How to label CT training data precisely (fast)

Deep-learning-based segmentation pervades the world of non-destructive evaluation. It enables the inspection of challenging CT scans: CT scans of parts made from materials that are hard to penetrate, fast CT scans that have to meet tight cycle times in production, or CT scans of multi-material components with heavily varying absorption coefficients. A crucial part in designing deep-learning-enabled systems is training the deep segmentation model, which requires a set of consistently and precisely labeled data – for training and for validation. Labeling data is always a tedious endeavor. With industrial CT scans, it becomes even harder: On the one hand, we need to deal with three-dimensional image data, which is hard to grasp, and, on the other hand, we are interested in finding features that have no inherent shape, like cavities or material detachments. Luckily, there are some tools and techniques that can simplify the labeling process in different scenarios. (i) While it is mandatory to use actual production data for training, you might want to use high-quality data for labeling. (ii) Combine different modalities like a CAD representation of the part under inspection with the CT scan to separate individual components. (iii) You don’t have to start deep. Train a shallow classifier that operates on simple image features to create a per-voxel classification. (iv) Iterate your way out to a perfectly labeled dataset, training a first deep segmentation model right after labeling the first CT scan. (v) Of course, you benefit the most when you combine different approaches. Note that this list is by no means exhaustive and is intended to spark some ideas on how you can simplify the unloved task of labeling machine learning data.