Abstract
Breast cancer is a leading cause of cancer-related deaths within the female population, making precise and timely diagnosis essential. Histopathological image analysis is considered the gold standard for diagnosis, yet is time-consuming and relies on expert pathologists. While prior work has generally studied individual CNN or transformer architectures in isolation, a unified, head-to-head comparison of state-of-the-art CNN and transformer models under identical preprocessing, training, and evaluation conditions remains limited. To address this gap, this study presents a controlled comparative evaluation of five deep learning architectures—ResNet50, DenseNet121, EfficientNetB0, InceptionV3, and Vision Transformer (ViT)—for binary (benign vs. malignant) breast cancer histopathological image classification using the BreakHis dataset. The novel contribution of this work is threefold: (i) all five architectures are benchmarked under an identical transfer-learning and fine-tuning protocol, removing confounds introduced by differing preprocessing or training regimes in prior studies; (ii) model performance is assessed using a complete set of metrics (accuracy, precision, recall, F1-score, and AUC) together with statistical significance testing between the top-performing models; and (iii) the generalizability of the benchmark is discussed in light of the dataset’s demographic and institutional composition. Experimental results showed that Vision Transformer achieved the best performance, attaining 95.24% accuracy and an AUC of 0.98, with EfficientNetB0 delivering the strongest results among CNN-based models while maintaining computational efficiency. The findings demonstrate the effectiveness of transformer-based architectures for capturing complex histopathological patterns and provide a benchmark comparison of CNN and transformer models for intelligent breast cancer diagnosis.
KEYWORDS
Breast Cancer Classification, Histopathological Image Analysis, BreakHis Dataset, Transfer Learning, Vision Transformer.
Sajal Kumar Kar1*, Achyut Pandey2
1Dept. of Computer Science & Engineering, APS University, Rewa-486003, M.P., India
²Dept. of Physics, TRS College, Rewa-486001, M.P., India