An image analysis tool for the classification of lesions suspicious for malignancy in autofluorescence bronchoscopy

Abstract

BACKGROUND: Autofluorescence bronchoscopy (AFB) has been utilized over the past decade, proving to be a new powerful tool for the detection and localization of premalignant and malignant lesions of the airways. AFB is, however, characterized by low specificity due to the high rate of false positive findings (FPFs) observed, caused mainly by inflammations which often produce abnormal fluorescence. According to several clinical trials, the percentage of the FPFs is about 30%.

OBJECTIVE: In this paper we present a computerized image analysis tool for the classification of lesions suspicious for malignancy, in order to help physicians to distinguish between true and false positive findings, and thus enhance the diagnostic value of AFB.

METHODS: For the development of the image analysis tool, several colour and texture analysis methods, feature selection techniques and pattern classification models were utilized and combined. 715 AFB images from 11 specific cases have been used; 6 of these cases corresponded to malignancy, whereas the other 5 corresponded to FPFs.

RESULTS: The presented system achieved to correctly classify all the cases studied, demonstrating correct classification rate as high as 95.4%.

CONCLUSIONS: The preliminary results suggest that the proposed system may improve the diagnostic accuracy of AFB.