Cervical cancer disproportionately affects low and middle income countries. Automated visual evaluation – using deep learning to analyze a digital cervix photograph – has been proposed for patient management. Image quality remains a key challenge, as it can be degraded by many types of image defects. A series of such defects were artificially added to a test set consisting of N=344 digitized cervigram images from existing studies. Replicate test sets were created for different image defects: blur, recoloring, obstructions of different colors and directions, rotations, and white Gaussian noise. The augmented images were evaluated by a classifier. The two most significant image defects were blur and Gaussian noise.
|