InceptionV۳-Based Breast Cancer Detection: A High-Accuracy Transfer Learning Framework for Full-Field Digital Mammography

Breast cancer can be successfully defeated if symptoms are detected in the early stages. Unfortunately, the full-field digital mammogram (FFDM) images are difficult to interpret. This study creates a deep learning framework based on the Inception V۳ transfer learning for the benign-malignant classification to determine the optimal method for the INbreast Database consisting of more than ۷۶۰۰ mammograms. The images were uniformly processed in a way to standardize the images for their intensity to be normalized. The images were also resized using a bicubic interpolation and then were augmented in a way to retain their diagnostic features, but increase the generalizability. The model uses the multi-stage dense classification head to add scale to the Inception V۳ multi-head and it focuses on features required for the detection of masses and microcalcification clusters. The model has class-weighted loss and uses stratified splits and adaptive learning rate scheduling. The model has a reproducibility for a fixed set of seeds. The model achieved ۹۷.۰۳% accuracy, ۹۸.۰% precision and also achieved an AUC of ۰.۹۹۶۵ on the test set. These scores are indicative of an almost perfect discrimination. The model error analysis showed the model had a low false negative rate, which proves it can be clinically relevant for use in screening devices. The results of the accuracy and reliability of the model confirm an Inception V۳ is a well-engineered and efficient model.

کلیدواژه ها:

Breast cancer detection ، Digital mammography ، INbreast dataset ، Deep learning ، Transfer learning ، InceptionV۳ ، Computer-aided diagnosis