Bone structure analysis in knee MRI

Specialeforsvar ved Rabia Granlund

Abstract

The pathogenesis of osteoarthritis (OA) includes a series of complex events in the whole joint. Cartilage loss and bone remodeling are central in the progression of OA.

In this paper, we investigated the feasibility of quantifying OA-characteristic bone structure from low-field MRI.

This is done in a fully automatic, uncommitted machine-learning based framework where classification is based on features selected by sequential floating forward selection (SFFS).

Six different classifiers were evaluated in the cross-validation and leave-one-out schemes, resulting in an AUC measure of the performance of the bone structure marker chosen by SFFS.

The results showed that a trabecular bone structure marker related to the presence of OA was possible to quantify.

The highest generalization AUC among the classifiers was 0.77 by the linear discriminant classifier.

The performance of the developed bone structure marker was comparable to other types of biomarkers related to OA diagnosis.

An aggregate marker based on the developed bone structure marker, a biochemical marker, and MRI cartilage markers demonstrated promising results yielding an AUC of 0.85.

This was higher than the aggregate marker without the bone structure marker, though not significantly.

In addition, the bone structure marker was able to significantly discriminate OA at different stages, which makes it a potential efficacy marker for OA treatment.

Supervisor: Professor Mads Nielsen, DIKU
Co-supervisor: Erik Dam, Nordic Bioscience
Censor: Professor Rasmus Larsen, IMM, DTU

The defense will be in Danish