|M.Sc Student||Ben-Lavi Azi|
|Subject||Improving Gamma Camera Imaging Based Renal Function|
Diagnosis by Using Temporal Data
|Department||Department of Biomedical Engineering||Supervisors||PROF. Haim Azhari|
|PROFESSOR EMERITUS Ora Israel|
|Full Thesis text - in Hebrew|
Purpose : Radionuclide imaging is a routine diagnostic tool in the evaluation of renal function. A dynamic acquisition protocol is used, as a rule, to produce a set of images demonstrating the presence and amount of radiotracer concentration in the kidneys over time. The purpose of our study was to develop and test supervised learning algorithms for automatic analysis of the dynamic image sequence. We have further applied these methods for diagnosis of acute tubular necrosis (ATN) in renal transplant patients.
Materials and Methods: Twenty-eight patients with normal functioning (N) transplanted kidneys and sixteen patients with acute tubular necrosis (ATN) early after transplant were imaged using a scintigraphic 99mTc-DTPA imaging protocol with 56 frames (32 X 2sec frames followed by 24 X 60sec frames). Scintigraphic data were digitally stored and transferred to a PC with Matlab ® software for analysis. Using the summed images regions of interest (ROI) was manually drawn over the analyzed kidney and time activity curves (TAC) were extracted by summing the counts in the ROI. TACs were then subjected to automatic analysis.Two algorithms were developed. The first method, further defined as “PCA +LDA”, is based on concatenation of Principal Components Analysis (PCA) followed by Linear Discriminant Analysis (LDA). Samples from a training set were used to calculate the PCA transformation matrix and to obtain LDA allocation rule. In order to classify a given TAC it is first projected onto the PCA space and then LDA allocation rule is applied to the projection coefficients.
The second suggested algorithm referred herein as the “PCA residual”, makes use of the fact that the process of projecting a TAC from the original data space into a PCA space of lower dimension, and then projecting back into the original space introduces a reconstruction error. PCA transformation was calculated to both N and ATN groups and a given TAC is assigned to the group for which minimum reconstruction error in the previously described calculation was obtained.
Results: Both algorithms have shown good performance indices for separation and differential diagnosis between the two groups of N and ATN kidneys. The “PCA +LDA” algorithm performed with Sensitivity=93%, Specificity=93% and Accuracy=93% and the “PCA residual” algorithm performed with Sensitivity=100%, Specificity=93% and Accuracy=96%.
Conclusion: Quantitative TAC and PCA indices obtained from dynamic scintigraphy are a promising tool for objective, automated detection, and accurate diagnosis of ATN in patients undergoing kidney transplant. These algorithms can be also applied to other clinical settings for optimized computer-aided diagnosis.