Automated Image Analysis of High-field and Dynamic Musculoskeletal MRI

Abstract

Abnormal knee biomechanics and injury to the menisci and posterior cruciate ligament (PCL) havebeen shown to increase the risk of developing knee osteoarthritis (OA). Recent advances in magneticresonance (MR) imaging introduced methods for in-vivo quantitative measurement of meniscus andligament morphology and biochemical integrity (water distribution and mobility) as well as 3D kneekinematics. Recent investigations have demonstrated the value of these imaging techniques for quantificationof the morphology (meniscus volume, subluxation and tibial coverage) and T2/T∗2 propertiesof the knee menisci and PCL for assessment of early changes associated with pre-osteoarthritic degeneration,or tissue health and function post-surgery. Likewise, quantitative measurements of kneekinematics such as patellar shift, tilt and cartilage contact areas from kinematic MR images have beenused to study the etiopathogenesis of pain and cartilage degradation in relation to abnormal knee function.Although these techniques can enhance assessments of the menisci, PCL and knee function, thereis currently no efficient method to extract these measurements from the MR images. Current methodsrequire significant time and resources from experts to perform manual analyses, which limits clinicalapplicability.In this thesis, we develop and validate a set of novel image analysis algorithms allowing the automatedsegmentation and quantitative analysis of the morphology and biochemistry of the medialmeniscus (MM), lateral meniscus (LM) (Aim 1) and PCL (Aim 2) from MR images of the knee joint,and the automated estimation of several important kinematic measurements (cartilage contact mechanisms,bone tracking) of the knee joint from kinematic MR images (Aim 3).An active-shape-model approach driven by a template matching process was developed to segmentthe MM and LM from MR images of the knee joint (Aim 1). Experimental MR datasets includedimages from patients with ligament and meniscus injuries (3T clinical scans) or knee OA (3T researchscans) and healthy subjects (7T MR scans). Extensive validation was performed against expert manualsegmentations using the Dice similarity index (DSI), a measure of spatial overlap. The results indicatedthat the automated method obtained accurate and robust segmentations of the MM and LM in allthe MR datasets (mean DSI between 74.5–84.3% for the MM and 76.5–85.1% for the LM). Quantitativemeasurements of the 3D morphology (volume, subluxation and tibial coverage) and biochemicalcomposition (T2-properties) of the MM and LM were automatically estimated from the segmentationvolumes. Good correlations were achieved between measurements derived from the automated andmanual segmentations of the menisci (r ≥ 0.7). Statistical comparison of these quantitative valuesacross clinically relevant groups of patients with variable knee pathologies obtained results in agreementwith the literature.A multi-atlas patch-based method was used to automatically segment the PCL in T2-maps fromhealthy and pathological knee joints (Aim 2). Quantitative validation of the method against expertmanual segmentations performed in T2-maps of healthy knee joints demonstrated good accuracy(mean DSI 74.5%). Qualitative inspections showed good PCL segmentation results in T2-maps frompathological knee joints. Correlations between the PCL T2-relaxation values derived from the automatedand manual segmentations were moderate to strong (r > 0.74).In-vivo 3D knee kinematics was evaluated automatically from MR images of the joint acquired atsix different degrees of knee flexion (“quasi-static” 3D) and in active motion (dynamic 2D+t) (Aim 3).The method extended an existing approach to segment the knee bones and cartilages from MR imagesof the joint at full extension. Two registration-based schemes were developed to align bone and cartilagesegmentations throughout the quasi-static and dynamic MR sequences. Automated segmentationsobtained throughout the quasi-static MR sequences were validated quantitatively against manual segmentations.Results showed good segmentation accuracy (mean DSI above 88.4% for the bones andabove 68.1% for the cartilages). Cartilage contact kinematics as well as quantitative measurements ofpatellar tilt and shift were estimated using features extracted automatically from the reconstructed bonesurfaces. The cartilage contact areas derived from the automated and manual segmentations showedgood spatial agreement (mean DSI above 84.0%). The correlations between quantitative measurementsestimated using the manual and automated segmentations were 0.46 ≤ r ≤ 0.93, with mostcorrelations being moderate to strong (> 0.60). Active knee kinematics estimated from the dynamicMR sequences was evaluated qualitatively.The automated MR image analysis algorithms can accurately and reliably evaluate the morphologyand T2 relaxometry of the knee menisci and PCL. This represents a considerable technical advancetowards clinical applicability of quantitative MR evaluation of these structures and can facilitate clinicalstudies. The results obtained from the automated analysis of kinematic MR images demonstrateits potential. However, further technical improvements in image acquisition and analysis are requiredbefore possible clinical applicability.