Interpretable Dimensionality Reduction in 3D Image Recognition with Small Sample Sizes

A systematic feature-engineering approach to generate informative 2D representations of 3D data is introduced. In this method, the sequences of voxels along one axis of the 3D image are treated as spatial variation sequences. These sequences are projected into a 783-dimensional feature space using algorithms from statistics, signal processing, complexity theory as well as time-series forecasting and financial time-series analysis. The resulting two-dimensional image has 783 layers from which the most relevant three layers are chosen using a combination of univariate and multivariate feature selection. This process effectively converts the volumetric data into a two-dimensional three-layer image which can then be used as input to established object detection models. The validation of the method is conducted on an object detection application, involving the identification of biomatter threats in 3D X-ray scans of international travellers’ baggage. The 3D scans were recorded at the Airport in Auckland, New Zealand, and comprised 1525 biomatter threats distributed over 690 different bags. Various object detection models from the YOLO series are tested on this dataset. The YOLOv5l model achieved the highest mAP@0.5 of 0.878 on the validation dataset. Our results demonstrate that the methodologies of time-series classification and pattern recognition can be combined to implement efficient pattern recognition on 3D data sets with small sample sizes.