Detection and Localization of Breast Lesion with VGG19 Optimized Vision Transformer

Abstract

Convolutional neural networks have been widely used in a variety of medical imaging tasks. Due to the inherent locality of convolution operation, CNNs typically perform poorly when modelling dependencies specifically long-range, which are necessary for accurately determining or recognizing corresponding breast lesion features. This motivates us to employ the Vision Transformer block along with VGG19 for the detection of breast cancer. We also offered a powerful model that effectively combines global and local features. Lastly, the model is trained independently using Database for Mastology Research and INbreast, two distinct modalities of datasets. Using transfer learning, we trained the model using data from both datasets, using 80% for training and 20% for testing. The network was trained over 100 epochs with a batch size of 50 and a learning rate of 0.01. For the INbreast and DMR datasets, the test accuracy was 98% and 89.9%, respectively. The results for the thermal image dataset are only slightly better than the very high results for the digital mammogram.