Histo-MExNet: A Unified Framework for Real-World, Cross-Magnification, and Trustworthy Breast Cancer Histopathology

Enam Ahmed Taufika; Md Ahasanul Arafatha; Abhijit Kumar Ghoshb; Md. Tanzim Rezab; Md Ashad Alamc

cs.CVMar 15, 2026

Histo-MExNet: A Unified Framework for Real-World, Cross-Magnification, and Trustworthy Breast Cancer Histopathology

Authors: Enam Ahmed Taufika, Md Ahasanul Arafatha, Abhijit Kumar Ghoshb, Md. Tanzim Rezab, Md Ashad Alamc

Organizations: European University of Bangladesh, Department of Computer Science and Engineering, Dhaka, Bangladesh · BRAC University, Department of Computer Science and Engineering, Dhaka, Bangladesh · Ochsner Center for Outcomes Research, Ochsner Research, Ochsner Clinic Foundation, New Orleans, LA, 70121, USA

Abstract

Accurate and reliable histopathological image classification is essential for breast cancer diagnosis. However, many deep learning models remain sensitive to magnification variability and lack interpretability. To address these challenges, we propose Histo-MExNet, a unified framework designed for scaleinvariant and uncertainty-aware classification. The model integrates DenseNet, ConvNeXt, and EfficientNet backbones within a gated multi-expert architecture, incorporates a prototype learning module for example-driven interpretability, and applies physics-informed regularization to enforce morphology preservation and spatial coherence during feature learning. Monte Carlo Dropout is used to quantify predictive uncertainty. On the BreaKHis dataset, Histo-MExNet achieves 96.97% accuracy under multi-magnification training and demonstrates improved generalization to unseen magnification levels compared to single-expert models, while uncertainty estimation helps identify out-of-distribution samples and reduce overconfident errors, supporting a balanced combination of accuracy, robustness, and interpretability for clinical decision support.

Figures & tables

Magnification Level	Benign	Malignant	Total
40 $\times$	625	1370	1995
100 $\times$	644	1437	2081
200 $\times$	623	1390	2013
400 $\times$	588	1232	1820
Total Images	2480	5429	7909
Total Patients	24	58	82

Table 1: Summary of the BreaKHis dataset, detailing the number of benign and malignant histopathology images across four magnification levels.

Benign Tumors	Malignant Tumors
Adenosis – Non-cancerous proliferation of glandular structures within the breast lobules.	Ductal Carcinoma – Malignant tumor arising from epithelial cells of the milk ducts.
Fibroadenoma – Well-circumscribed benign mass composed of fibrous and glandular tissue.	Lobular Carcinoma – Invasive malignancy originating from the lobules, often with diffuse infiltration.
Phyllodes Tumor – Rare fibroepithelial lesion with variable behavior and potential for recurrence.	Mucinous Carcinoma – Malignancy characterized by abundant extracellular mucin production.
Tubular Adenoma – Benign tumor consisting of tightly packed tubular structures with minimal atypia.	Papillary Carcinoma – Malignant tumor exhibiting papillary architectural patterns.

Table 2: Overview of the benign and malignant histopathological subtypes represented in the BreaKHis dataset.

Figure 1: Representative subset of the BreaKHis dataset illustrating all eight tumor subtypes across four magnification levels ( $40\times$ , $100\times$ , $200\times$ , and $400\times$ ). Each column stands for a particular form of tumor, and each row stands for a different level of magnification. This shows how tissue morphology and cellular detail change at different scales.

Figure 2: Overview of the proposed Histo-MExNet framework.Features are extracted using DenseNet201, the most effective backbone, as part with CBAM attention. These features are mapped to a prototype-aligned space and processed by several expert classifiers. The outputs of these classifiers are adaptively fused by a gating network to produce the final prediction.

\mathcal{L}_{\text{morph}},\ \mathcal{L}_{\text{spatial}},\ \mathcal{L}_{\text{bio}}

Algorithm 1 Histo-MExNet: Multi-Expert Classification Framework

Backbone Model	Metric	40X	100X	200X	400X
VGG16	Accuracy	67%	58%	52%	52%
VGG16	F1-Score	0.59	0.42	0.26	0.37
ResNet50	Accuracy	49%	49%	53%	47%
ResNet50	F1-Score	0.36	0.32	0.38	0.33
EfficientNetB0 $\downarrow$	Accuracy	48%	49%	46%	45%
EfficientNetB0 $\downarrow$	F1-Score	0.23	0.22	0.20	0.19

Table 3: Comparison between baseline CNN architectures evaluated under Type 1 experimental settings, where models are trained and tested on the same magnification level using a random split on each type of magnification.

Panel of Figure 3 — Figure 3: t-SNE visualization comparing how well the baseline CNNs and the proposed model using the DenseNet201 backbone separate features at a $100\times$ magnification.

Backbone Model	Metric	Type 1: Same-Magnification				Type 2: Cross-Magnification
Backbone Model	Metric	40X	100X	200X	400X	40X	200X	400X
DenseNet201	Accuracy	98.25%	96.88%	92.80%	88.46%	73.08%	71.93%	51.15%
DenseNet201	F1-Score	0.983	0.969	0.927	0.881	0.716	0.690	0.393
ConvNeXt-Tiny	Accuracy	95.74%	97.84%	92.8%	90.66%	78.25 %	73.27%	51.37%
ConvNeXt-Tiny	F1-Score	0.956	0.978	0.926	0.906	0.776	0.708	0.397
EfficientNetV2-S $\downarrow$	Accuracy	95.99%	92.09%	93.55%	89.84%	67.92%	72.03%	55.22%

Table 4: Performance of the proposed Histo-MExNet framework under Same-Magnification (Type 1) and Cross-Magnification (Type 2) experimental settings.

Panel of Figure 4 — Figure 4: Comprehensive visual analysis of predictive reliability. Row 1 (a-d): Cross-magnification generalization (trained at $100\times$ , evaluated at $40\times$ ) comparing uncertainty and confidence distributions. Row 2 (e-h): Multi-magnification evaluation comparing variance and predictive confidence (U vs C) correlations for models trained and evaluated on data across all magnification levels.

Backbone Model	Accuracy	W-Precision	W-Recall	W-F1 Score
DenseNet201 $\star$	96.97%	0.9703	0.9697	0.9698
ConvNeXt-Tiny	93.49%	0.9352	0.9349	0.9347
EfficientNetV2-S	93.93%	0.9395	0.9393	0.939

Table 5: Performance metrics of the proposed Histo-MExNet framework under the Type 3 experimental setup, which combines all magnification levels for training and tests on a stratified split.

Figure 5: Confusion matrix for the DenseNet201 backbone under Type 3 experimental condition, trained on mixed data from all magnifications and assessed on a stratified test split.

Histological Subtype	40X		100X		200X		400X
Histological Subtype	mean	max	mean	max	mean	max	mean	max
Benign
Adenosis	1.44	4.6	0.96	3	1.68	4	1.16	4.8
Fibroadenoma	.94	2.4	0.86	3.4	.96	2.8	2.66	5.6
Phyllodes Tumor	0.9	3	1.12	3.6	1.2	4.6	0.58	1.2
Tubular Adenoma	1.26	3.2	1.78	3.28	0.98	3.2	1.68	5.8

Table 6: Quantitative Analysis of Occlusion Sensitivity Features by Class and Magnifications

Figure 6: Occlusion sensitivity study in Ductal Carcinoma. The rows indicate magnification levels of $40\times$ , $100\times$ , $200\times$ , and $400\times$ . Columns display the input picture, occlusion sensitivity map, and overlay. The red zones emphasize the individual tissue properties that have the most influence on the model’s categorization choice, with confidence ratings for each scale.

Method	40X	100X	200X	400X	All-Mag
ES-CNN [ 50 ] $\downarrow$	83.46%	79.62%	82.13%	78.3%	79.01%
HF+DNN [ 51 ]	90.87%	89.57%	91.58%	88.67%	-
CSDCNN [ 52 ]	89.4%	90.8%	88.6%	87.6%	-
Histo-MExNet $\star$	98.25%	96.88%	92.8%	88.46%	96.97%

Table 7: Comparative analysis of classification accuracy (%) between existing state-of-the-art methods and the proposed Histo-MExNet across four individual magnifications (

40\times,100\times,200\times,400\times

) and the combined all-magnification setting.

Variant	Accuracy (%)	W-Precision	W-Recall	W-F1
Histo-MExNet	96.97	0.9703	0.9697	0.9698
A1	93.05	0.9303	0.9305	0.9296
A2	90.49	0.9051	0.9049	0.9048
A3	91.81	0.9175	0.9181	0.9176
A4	92.62	0.9259	0.9262	0.9259

Table 8: Ablation study of Histo-MExNet evaluating the impact of key components on classification performance. The table reports Accuracy, Weighted Precision, Weighted Recall, and Weighted F1-Score for the full model and variants with specific components removed, including ensemble learning (A1), hybrid loss (A2), attention module (A3), and prototype module (A4). This analysis highlights the individual contribution of each module to the overall model effectiveness.

Explore similar work

MagViT: Interpretable Multi-Magnification Transformers with Patient-Level Model Selection for Breast Histopathology
Aug 16, 2026Nabil Ashab, Soumit Kumar Kundu, Saif Mahmud Parvez +3Histopathology Image Classification Vision Transformer
Department of Computer Science and Engineering, Dhaka International University, Dhaka, Bangladesh · Department of Microbiology, TMSS Medical College, Bogura, Bangladesh
Magnification-Invariant Image Classification via Domain Generalization and Stable Sparse Embedding Signatures
Apr 28, 2026Ifeanyi Ezuma, Olusiji MedaiyeseHealthcare Cross-Domain Generalization
Ifeanyi Ezuma, Olusiji Medaiyese
DeepHistoViT: An Interpretable Vision Transformer Framework for Histopathological Cancer Classification
Mar 12, 2026Ravi Mosalpuri, Mohammed Abdelsamea, Ahmed Karam EldalyHistopathology Image Classification Computational Pathology

Aug 16, 2026J/K move · Enter open · S save

MagViT: Interpretable Multi-Magnification Transformers with Patient-Level Model Selection for Breast Histopathology

Nabil Ashab, Soumit Kumar Kundu, Saif Mahmud Parvez, Shahadat Hossain Sohag+2

Department of Computer Science and Engineering, Dhaka International University, Dhaka, Bangladesh · Department of Microbiology, TMSS Medical College, Bogura, Bangladesh

Histopathology Image Classification Vision Transformer Explainable Medical Image Analysis Domain Generalization

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

MagViT: Interpretable Multi-Magnification Transformers with Patient-Level Model Selection for Breast Histopathology

Magnification-Invariant Image Classification via Domain Generalization and Stable Sparse Embedding Signatures

DeepHistoViT: An Interpretable Vision Transformer Framework for Histopathological Cancer Classification