NE-AUA 2006 Annual Meeting, September 28 - 30, 2006, The Westin Hotel & Rhode Island Convention Center Providence, Rhode Island
Back to Scientific Program
Back to Annual Meeting
The Use of Spectral and Spatial Analysis to Improve the Utility of Urine Cytology in the Diagnosis of Transitional Cell Carcinoma (TCC) of the Bladder
Rohit Garg, MBBS, MPH1, Cesar Angeletti, MD, PhD2,Yan Li, MD, PhD3, Joseph Renzulli, MD1, Edward M. Uchio, MD1, David L. Rimm, MD, PhD2.
1Department of Surgery, Yale School of Medicine, Yale University, New Haven, CT, USA, 2Department of Pathology, Yale School of Medicine, Yale University, New Haven, CT, USA, 3Cancer Institute, Sun Yat-Sen University, Guangzhon City, China.

Background: The gold standard for the diagnosis of TCC of the bladder is cystoscopy with voided cytology. Urine cytology provides limited information (low sensitivity) and does not obviate the need for bladder visualization by cystoscopy. Although color is information rich, current morphologic diagnoses in urine cytology are based almost exclusively on spatial information. We hypothesize that using the quantitative assessment of color (spectral imaging) with algorithm based morphologic (spatial) analysis in a machine learning environment will improve the ability of urine cytology to detect TCC in voided specimens.
Methods: Papanicolaou-stained urine cytology Thin Prep slides from patients presenting for the diagnosis of TCC were analyzed. High power field (400X) multispectral images were acquired using a high quality specialized microscope, charge-coupled device (CCD) camera, liquid crystal tunable filter, and acquisition software. Multispectral image data was analyzed using the Los Alamos-developed GENetic Imagery Exploitation (GENIE) package, a hybrid genetic algorithm that classifies images using automatically learned spectral-spatial features. Proprietary sets of mathematical algorithms for voided cytology were previously developed using urine cytology slides from patients with the unequivocal diagnosis of urothelial carcinoma. These algorithms were used to classify specimens as benign or malignant. Statistically analysis was performed using receiver operator characteristic (ROC) curves.
Results: Using prior voided urine cytology specimens from two different institutions, the GENIE algorithm utilized spectral and spatial information to diagnosis urothelial carcinoma with a sensitivity and specificity of 85 and 95%, respectively. Further analysis of specimens from 66 independent patients revealed the area under the ROC curve of 0.79.
Conclusions: Spectral-spatial analysis in a machine learning environment (GENIE) is a promising methodology to increase the utility of voided urine cytology in the diagnosis of TCC of the bladder. New algorithms are currently under development to further increase the sensitivity and specificity of urine cytology with the intent of minimizing the need for invasive procedures to diagnose TCC.


Back to Scientific Program
Back to Annual Meeting