Application of a transformer-based deep learning model inbreast cancer imaging

This thesis investigates a Transformer-based deep learning model, KD-MWINet, developed for breast cancer detection using microwave imaging. The proposed model follows an encoder–decoder architecture designed to process complex-valued scattered electric field (Es) data and generate tumor probability maps. A fully synthetic dataset was created within the 1–8 GHz frequency range, using 1001 frequency points and 72 antenna positions, resulting in 10,000 scenarios with randomly assigned tumor numbers, locations, and radii. The input data were represented as real and imaginary components, while the outputs were 64×64 single-channel segmentation maps. Training was carried out with the Adam optimizer and cosine annealing scheduler, and model performance was assessed using metrics such as Dice, SSIM, PSNR, IoU, Precision, and Recall. The results demonstrated that the proposed model achieved competitive accuracy in tumor segmentation, particularly in detecting small lesions, and performed on par with state-of- the-art methods in the literature. This study highlights the potential of complex-valued neural networks enhanced with Transformer modules for biomedical imaging applications, contributing to the development of non-ionizing and cost-effective diagnostic tools.