3D deep learning threat detection for real-time computed tomography baggage screening
File(s)
Author(s)
Eftychios-Angelos, Malandrakis
Type
Thesis or dissertation
Abstract
X-ray Computed Tomography (CT) baggage screening has become widely used in aviation security, offering a non-destructive 3D imaging method for identifying prohibited items. However, human-operated screening faces critical challenges: it relies on individual operators' recognition abilities, introduces time inefficiencies resulting in passenger delays, and places heavy cognitive demands on screeners that may result in missed threats. While machine learning automation shows promise, previous attempts have been limited by the complexity of 3D object analysis and insufficient annotated training data.
This thesis proposes a set of novel deep-learning solutions designed with three core objectives: statistical efficiency to work with limited annotated datasets, computational efficiency for rapid processing of large data volumes, and flexibility for easy adaptation to emerging threats.
The first approach employs a deep learning model analysing multiple 2D projections of 3D CT scans. By utilising established architectures from natural image recognition, combined with strategic image preprocessing and multi-view information aggregation, this model achieves both computational and statistical efficiency. Critically, it can detect threats even when obscured by high-density objects, overcoming a significant limitation of earlier research. An additional multi-view method introduces the domain's first few-shot learning pipeline, enabling threat recognition from as few as three to five samples per class, allowing rapid adaptation to new threat types.
Despite strong performance, these 2D-based models cannot capture essential 3D features like sharp edges. To address this gap, the thesis proposes two novel 3D architectures: a 3D convolutional neural network and a 3D Vision Transformer. To overcome the data scarcity problem, a novel 3D, contrastive self-supervised learning pipeline is developed. This method pretrains the models on vast quantities of unlabelled data collected at airports, substantially reducing the dependency on expensive annotated datasets and facilitating the creation of larger, more powerful, and generalisable 3D threat detection models.
This thesis proposes a set of novel deep-learning solutions designed with three core objectives: statistical efficiency to work with limited annotated datasets, computational efficiency for rapid processing of large data volumes, and flexibility for easy adaptation to emerging threats.
The first approach employs a deep learning model analysing multiple 2D projections of 3D CT scans. By utilising established architectures from natural image recognition, combined with strategic image preprocessing and multi-view information aggregation, this model achieves both computational and statistical efficiency. Critically, it can detect threats even when obscured by high-density objects, overcoming a significant limitation of earlier research. An additional multi-view method introduces the domain's first few-shot learning pipeline, enabling threat recognition from as few as three to five samples per class, allowing rapid adaptation to new threat types.
Despite strong performance, these 2D-based models cannot capture essential 3D features like sharp edges. To address this gap, the thesis proposes two novel 3D architectures: a 3D convolutional neural network and a 3D Vision Transformer. To overcome the data scarcity problem, a novel 3D, contrastive self-supervised learning pipeline is developed. This method pretrains the models on vast quantities of unlabelled data collected at airports, substantially reducing the dependency on expensive annotated datasets and facilitating the creation of larger, more powerful, and generalisable 3D threat detection models.
Version
Open Access
Date Issued
2025-07-23
Date Awarded
2026-01-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Chana, Deeph
Ochieng, Washington
Sponsor
Great Britain. Dept. of Transport
Connected Places Catapult
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)
Rights Embargo Date
2026-06-30