Food Detection, Visualisation and Motion Reconstruction in MRI using Machine Learning Techniques

Abstract

Research into human digestion is important for gaining a greater understanding of how we can optimise our nutrition. However, most digestion studies focus on the organs and neglect the content undergoing digestion. The work presented in this thesis builds upon previous research aimed to close this gap in knowledge by utilising a magnetic resonance imaging (MRI) experiment involving the ingestion and classification of frozen peas. The previous research using this data lacked effective visualisation techniques and had poor classification of contents since only ten datasets are available for training and validation. Furthermore, no attempts were made to study the motion of the food. Therefore, this work attempts to provide an effective visualisation tool, an improved pea classification model, and a method to model the motion of the peas.

Firstly, we created a tool that visualised the human stomach and the ingested peas while providing facilities to correct the previous pea classification. Our tool featured visualisation and correction areas, including a 2D pea correction area and a 3D model. Aside from difficulties interpreting the data due to acquisition artefacts, our tool was found to be very easy to use. In one of our datasets, our collaborators used the tool to label 408 good, 100 bad and 113 unsure peas. In another, they labelled 234 good, 120 bad and 80 unsure peas. Overall, the software achieved its aims of visualising the peas, stomach and providing methods to correct the pea labels.

Next, we provided an improved semi-automatic pea classification model that accounted for the lack of available training data, that would have proven difficult for a state-of-the-art convolutional neural network (CNN) to be trained on. We proposed an ensemble classification model, involving an iterative human in the loop (or active learning) method. Our ensemble model to classify the peas contained quadratic discriminant analysis (QDA) and support vector machine (SVM) classifiers that take radiomic features, as well as the CNN AlexNet. Radiomic features were selected with the non-dominated sorting genetic algorithm II (NSGA-II). At each iteration, the HITL corrected the results of the previous classification. The corrected labels were then added to our training data for the next iteration, involving 5 iterations. In each iteration the HITL ensemble had a greater F1 score (0.972 - 0.982) than the individual classifiers alone (0.939 - 0.971), and had greater F1 scores than the traditional labelling approach (0.898 - 0.951), while reducing the workload for our manual labellers.

Subsequently, we provided a method to reconstruct the motion of the peas. We believe that many peas have been duplicated at different timesteps due to a time-delay in the MRI sequence. We therefore aimed to reconstruct their motion by clustering these duplicates. We chose NSGA-II to optimise the starting positions for a modified k-means using two criteria: acceptable pea velocities (range accuracy) and movement correlation (mean dot product). We compared our optimisation approach with a pure random search (PRS) with an equal processing time. We found that the optimised version was an improvement over the PRS for both the range accuracy objective (54.58% - 63.25% optimised, 50.00% - 57.25% for PRS) and the mean dot product objective (0.12 - 0.17 optimised, 0.07 - 0.11 for PRS) for each individual on the Pareto front from each dataset tested. Furthermore, we found many interesting food motion behaviours, such as collisions.

Overall, we found that the visualisation tool, semi-automatic classification and subsequent motion reconstruction provided valuable insight into the dynamics of food digestion. Future work could be done to provide an automatic segmentation of the stomach boundary, or to continue to follow the motion of contents as they leave the stomach.

Date of Award	3 Nov 2025
Original language	English
Awarding Institution	Bangor University
Supervisor	Franck Vidal (Supervisor), Panagiotis Ritsos (Supervisor) & Peter Butcher (Supervisor)