Synthetic data for X-ray CT of healthy and disordered pear fruit using deep learning

Abstract

Over the last years, deep learning (DL) models have led to an enormous breakthrough in a wide range of computer vision tasks, including the classification and quantification of internal defects in fruit using X-ray CT images. However, developing these models typically requires large annotated datasets. Obtaining such datasets faces multiple challenges as the availability of defect fruit is unpredictable and it is unclear whether the disordered fruit in the dataset are representative of all possible defects. This work proposes a method to circumvent these problems by generating synthetic CT images using a conditional Generative Adversarial Network (cGAN) that learns the underlying distribution of gray values for different classes (“healthy tissue”, “internal browning”, “background, core and cavities”) in manually annotated images of pear fruit. The performance of the predictor was evaluated by quantitative metrics and visual inspection. To generate new synthetic data, operations are suggested and tested on annotated data to create as much variation as possible in the generated pears. The Fréchet Inception Distance (FID) and visual observations showed that the cGAN was effective in image-to-image translation, generating CT images of both healthy and disordered fruit based on their annotations. The best performing model obtained a low FID score of 10.2 and produced detailed visual features in the synthetic CT data, such as vascular bundles. The high potential and broad applicability was further demonstrated on augmented image annotations for generating new healthy and disordered pears with different appearances of defects. The proposed method can be used to produce additional training data when datasets are limited and laborious to manually annotate. Furthermore, the robustness of other DL models might increase by including new synthetic samples with a high variety of disorders during training.