Study On Projection Driven Reconstruction Technique For Medical CT Analytical Formulae

Analytic reconstruction formulae are widely adopted in modern medical CT scannerfor CT image reconstruction. Normally, the"Voxel–by-Voxel"algorithm, by which theX-Ray attenuation rate of the voxel is reconstructed one by one, is used to solve theformulae. The advantage of the algorithm is the well developed technique and it's easyto implement the algorithm on the general processor based computing platform. Thedisadvantage of the algorithm is that it consumes long computation time, and a lot ofsoftware and hardware processing elements should be used. The disadvantage of"Voxel–by-Voxel"algorithm is that, the reconstruction system is in idle state when the projectiondata are being transferred. Moreover, access of multiple frames of projection data mustbe performed to reconstruct one single voxel. Also, one projection frame contains theprojection data for many voxels, the data coupling problem occurs. These two aspectslead to too much random access of the projection data memory.In this dissertation, the numerical implementation for medical CT analytic recon-struction formulae is brie?y reviewed firstly. Secondly, a new algorithms is proposedfor solving the medical CT analytic reconstruction formulae, the behavior level modelingtechnique is implemented for the hardware acceleration of the analytic reconstruction for-mulae. At last, the improvement of one medical CT scanner reconstruction subsystem iscarried out using the projection driven algorithm. In detail, the main contributions of thedissertation are listed as following:(1) To overcome the shortcomings of the"Voxel-by-Voxel"algorithm, we derivea common analytical formula firstly, and the formula is proved. Then, the projectionreconstruction algorithm is proposed by changing the calculation order and the integralset. The advantage of the algorithm is that the image reconstruction is not in idle statewhile data are being transferring, and the projection data memory can be reduced to storeone frame, moreover, the projection data memory is accessed on only one time. Forimplementing the formula in different CT scanners with different data transfer methods,we give two different timing model, one is the direct transfer model, the other is theindirect transfer model. According to different timing model, the relationship betweenthe data transfer speed and the data processing speed is given; (2) In some analytic reconstruction formulae, there exists a weighted distance whichis normally the function of the distance between the X-ray source and the voxel. To speedup calculation of the distance, we propose an improved 3D image ray traverse algorithm.In the algorithm, the comparison of the distance increment is carried out firstly, then thedistance and the next voxel information is produced in the next step. Compared withthe classic ray traverse algorithm, the algorithm complexity is not increased. But all theresults are achieved in two computing cycle while the classical algorithm needs threecycle;(3) It costs quite much time to solve the exact reconstruction formula for spiralcone-beam CT , and it's really hard to be implemented in real-time, we propose a newalgorithm to solve the formula based on the projection driven algorithm. The two inter-polation algorithms for calculating the filtered projection data and the X-ray attenuationare also given;(4) We introduce the VLSI front-end design ?ow to the hardware acceleration of CTreconstruction. We also introduce the transaction level modeling technique to the behaviormodeling of the acceleration hardware for medical CT reconstruction. Then, we proposea VLSI architecture for the spiral cone-beam acceleration. Also, we implement the be-havior model of the architecture. Then, we point out that the on-chip memory storingthe projection data is one performance bottleneck. We take out the performance analysisand the experiment results show that double the on-chip memory size can increase theperformance of the architecture up to 3 times;(5) One FBP formula for circular fan-beam reconstruction formula is implementedusing the projection driven algorithm and the"Voxel-by-Voxel"algorithm. With the sameconfiguration, we implement both the"Voxel-by-Voxel"algorithm and the projection al-gorithm. The experimental results show show that both the algorithm have a performanceof up to 6 line-pairs/centimeter. But with the projection driven method, the CT reconstruc-tor consumes 4s to finish the computation, compared with 5s of the"Voxel-by-Voxle"algorithm, which can save up to 20% of time.