A Hybrid Architecture for Benign-Malignant Classification of Mammography ROIs

Abstract: Accurate characterization of suspicious breast lesions in mammography is important for early diagnosis and treatment planning. While Convolutional Neural Networks (CNNs) are effective at extracting local visual patterns, they are less suited to modeling long-range dependencies. Vision Transformers (ViTs) address this limitation through self-attention, but their quadratic computational cost can be prohibitive. This paper presents a hybrid architecture that combines EfficientNetV2-M for local feature extraction with Vision Mamba, a State Space Model (SSM), for efficient global context modeling. The proposed model performs binary classification of abnormality-centered mammography regions of interest (ROIs) from the CBIS-DDSM dataset into benign and malignant classes. By combining a strong CNN backbone with a linear-complexity sequence model, the approach achieves strong lesion-level classification performance in an ROI-based setting.

• The paper introduces a hybrid model that combines local CNN feature extraction (EfficientNetV2-M) with a global SSM (Vision Mamba) for improved lesion classification.
• The proposed architecture achieves an AUC of 0.875 and demonstrates a +2.5% AUC improvement over CNN-only models.
• The study employs rigorous preprocessing and balanced training protocols on CBIS-DDSM to accurately distinguish benign and malignant mammography ROIs.

Hybrid CNN–State Space Modeling for Mammography Lesion Classification

Introduction

Accurate discrimination between benign and malignant breast lesions in mammography is a persistent challenge, compounded by the limitations of both human interpretation and classical computer-aided diagnostic systems. While deep learning, especially convolutional neural networks (CNNs), has advanced automated feature extraction in medical imaging, these architectures remain limited in modeling long-range contextual dependencies due to the locality of convolutional receptive fields. Vision Transformers (ViTs) address global context via self-attention mechanisms but suffer from quadratic scaling, making them inefficient for high-resolution medical images. The paper "A Hybrid Architecture for Benign-Malignant Classification of Mammography ROIs" (2604.12437) introduces a hybrid paradigm, integrating EfficientNetV2-M for local feature acquisition with Vision Mamba—an efficient State Space Model (SSM) for global context integration—to achieve improved lesion-level classification of mammography regions of interest (ROIs).

Dataset, Preprocessing, and Task Formulation

The study utilizes the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM), focusing on abnormality-centered ROIs rather than full-field mammograms. The binary classification task consolidates all malignant lesions into one class and benign lesions into another, concentrating on lesion-level discrimination—a more granular classification scenario than full-image screening. Preprocessing is rigorous: image crops are resized, normalized to ImageNet statistics, and lightly augmented through flips and rotations to induce invariance to minor transformations. Stratified train/validation/test splits are established at the patient level, with class imbalance addressed via weighted loss functions.

Hybrid Architecture: Local-Global Feature Synergy

The central contribution of the paper lies in its hybridization of high-performance CNN feature extraction and efficient, scalable global context modeling. The architectural design draws upon the observation that radiological assessment often progresses from examining local lesion morphology to considering broader anatomical context.

EfficientNetV2-M operates as the local feature extractor, leveraging Fused-MBConv blocks for balanced accuracy, parameter efficiency, and fast throughput. The model is truncated before classification, and feature activations from the deepest convolutional layer serve as the foundation for downstream modeling—yielding spatial feature maps encapsulating rich local information.

Vision Mamba, a bidirectional State Space Model, processes patchified, flattened feature maps for long-range dependency modeling. By operating at linear time and space complexity with respect to sequence length, Vision Mamba allows for practical exploitation of global context in high-resolution medical images where Transformer-based architectures would be either prohibitive or require undesirable downsampling.

The sequential fusion is as follows: features from EfficientNetV2-M are patchified and embedded with learned positional encodings before being processed by stacked Vision Mamba blocks. The resultant sequence is globally averaged and passed through a fully connected sigmoid classifier, outputting malignancy probability per ROI.

Figure 1: Vision Mamba architecture, illustrating bidirectional state space modeling for efficient global context capture.

Figure 2: Proposed hybrid architecture combining EfficientNetV2-M for local feature extraction and Vision Mamba for global context modeling.

Experimental Protocol

Transfer learning is employed, initializing EfficientNetV2-M with ImageNet weights (frozen during an initial training phase, then fine-tuned). The Vision Mamba stack and classification layers are trained from scratch. Optimization utilizes AdamW for stability, leveraging class-balanced binary cross-entropy loss to counteract dataset imbalance. Learning rates are managed via cosine annealing with warm restarts, and performance monitoring employs standard metrics including AUC-ROC, sensitivity, specificity, F1-score, and accuracy.

Empirical Results and Analysis

The hybrid model outperforms a spectrum of established baselines—CNNs (VGG-16, ResNet-50, EfficientNet-B3, DenseNet-121), ViT, and two-view EfficientNet variants—for ROI-based classification on the CBIS-DDSM test split. Notably, it achieves an AUC of 0.875, accuracy of 94.2%, sensitivity of 0.89, specificity of 0.95, and an F1-score of 0.90. The ablation study indicates that the addition of Vision Mamba to the EfficientNetV2-M backbone yields a +2.5% AUC gain, underscoring the benefit of explicit long-range context modeling: standalone Vision Mamba lags both hybrid and CNN-only models, but its synergistic integration with CNN features outperforms all baselines considered.

Key numerical results and claims:

• AUC of 0.875: Exceeds all CNN and Transformer-based single-view baselines in the evaluated setting.
• Ablation improvement of +2.5% AUC over EfficientNetV2-M alone: Explicitly quantifies the impact of global state space modeling given identical local feature representations.
• Validation that bidirectional SSMs provide complementary information to conventional CNN processing in medical ROI classification.

Implications and Future Directions

Practically, this work demonstrates that SSM-based global modeling is a viable and computationally efficient alternative to Transformers for high-resolution medical imaging. The framework is directly extensible to other modalities or imaging tasks where large spatial context is important but quadratic self-attention is computationally infeasible. Theoretically, the observed synergy between dedicated local (CNN-based) and global (SSM-based) modules supports the perspective that architectural modularity—a principle gaining traction in vision-language modeling—can yield statistically meaningful improvements in constrained-data medical domains.

Potential avenues for future work include:

• Extension to screening-level scenarios employing full whole-image input and multi-view fusion.
• Exploration of Mamba model variants (e.g., MedMamba) tailored for imaging.
• Incorporation of richer clinical priors or multimodal input streams to further enhance lesion characterization and reduce false positives.
• Systematic investigation of model calibration and external dataset generalization, supporting clinical deployment protocols.

Conclusion

The hybrid EfficientNetV2-M + Vision Mamba architecture delivers state-of-the-art lesion-level classification performance on CBIS-DDSM mammography ROIs by integrating local CNN feature extraction with efficient SSM-based global context modeling. The demonstrated improvements over both CNN and Transformer baselines reinforce the value of architectural hybridization for medical image analysis. Future research should move beyond the ROI regime, targeting more complex, real-world screening applications and further modular architectural developments.