Evaluation of Functional Data Clustering Algorithms on Renogram Curves to Aid in the Diagnosis of Kidney Obstruction Open Access

Park, Kevin (Spring 2018)

Permanent URL: https://etd.library.emory.edu/concern/etds/f4752g745?locale=en


Kidney obstruction is a serious condition where the body’s urinary system develops resistance to urine outflow. Radionuclide renal imaging plays a major role in evaluating a kidney with suspected obstruction. Current practices model 99mTc-mercaptoacetyltriglycine (MAG3) gamma tracer concentration across time through a renogram curve to help in the diagnosis of kidney obstruction. One important issue that exists in interpreting kidney obstruction is that there are high misclassification rates of whether a kidney is obstructed or not because of the lack of training in diuretic renography among radiologists. The objective of this work is to provide another perspective on the statistical and computer-assisted diagnosis for kidney obstruction by assessing functional data clustering methods in the classification of renogram curves. We considered seven existing algorithms with the training dataset (N=147) where 23.80% of the renogram curves were from obstructed kidneys. We first use the training dataset to evaluate the unsupervised clustering algorithms against the consensus on kidney obstruction from three experts. We then evaluate the accuracy of prediction by using another dataset where we predict the obstruction status for each kidney from clustering methods developed with the training data. We also assess the performance of the best performing clustering methods in a group of kidneys which are difficult to interpret. The clustering algorithms of fscm, waveclust, and itersubspace provide reasonable results for the training set with a kappa of 0.7917, 0.7322, and 0.522 respectively. These three methods resulted with a sensitivity of 82.86%, 77.14%, and 91.43% and specificity of 95.54%, 94.64%, and 74.11% in the training set respectively. With the validation set, we find that only the two algorithms of fscm and waveclust perform strongly with a kappa of 0.7273 and 0.6957, sensitivity of 75.00% and 100.00%, and specificity of 95.00% and 80.00% respectively. With the difficult renogram curves, we saw that fscm and waveclust gave ratings most similar to expert one while often not aligning with the majority expert rating. These two algorithms have shown potential to separate obstructed and unobstructed kidneys. Studies with larger sample sizes may provide further insight to the success of these algorithms in assisting kidney interpretations.

Table of Contents

1) Introduction ….                                                                                                                         1

2) Data                                                                                                                                         4

3) Methods and Clustering Method Backgrounds                                                                5                                                                    

4) Methods of Analysis                                                                                                            17

3) Results                                                                                                                                 23

4) Discussion …                                                                                                                            29

5) References …                                                                                                                           30

6) Appendix                                                                                                                              33


About this Master's Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Partnering Agencies
Last modified

Primary PDF

Supplemental Files