Systematic Evaluation of Hybrid Vision Transformers and Convolutional Neural Networks for Enhanced Breast Cancer Histopathology Classification
Amesia · 30.06.2026 · 7 (1) · 72-91
DOI: 10.54559/amesia.1958087 ↗Kurz gefasst
Fachartikel in der Zeitschrift Amesia, online erschienen am 30.06.2026. Verfasst von Khaled Hejja und 1 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Gefördert von: The authors would like to thank Abdullah Gül University for providing the computational resources used in this study..
Abstract (Originalsprache)
The classification of breast cancer from histopathological slides is a critical task in oncology, requiring a balance between fine-grained cellular texture recognition and global tissue architecture. While Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have achieved high performance, their relative behavior under matched patient-wise evaluation remains important for model selection. This paper presents a systematic evaluation of five ImageNet-initialized architectures on the BreaKHis 40X dataset, including a ConvNeXt-Hybrid variant that adds one Global Multi-Head Self-Attention block to the final ConvNeXt feature map. Across the dataset-provided patient-wise train/test partitions repeated over five splits, ConvNeXt-Hybrid achieved 89.77% ± 1.62% accuracy, 0.9530 ± 0.0185 Receiver Operating Characteristic Area Under the Curve (ROC-AUC), and 0.9681 ± 0.0194 Precision-Recall Area Under the Curve (PR-AUC). This performance was statistically indistinguishable from ConvNeXt-Base (89.63% ± 2.95% accuracy; exact Wilcoxon p = 1.00), but consistently higher than Swin-Base (85.68% ± 2.67% accuracy; paired t-test p = 0.0028; exact Wilcoxon p = 0.0625). Since 0.0625 is the smallest attainable two-sided exact Wilcoxon p-value at n = 5, the Swin comparison represents maximal non-parametric evidence at this sample size. Additional analyses include computational-cost profiling, attention-head sensitivity, Gradient- weighted Class Activation Mapping (Grad-CAM)visualization, a frozen-feature multi-instance patch aggregator, within-BreaKHis magnification-shift tests, and BreAst Cancer Histology (BACH) external validation using zero-shot and frozen-head protocols. Overall, the results support a bounded conclusion: explicit global attention is useful for a window-constrained Swin backbone, but it does not provide a statistically detectable improvement over a strong ConvNeXt backbone in this setting.
Förderung
- The authors would like to thank Abdullah Gül University for providing the computational resources used in this study.
Zitieren
APA
Hejja, K., & Kurban, R. (2026). Systematic Evaluation of Hybrid Vision Transformers and Convolutional Neural Networks for Enhanced Breast Cancer Histopathology Classification. Amesia, 7(1), 72-91. https://doi.org/10.54559/amesia.1958087
Vancouver
Hejja K, Kurban R. Systematic Evaluation of Hybrid Vision Transformers and Convolutional Neural Networks for Enhanced Breast Cancer Histopathology Classification. Amesia. 2026;7(1):72-91. doi:10.54559/amesia.1958087
BibTeX
@article{Hejja2026,
author = {Hejja, Khaled and Kurban, Rifat},
title = {Systematic Evaluation of Hybrid Vision Transformers and Convolutional Neural Networks for Enhanced Breast Cancer Histopathology Classification},
journal = {Amesia},
year = {2026},
volume = {7},
number = {1},
pages = {72--91},
doi = {10.54559/amesia.1958087}
}
EUVIMED