Multimodal deep learning for diagnosing sub-aneurysmal aortic dilatation
Typ
Examensarbete för masterexamen
Program
Publicerad
2019
Författare
FINATI, SARA
LJUNGGREN, ELIN
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Abdominal Aortic Aneurysm (AAA) is a localized enlargement of the abdominal
aorta that can progress to a rupture, which will cause an internal bleeding that is
fatal in the majority of the cases. To save more lives, a nationwide screening program
invites all men at 65 to measure their largest aortic diameter during an ultrasound
examination. The diagnosis is based solely on this diameter and if it is below 30
mm the patient is declared healthy and discarded from any follow-up monitoring.
However, recent studies have shown that patients with a diameter within 25–29 mm,
a sub-aneurysm, run an elevated risk of developing a full-size aneurysm and hence
might need further surveillance.
This thesis is a collaboration between the product development company QRTECH
AB and Västra Götalandsregionen (VGR). It proposes a novel solution for predicting
which sub-aneurysms that might grow into a full-size aneurysm with the help of an
ultrasound image complemented by patient data including aortic diameter, number
of years of smoking and snus, blood pressure and medications. The solution consists
of a multimodal deep learning algorithm that classifies the sub-aneurysms as either
healthy or sick and thereby suggests patients that should be kept under surveillance.
Due to lack of any follow-up data for the men with sub-aneurysms a comparison
with meta-studies, examining how many sub-aneurysms that progressed into a fullsize
aneurysm, was carried out. The results from those studies did not agree with
the results obtained from classifying the sub-group in this project. A feasible explanation
is the limited data set which most likely affected the learning. However, the
evaluation of the model’s performance was still promising and indicates the potential
of using neural networks for diagnosing AAA.
Beskrivning
Ämne/nyckelord
multimodal deep learning , abdominal aortic aneurysm (AAA) , subaneurysmal aortic dilatation , VGG19 , Keras , heatmaps , permutation importance