Design GAN-Based Data Augmentation for Enhanced Test Tube Classification in Medical Automated Liquid Handling Systems

Synthetic images have become a part of life in recent years, moving from use in entertainment and marketing into healthcare as fields where the impact is greater. One attractive application is the production of AI-generated data in order to reduce the time and effort needed for creating large and diverse datasets. However, for effectiveness, these synthetic images need to closely represent real-life conditions. This thesis discusses a GAN-based data augmentation pipeline to enhance test tube classification in the Automated Liquid Handling System. The main challenge faced is the minimal variation in the available datasets. Here, the system generates realistic images of test tubes with the desired levels of variation in shape, size, lighting, and orientation. Multiple GANs were implemented: Pix2Pix and PatchGAN for paired image generation by applying Canny edge detection, and CycleGAN for unpaired image translation. A novel Hybrid GAN was introduced by integrating strengths from both methods, visually improving and enabling reductions in training time. Experiments with Auxiliary Classifier GAN were conducted on labeled data, albeit ineffectively due to poor label consistency. Finally, perceptual loss, based on VGG16, was integrated into Pix2Pix and Hybrid GAN to further improve the image realism. The original dataset was comprised of five categories: empty, small, medium, large, and cap; new manual and mixed datasets were introduced. Fréchet Inception Distance (FID), Inception Score (IS), and MS-SSIM confirmed the fidelity of the generated images, evaluating quality and diversity. The synthetic images were then tested using the M31 S.r.l. classification algorithm, showing increased generalization to unseen cases. The Hybrid GAN was also validated on the CIFAR-10 dataset for its broader applicability, with very promising results even with limited training. In all, this work presents the potential of structured GAN-based augmentation for addressing limitations in the datasets and improving performance in medical automation, especially within the GQ-2050 ALHS platform developed by M31 S.r.l. It also shows the increasing contribution of synthetic data in the further development of healthcare AI applications. Considering the strength of the proposed Hybrid GAN in synthesizing images.