Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Neural network on the assessment of danger of abdominal aortic aneurysms (AAA)

Siaterlis Kyriakos

Full record

---

Summary

In this thesis, we study the effectiveness of using neural networks to assess abdominal aortic aneurysms (AAA) risk by predicting maximum von Mises stress and maximum displacement at the walls of the aorta. The datasets, derived from finite-element method (FEM) simulations on geometric models reconstructed from MRI scans of real patients, come from two independent studies.We develop and evaluate two multi-output regression architectures: (a) an architecture with specialized neurons in the hidden layers (separate pathways toward each output), and (b) a fully connected feedforward network with shared neurons in the hidden-layers. In addition, we examine output normalization due to the different orders of magnitude between the two outputs, aiming to counter the training bias toward the higher scale output, as well as feature-selection strategies.Evaluation is performed with exhaustive leave-2-out cross validation, i.e., for n samples we train independent models. Experiments are run using the AdamW and Levenberg-Marquardt (LM) optimizers.Overall, AdamW exhibits more stable generalization on small/heterogeneous datasets like ours, whereas LM can offer faster convergence but is more prone to overfitting. Output normalization achieves its goal, and the architecture with specialized neurons is beneficial when it reduces interference between targets, meaning that its effectiveness depends on how the targets are produced by the FEM. Moreover, using targeted feature subsets yields significant improvements compared to using all available features, underscoring the importance of feature selection.Results on both datasets show good target prediction given the number of samples relative to the spread of the target values we aim to predict. The results highlight the potential for accessible medical applications for AAA risk assessment, but more work is needed before it can be used for clinical decision-making.

Available Files

Services