Table 1 Summary of related studies on breast cancer classification and diagnosis.

Golatkar et al.²³

Fine-tuned Inception-v3 CNN with patch extraction based on nuclear density

Targeted extraction of informative regions, improved classification accuracy

Accuracy: 85% (average), 93% (non-cancerous vs. malignant)

Spanhol et al.²⁴

CNN-based classification using image patches from BreaKHis dataset

Significant improvement in recognition rates

Patient-level: 90%, Image-level: 89.6%

Motlagh et al.²⁵

Pretrained ResNet50V1 and ResNet152 architectures

High accuracy in benign/malignant classification

ResNet50V1: 99.8%, ResNet152: 98.7%

Vang et al.²⁶

Inception V3 with majority voting ensemble

Improved performance over state-of-the-art methods

12.5% improvement over prior methods

Xiang et al.²⁷

Data augmentation and fine-tuning on BreaKHis dataset

Suppressed overfitting, improved generalizability

Patient-level: 97.2%, Image-level: 95.7%

Zhou et al.²⁹

Resolution-adaptive network (RANet) + ADSVM

Excluded mislabeled patches, optimized computational efficiency

Multiclass: 97.75%, Binary: 99.25%

Vesal et al.³⁰

Transfer learning with Inception-V3 and \(\text {ResNet50}\)

Addressed color variations, robust classification

Inception-V3: 97.08%, \(\text {ResNet50}\): 96.66%

Vizcarra et al.³¹

Fusion of SVM (shallow learner) and CNN (deep learner)

Improved accuracy through model fusion

Fused model: 92%

Kone et al.³²

Hierarchical CNN system for classifying BC pathologies

Automated hierarchical classification, high accuracy

BACH: 99%, Extension dataset: 96%

Thomas et al.³³

ViTs on BreaKHis dataset

Stable performance across magnifications

Accuracy: 96%

Zou et al.³⁴

DCET-Net (CNN + Transformer)

Combined local and global feature modeling

Image-level: 98.79%, Patient-level: 98.77%

Wu et al.³⁵

ScATNet (multiscale self-attention-based network)

Comparable to human pathologists, real-world applicability

Performance comparable to 187 pathologists