Transfer Learning vs. Training from Scratch for Breast Cancer Histology Image Classification
Janhvi Chauhan -, Dhaval Modi - · Journal of Advances in Developmental Research · 2022
We evaluated the effectiveness of transfer learning compared to fully trained networks for histopathological image classification, using three pre-trained models: VGG16, VGG19, and ResNet50. Their performance was analyzed in the context of magnification-independent breast cancer classification. Additionally, we assessed how varying the training–testing data split impacts model performance. Among the tested configurations, the fine-tuned VGG16 model combined with a logistic regression classifier achieved the highest accuracy of 93.50%, an AUC of 96.00%, and an average precision score (APS) of 96.05% using 85%–15% train–test split. Future work may explore layer-wise fine-tuning and alternative weight initialization strategies to further enhance performance.