DeepFake Technology for Breast Cancer Dataset Generation Using Autoencoders and Deep Neural Networks

Suzan Anwar¹*, Shereen Abdullah², Dalya Ali³, Mardin Anwer³ and Daniah Al-Nadawi^4
1Philander Smith University, LittleRock, Arkansas, USA
²Erbil Polytechnic University, Erbil, Iraq
³Salahaddin University, Erbil, Iraq
⁴Medical University of Lublin, Lublin, Poland
*Corresponding Author: Suzan Anwar, Philander Smith University, LittleRock, Arkansas, USA.

Published: March 07, 2024

View Pdf

Abstract  

In the emerging field of radiogenomics, the primary challenge is the high cost of genetic testing, which restricts access to large, paired datasets of imaging and genetic information. Such datasets are essential for the effective training of machine learning algorithms in radiogenomic analyses. This research aims to bridge the gap between gene expression in tumors and their morphological representation in MRI scans of breast cancer patients. In this work an advanced autoencoder for processing gene expression data, and the derived weights from this autoencoder utilized were then employed to initialize a supervised Deep Neural Network (DNN). This network extracted distinct morphological markers from each MRI scan. This study introduces an innovative approach that utilizes deepfake technology, employing dual Generative Adversarial Networks (GANs) to generate synthetic imaging data from a radiogenomic dataset. This synthetic data, nearly indistinguishable from real data, is produced using a supervised neural network and is aimed at enhancing breast cancer diagnostics. Notably, the proposed neural network, when enhanced with an autoencoder and dropout techniques, demonstrated superior predictive accuracy over linear regression models. Specifically, it reduced errors by an average of 1.8% in mean absolute percent error. These findings underscore that the images generated by the proposed model are virtually indistinguishable from authentic images and exhibit high reliability in applications through the PyTorch framework. The results of this study underscore the potential of the proposed methodology to significantly contribute to advancements in breast cancer diagnostics.

Keyword: Generative Adversarial Networks (GANs); Autoencoder; Radiogenomic; DeepFake; Supervised Deep Neural Network DNN

.