Advanced Deep Learning Approaches for Predicting Breast Cancer Genes Using Spectrographic Data

Abstract

Breast cancer is the most frequently diagnosed cancer among women in urban regions of India, such as Mumbai, Delhi, Bengaluru, Bhopal, Kolkata, Chennai, and Ahmedabad, accounting for 25% to 32% of all female cancer cases. To address this growing concern, our study leverages advanced deep learning techniques to enhance the automated prediction of breast cancer-associated genes using spectrograms. Gene data, collected from reliable sources like the National Center for Biotechnology Information (NCBI), are transformed numerically using frequency-of-occurrence mapping and the VOSS representation method, which employs binary sequences and long-range fractional correlation analysis. These numerical representations are converted into spectrograms through Short-Time Fourier Transform (STFT), enabling their analysis using deep learning models, including 1D Convolutional Neural Networks (1DCNN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Graph Neural Networks (GNN). Among these, GNN achieved the highest accuracy of 97.97%, followed by RNN with 93.88%, and CNN and 1D-CNN with 91.02% and 89.38%, respectively. These results highlight the potential of GNN as a precise and reliable tool for breast cancer gene prediction, offering significant promise for clinical applications.