Computer Aided Detection and Measurement of Peripheral Arterial Diseases from CTA Images Professor Jamshid Dehmeshki Kingston University, London The University Hospital of Lausanne
Outline ➢ Anatomical overview of Peripheral Arteries Diseases (PAD) ➢ Investigation of PAD in Computed Tomography Angiography (CTA) ➢ Methodology – Automatic Computer Aided Detection (CAD) and Automatic Computer Aided Measurement (CAM) of PAD ➢ Evaluation – computational implementation and evaluation ➢ Conclusıon
Anatomical overview of PAD What is Peripheral Arterial Disease (PAD)? • Obstruction of arteries in lower extremities • Afflicts more than 2.7 million people in the U.K and 10 million of Americans per year
Anatomical overview of PAD
Anatomical overview of PAD Research motivation ➢ PAD has a mortality rate higher than breast cancer . ➢ PAD is a marker of coronary artery disease (CVD) , ischemic heart disease (IHD) and cerebrovascular disease ➢ The disease is undertreated and under detected
Anatomical overview of PAD Treatment • Minimizing the risk factors of atherosclerosis • Pharmacological therapy o Antiplatelet, statins
Anatomical overview of PAD Treatment • Minimizing the risk factors of atherosclerosis • Pharmacological therapy o Antiplatelet, statins • Endovascular o Angioplasty o Stent placement o Atherectomy
Anatomical overview of PAD Treatment • Minimizing the risk factors of atherosclerosis • Pharmacological therapy o Antiplatelet, statins • Endovascular o Angioplasty o Stent placement o Atherectomy • Surgical o Endarterectomy o Peripheral bypass grafting o Amputation
Anatomical overview of PAD Diagnosis Imaging techniques • Duplex ultrasound scanning • Magnetic resonance angiography (MRA) • Transcatheter angiography (TCA) • Computed Tomography Angiography(CTA)
CTA for PAD Statement of the problem ➢ CTA a common imaging technique for investigation of the disease Current problems ➢ Large amount of data • Confined to two-dimensional (2D) views ➢ Variety of arteries size and shape, abnormalities ➢ Image quality: parameters of CT scanning, contrast agent, PVA always present Radiologist ➢ High inter-observer and intra-observer variability ➢ Error, fatigue vs. performance ➢ Cost
Scope and aim of the project CAD - CAM ➢ A system that can fully detect and measure peripheral arterial diseases in CTA datasets automatically ➢ A tool for radiologist to reduce investigation time, human error improving clinical assessment
Methodology Proposed Methodology
Methodology ➢ Automatic seed selection ➢ I - the input Dicom image ➢ l f0 - initial foreground value ➢ lb0 - background value ➢ Region growing criteria: ➢ Optimum threshold is - Connectivity among voxels calculated iteratively - Intensity homogeneity
Methodology ➢ Erosion is not sufficient to ➢ Over segmentation from previous delineate areas of continuity step (region growing): artery and bone between bone and artery are connected ➢ Use a prior information and spatial information - Variations in bone density - Arteries intensity similar with bone ➢ Anatomical prior: properties of or soft tissue artery and bones - bones are larger structures - bones show inhomogeneous intensity values - arteries have more homogeneous intensity(exception: calcifications)
Methodology False Positive Reduction ➢ Erosion is not sufficient to delineate areas of continuity between bone and artery ➢ Use a prior information and spatial information ➢ Anatomical prior: properties of artery and bones - bones are larger structures - bones show inhomogeneous intensity values - arteries have more homogeneous intensity(exception: calcifications)
Methodology ➢ We have extracted the peripheral arteries and we want to investigate the presence of stenosis ➢ Requires a centreline that preserves geometry and topology
Methodology ➢ Centreline contains spurious branches and redundant voxels ➢ Removal of these unwanted voxels is done using: • Pruning: A 3D directional connectivity search to remove branches • Refinement: remove redundant voxels ➢ ➢ Create continuous smooth line • Interpolation using Catmull Rom splines
Methodology ➢ The obtained centreline is a discrete representation of voxels ➢ Artery measurements are performed in cross section images ➢ Cross section images are orthogonal planes to the arteries ➢ Create continuous smooth line • Interpolation using Catmull Rom splines ➢ Cross section images are obtained using a sliced- based rendering technique
Methodology ➢ We have created the cross section images where the measurement can be performed ➢ Area ➢ Equivalent Diameter
Methodology ➢ The vessel profile is a 1D representation of the measurements ➢ Looking for significant extrema points ➢ Maxima (p max )and Minima (p min ) are chosen ➢ A stenotic area is automatically identified if
Methodology ➢ A MAP-MRF expectation maximization method is used ➢ Partial volume effect is a mixture of tissues in one pixel ➢ In the current context is mixture between the peripheral artery and surrounding tissue
Methodology ➢ Finding the contribution of each tissue in one pixel would provide an accurate measurement of area/diameter ➢ By using an a priori penalty as a spatial constraint within a Markov Random field framework we can model the tissue mixture ➢ The EM algorithm is used to estimate the tissue mixture
Methodology ➢ A new measurement is performed ➢ Stenosis is measured based on a taking into account the partial volume reference measurement in a healthy effect, reflected by the determined artery distal or proximal to a percentage contribution of each tissue stenotic site and the measurement , λ , in one pixel in the most severe site of stenosis
Methodology ➢ Degree of stenosis is determined according to the area percentage, by a scale 5 point : • 0 healthy artery, • 1 (1-49%) • 2 (50-69%) • 3 (70-99%) and • 4 occlusion ➢ Length of stenosis shows severity of the disease and is expressed through two categories: • short (< 1cm), 1-3 cm, 3-5 cm, 5-10 cm • long (>10)
Evaluation Real patient data ➢ Twenty CTA datasets provided by CHUV, Lausanne but 18 were used ➢ GE Multi-slice helical CT ➢ Different sizes from 512 x 512 x 900 to 512 x 512 x 1050 ➢ Resolution: 0.703125 x 0.703125 x 1.25 Phantom data ➢ Mimics the shape and attenuation properties of the artery ➢ Diameter ranges from 1 to 8 mm ➢ Stenosis is simulated ➢ Different acquisition protocols (10)
Evaluation Threshold Applied on 18 datasets ➢ Optimal threshold: 140-200 HU ➢ Results Visual evaluation showed that all ➢ arterial segments were present
Evaluation Region Growing ➢ Applied on 18 datasets ➢ 6 neighbourhood connectivity system Results ➢ The arteries were extracted correctly in 2 datasets ➢ In 16 datasets the extracted arteries were connected to the bone
Evaluation
Evaluation Vessel profile analysis ➢ Applied on 15 segmented data, each partitioned in 35 segments (total of 525 segments) ➢ Ground truth: 149 segments with stenosis: 132 soft plaque, 17 calcifications Parameters ➢ The width of the 1D Gaussian filter was set to 9 ➢ The parameter k in the discrete curve analysis was fixed to 20 ➢ The parameter θ s was set to 0.55 Results ➢ The method identified 116 stenosis caused by soft plaque ➢ 33 stenotic areas were missed (17 calcified) ➢ Identified 15 false positives stenoses
Evaluation Vessel profile analysis
Evaluation ➢ The peripheral arteries were extracted with a sensitivity of 83.3% ➢ The undetected stenosis in the calcified arterial segments influences the sensitivity
Evaluation A MAP-MRF method for partial volume effect correction ➢ Evaluated on phantom data and compared to threshold method Parameters ➢ A MAP-MRF method • Number of tissues, K=2 • Degree of penalty parameter, β was set to 0.85 • Mean and variance for artery class: µ1= 200, ν 1=100 • Mean and variance for surrounding tissue: µ2= 60, ν 1=10 • EM convergence parameter δ was set to 0.05 ➢ Threshold method • Threshold value was selected 200 HU
Evaluation A MAP-MRF method for partial volume effect correction ➢ Evaluated on phantom data and compared to threshold method Results ➢ The measurement of the simulated artery in the phantom, using MAP- MRF method showed a percentage error of 8%.
Conclusıon ➢ A fully automatic CAD-CAM system for detection and measurement of peripheral arterial diseases in CTA images has been developed and implemented ➢ Results for stenosis detection showed a sensitivity of 78% and specificity of 96% (15 false positive in all datasets), with an increase in sensitivity (to 88%) if calcified areas were excluded ➢ The partial volume effect was corrected and CAM component showed an average errot of 8% when evaluated on phantom data ➢ The computation time for processing 1000 slices is 8 min, while radiologist examination is 1- 4 hours
Recommend
More recommend