Longitudinal CT scanning for explainable early detection of postharvest disorders: The ‘Braeburn’ browning case

Longitudinal computed tomography (CT) datasets quantify the internal state of agricultural products in 3D over time, making them very suitable for studying the progression of postharvest disorders. In this paper, we present a workflow for developing early detection systems using longitudinal CT datasets. In our workflow, each CT-voxel is treated as a time series, which facilitates the analysis and detection of gradually progressing disorders in four different ways. Firstly, visualizing the difference between two CT scans provides higher contrast visualizations to study the progression of a disorder. Secondly, features derived from regional changes can be more discriminative than features derived at one time point over the whole product. Thirdly, detection setups can be simulated at different scanning times, making it possible to compare potential factory setups. Fourthly, longitudinally explainable artificial intelligence answers a specific research question about the decision-making of a neural network. The workflow is showcased on a dataset of CT scans of 80 ‘Braeburn’ apples that developed core browning. We provide a detailed visualization of the progression of core browning during and after controlled atmosphere (CA) storage. Moreover, a deep neural network classifier was developed that had more than 90% accuracy in the early detection of browning. The presented workflow provides powerful tools to study and automatically detect disorders that develop over time, thereby aiding in waste reduction and marketability.