PURPOSE: We present first results of a knowledge-based computer-assisted diagnosis (CAD) system for white light bronchoscopy.
METHODS: Based on a query image, a lesion of interest is interactively marked. Using a similarity search between this lesion and the annotated lesions in a reference database a suggestion for a diagnosis will be provided. Our current reference database comprises 260 images of 173 patients. Within these images 305 image regions have been annotated by clinical experts. Theregions have been assigned to three diagnostic classes ("normal mucosa", "chronic bronchitis", "tumor"). A comparison of the system’s classification performance applying and combining different color-extended textural features (cooccurrencematrices, sum and difference histograms, Gabor filters) was performed.
RESULTS: The best results could be achieved byusing statistical features derived from co-occurrence matrices of each color channel in the RGB color space. Classification rate for normal mucosa was 89%, for chronic bronchitis was 80% and for tumor was 72%. The maximumtotal classification rate is 80%.
CONCLUSION: CAD system could be applied later to diagnostic bronchoscopy technology as well as for teaching purposes.
CLINICAL IMPLICATIONS: The next step concerning the validation of the proposedCAD system will be an observer study in order to comparethe performance of physicians with and without the CADsystem.
DISCLOSURE: Heinrich Becker, No Financial Disclosure Information; No Product/Research Disclosure Information