Abstract Manual detection of breast cancer in histopathology images is a highly complex task due to variations in tissue appearance and the requirement for analysis across multiple magnifications levels. The often overlooked problem in most existing approaches is the number of models required for different magnification levels. Most works in this field use separate models for each magnification, which can limit scalability and practical real-world deployment. This proposed study leverages Transfer Learning and Knowledge Distillation with transformers to develop a unified model capable of handling histopathology images across multiple magnifications. The model is trained sequentially across magnification levels, enabling the transfer and refinement of learned representations. The proposed model effectively generalizes across magnifications by prioritizing common features, achieving an average accuracy of 95.43%, average precision of 94.45%, average recall of 99.20%, average F1 score of 96.76% and AUC of 0.9930 across all magnifications of the BreakHis dataset.
Mankodi et al. (Tue,) studied this question.