direct zero-shot transfer produced substantial performance degradation across all architectures, with mean false-negative rates ranging from 75.85% to 90.11%, highlighting the limited transferability of source-domain representations under heterogeneous acquisition conditions. Both adaptation strategies substantially improved performance and reduced missed malignant cases to below 10%. Swin-Tiny under head-only adaptation achieved the most favorable malignant detection profile, reaching a mean sensitivity of 97.36% while reducing the average false-negative rate to 2.64%. In contrast, MaxViT-Tiny achieved the highest mean ROC–AUC value (0.849) after full fine-tuning, although this did not correspond to the lowest false-negative burden. Conclusions: The findings demonstrate that maximizing global discrimination performance does not necessarily correspond to optimal malignant detection under cross-domain conditions. Sensitivity and missed-case analysis provide complementary information beyond conventional discrimination metrics and may support more informed model assessment. Furthermore, the proposed methodology provides a reproducible framework for investigating adaptation performance in AI-assisted breast histopathology systems.
L. Stanescu, Cosmin Stoica Spahiu· Diagnostics· 0 citations
Histopathological imaging of breast cancer remains a vital
clinical task. While deep learning techniques have shown strong diagnostic accuracy, most existing models suffer from poor interpretability, high computational demands, and limited clinical transparency. Therefore, this research aims to develop a simple, streamlined, and interpretable hybrid system for early, accurate, and reliable breast cancer detection to support real-time
clinical diagnosis.
A new hybrid architecture, Light-PSO-Net, was proposed by integrating a lightweight
convolutional neural network (MobileNetV2), Particle Swarm Optimisation (PSO) for global
hyperparameter optimisation, Principal Component Analysis (PCA) for dimensionality reduction, and a Fuzzy Inference System (FIS) to stratify risks in an interpretable manner. The publicly available BreaKHis dataset of breast cancer histopathology images at multiple magnifications, comprising benign and malignant tissues, was utilised to train and test the model. GradCAM visualisation highlighted spatial features used for classification, while PCA-FIS offered
semantic interpretability of the results. Model performance was evaluated using accuracy, F1-
score, confusion matrices, and ablation analyses.
The proposed framework achieved 95.5% classification accuracy and an F1 score of
0.9682, outperforming many other modern lightweight and hybrid deep learning systems. The
ablation analysis showed that PSO outperformed PCA in terms of convergence stability, predictive performance, and feature compactness and separability. The fuzzy reasoning module enhanced sensitivity to malignancy and provided categorisation of risk into low, medium, and
high risk. The model also demonstrated consistent results across different magnifications and
external validation conditions.
The combination of global maximisation, dimension minimisation, and rule-based
solutions helped the proposed model overcome key challenges in breast cancer diagnosis, such
as computational efficiency, interpretability, and clinical relevance. Unlike traditional black-box
models, Light-PSO-Net provides both pixel-level and decision-level explainability using GradCAM and fuzzy logic. Its lightweight design makes it suitable for resource-limited pathology
laboratories and edge healthcare environments. These findings emphasise the importance of integrating deep learning with explainable and optimisation-based systems for translational medical applications.
Light-PSO-Net provides a genuine, effective, and interpretable approach for classifying histopathological breast cancer. The proposed hybrid learn-then-reason pipeline shows
strong potential to assist pathologists in early diagnosis and improve clinical outcomes. Future
research will focus on multimodal integration, federated learning, adaptive fuzzy systems, and
multi-centre validation to enhance generalisability and support wider clinical adoption.
N. Mishra, Praveen Kumar· Recent Advances in Computer...· 0 citations
Breast cancer (BC) is one of the harsh cancers which lead to severe health issues worldwide. Premature detection and treatment implementation is important for treating the BC with a suitable clinical handing process. Proposed research implements a deep-learning (DL) tool for identifying benign/malignant BC from the histopathology data. This work considered benign/malignant class image data for examination and outcome is verified to substantiate the developed system's merit. Various phases found in this scheme includes; (i) image augmentation, colour normalization and resizing, (ii) features extraction and image classification with softmax, (iii) best features identification and fused-features generation, (iv) performance confirmation with fused-features along with 3-fold cross validation. Proposed work implemented conventional and lightweight DL-models for verifying its performance and the experimental outcome confirms that the developed system provides up to >99% accuracy. The performance of lightweight and conventional models is separately evaluated and this research confirms that the conventional scheme offers a better result.
A. S. Vickram, Kavitha, D. P· 2026 11th International Conf...· 0 citations
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
