1774 INTEROBSERVER AGREEMENT AND ACCURACY OF CONFOCAL LASER ENDOMICROSCOPY FOR IN VIVO DIAGNOSIS OF BLADDER CANCER
INTRODUCTION AND OBJECTIVES
Probe-based confocal laser endomicroscopy (CLE), an emerging optical imaging technology that enables in vivo microscopy, is a promising intraoperative adjunct to white light cystoscopy (WLC) to better characterize bladder lesions. As interobserver variance is well documented in standard histopathology, this study aims to determine the interobserver agreement and accuracy of novice users utilizing CLE to diagnose bladder cancer.
METHODS
Urologists and non-clinical researchers with no experience in CLE were recruited to participate in a two-hour computer-based training that consisted of a teaching and validation set of intraoperative WLC and CLE video images from patients undergoing transurethral resection of bladder tumor at the VA Palo Alto Health Care System since 2008. The teaching set included 15 CLE images to train the participants to identify key CLE features observed in benign, low grade and high grade bladder cancers. Participants then reviewed a validation set split into 3 sections: CLE images alone, WLC images alone, and WLC+CLE images combined. Each section contained 32 corresponding but randomized images of benign, low grade and high grade cancer. Accuracy was determined by comparing the diagnoses obtained from the participants to histopathologic results.
RESULTS
In assessing interobserver agreement for the urologists (n=7), κ statistic and percent agreement was 0.43 and 62% for WLC, 0.44 and 63% for CLE, and 0.49 and 66% for WLC+CLE. Accuracy was 60% for WLC, 63% for CLE, and 67% for WLC+CLE with significant difference between WLC and WLC+CLE (p<0.01). In detecting HG carcinoma, sensitivity was 57% for WLC, 63% for CLE, and 73% for WLC+CLE with significant increase from both WLC to WLC+CLE and CLE to WLC+CLE (p<0.02), and specificity was 80% for WLC, 80% for CLE, and 79% for WLC+CLE. For non-clinical researchers (n=4), κ statistic, percent agreement, and accuracy for CLE was 0.56, 71%, and 63%, respectively.
CONCLUSIONS
We have shown that novice users can be trained to interpret CLE images to similar levels of interobserver agreement (moderate) and accuracy as standard WLC after participating in a two-hour training. Also, compared to standard WLC, accuracy improves with the adjunctive use of CLE. Moreover, sensitivity in detecting high grade cancer increases from WLC to WLC+CLE with no change in specificity. Interestingly, compared to urologists, non-clinical researchers exhibited a higher level of agreement with similar accuracy for CLE, suggesting that clinical training is not a prerequisite for the accurate interpretation of CLE images.