Diagnosis of breast malignancy with microwave breast imaging method and utilizing EfficientNet-B4 deep learning model

In this literature, an experimental study on non-invasive microwave breast imaging technology has been proposed to diagnose breast malignancy at its primary stage. The work utilizes a novel methodology to identify small tumors present in the female breast showing benign or malignant symptoms. The proposed antenna, designed for this imaging application, features a defected ground structure and exhibits multi-resonance behavior. It functions within 2 to 12 GHz frequency range that is suitable for medical imaging applications. The performance of the model has been tested with fabricated breast phantom models mimicking the dielectric properties of biological breast tissues. Two identical antennas are located around the phantom in closer proximity to acquire reflected and transmitted signals in the form of scattering parameters. The obtained dataset has been converted to high resolution two-dimensional Scalogram images utilizing proficient signal processing with continuous wavelet transform that provides both spatial and temporal information of the signal. The images are exploited with fine-tuned EfficientNet-B4 model to classify benign and malignant tumors. The issue of overfitting caused by a small experimental dataset has been addressed through the use of image augmentation and transfer learning techniques. Using compound scaling, the proposed network creates an optimal and balanced network model that achieves 99.56% accuracy and 99.40% precision in classifying benign and malignant breast phantom images. Overall, the model outperforms other reported CNN models while being compared in terms of its performance accuracy.