Image Synthesis for Prostate Cancer Biopsies Using Conditional Deep Convolutional Generative Adversarial Network
Paul Kariuki, Patrick Gikunda, John Wandeto · 2025
Deep learning models have shown promising results in complementing computer-aided diagnosis. However, the availability of quality medical image data sets is inhibited by the limited and imbalanced data sets available, privacy issues, and the high cost of generating labeled medical imaging data sets. When a deep learning model is trained on such data, its performance is greatly impeded, with a high chance of overfitting, and in turn, limiting the model's ability to scale and generalize well to previously unseen data. More recently, empirical studies have set the stage for image synthesis as a solution to these problems using generative adversarial neural networks. This paper presents a novel conditional DCGAN capable of growing synthetic prostate cancer biopsy whole slide images augmented from the PANDA histopathology data set. From the results, the novel Conditional Deep Convolutional Generative Adversarial Network model achieved an average Fr échet Inception Distance score of 1.3 with high quality synthetic images. The results validated the model's exceptional ability to generate realistic images with high discriminatory accuracy and low generator loss.