Introduction Autofluorescence endoscopy (AFE) is a novel technique that identifies early neoplasia in Barrett's oesophagus (BE) by highlighting differences in tissue autofluoresence (AF). It has high sensitivity but is associated with false positive rates up to 80%. We aimed to develop numerical measures of AF to reduce the false positive rates of AFE.
Methods Images of AFE lesions in patients with BE were prospectively collected. Blinded anonymised images were de-gamma corrected and average grayscale values in the red, green and blue channels of the abnormal and background normal area were quantified. The autofluoresence intensity (ratio of average red to green channel greyscale value of lesion compared to background), colour contrast index (between lesion and background), hue, saturation and lightness (of the lesion) were calculated. A decision tree based on the training set was developed with the J48 algorithm in WEKA3.2.4, using a 10-fold cross validation strategy. The performance of the model developed was assessed on an independently collected test dataset.
Results There were 82 images (37 high grade dysplasia/cancer) in the training set and 164 images (51 high grade dysplasia/cancer) in the test set. The decision tree classifier developed utilised only autofluoresence intensity and colour contrast index and had a sensitivity of 97%, specificity of 77%, and negative predictive value of 98% in detecting high grade dysplasia/cancer in the independent test set. The false positive rate of AFE was reduced from 70% to 16%.
Conclusion Numerical analysis of colour fluorescence and contrast is a reliable, objective and accurate method of reducing the false positive rate of AFE and can be easily incorporated into real time endoscopy.
Competing interests V Subramanian: None declared, J Mannath: None declared, D Boerwinkel: None declared, L Alvarez-Herrero: None declared, W Curvers: None declared, C Hawkey: None declared, J Bergman Grant/Research Support from: Olympus Medical, K Ragunath Grant/Research Support from: Olympus Medical.